THE ORGANIZATIONAL ANALYTICS E-BOOK

A GUIDE TO DATA DRIVEN ORGANIZATION DESIGN

PHANISH PURANAM (INSEAD) JULIEN CLEMENT (STANFORD UNIVERSITY)

Version 2.0 (October 2023)

(Send feedback to [Link]@[Link] or jclement@[Link])

PREFACE
Knowing how to organize may be humanity’s most significant and oldest accomplishment.

But our lives today are also influenced by another, more recent, human achievement:

algorithms. This book is about using the power of data and algorithms to design better

organizations. We were inspired by our experiences teaching the Org2.0 elective to

MBAs at INSEAD and Stanford.

The ideas are drawn from our research in organization science, in which we have

extensively relied on algorithms to analyze data from organizations and simulate

organizations using computer models. The inputs of numerous industry experts who

contributed to our classes as guest speakers and lecturers are also very gratefully

acknowledged as they helped build the bridge from theory to practice.

This book is free. We’ll be happy if you use it for teaching, research, consulting, or even

entertainment. “How to organize” seems too important a problem today for us to put the

little knowledge we have gathered on the topic behind a paywall, especially as we

struggle to adapt to major shocks like the COVID-19 pandemic and the rapid

developments in generative AI (such as ChatGPT and DALL-E2) aspire to make our

organizations resilient to future shocks. We do request you cite the book if you use it.

The book is a living document— we hope to keep revising it based on your feedback. Do

send us your thoughts: [Link]@[Link] / jclement@[Link]

THE ORGANIZATIONAL ANALYTICS E-BOOK 2

TABLE OF CONTENTS

Chapter 1: What this book is about 4

Chapter 2: The core problems of Organization Design 11

Chapter 3: Organization Design Analytics 16

Chapter 4: The first lever of organization design - Structuring 28

Chapter 5: The second lever of organization design - Sorting 47

Chapter 6: The third lever of organization design – Sensemaking 56

Chapter 7: An eye on the future 67

THE ORGANIZATIONAL ANALYTICS E-BOOK 3

CHAPTER 1
What This Book Is About

Organization design matters. Irrespective of size, sector, strategy, and goal, every

organization must be designed in a way that helps its members pull together effectively

to accomplish its objectives.

Organization design decisions come in all shapes and sizes:

● how (and whether) to form organizational units like teams, departments, and

divisions;

● how to set up reporting and accountability;

● what incentives to use;

● which employees to hire, retain and promote;

● where to physically locate employees;

● how to best organize remote collaboration...

Notice that some of these topics are at the group level whereas others are primarily at the

individual level. Both matter because an organization works well when individual efforts

come together to reach group objectives. But in any given project we might put more

emphasis on one level or the other. All of these are organization design decisions that will

tremendously impact how people and groups in your organization perform today,

tomorrow, and beyond.

THE ORGANIZATIONAL ANALYTICS E-BOOK 4

 ORGANIZATIONAL STRUCTURE & DEVELOPMENT PROBLEMS

Yet the way we currently make these decisions is, all too often, shockingly unsophisticated.

That’s why we wrote this book. Imagine having a headache and hearing the following from

your doctor:

“I have what you need! Ancient wisdom says that herbs gathered under the harvest moon

cures just your symptoms. Mr. X and Ms. Y, my other patients, have been using it regularly

and they’re really quite healthy for their age! Clinical trials? No there aren’t any.”

We wish this were a gross caricature. But this is roughly how many organization design

recommendations are currently made, whether by managers within companies, advisers

THE ORGANIZATIONAL ANALYTICS E-BOOK 5

from consulting firms, or bureaucrats within governments. (Let’s be honest; we probably did

it a few times ourselves). Clichés such as “It’s all about incentives” and “culture eats [pick

your least favourite function] for [pick your favourite meal]” take the place of a science based

approach to the problem. Trials to validate whether the recommended design can actually

work for a particular organizational context are unheard of. Examples of “best practice” take

the place of Evidence.

Copying industry best practice is a major driver of decisions in today’s corporate world,

and not just in the domain of organization design. This is a problem for two reasons.

● First, the evidence that practice truly worked well elsewhere is usually fig-leaf thin.

Correlation is not causation, and it is easy to be fooled by randomness into thinking

that a particular structure or practice drives success when in fact it does not (or

perhaps even harms performance!). Mr. X and Ms. Y’s health may have nothing to do

with the pills they are popping, or worse, perhaps they are healthy despite them.

● Second, all organizations are different. Just because something worked well

elsewhere does not mean it will work for your organization. Even what really keeps

Mr. X healthy (or at least does no harm to him) may be poison for you. At best, you

may be paying for useless pills. The biological variation that exists between people is

almost nothing compared to the social variation that exists between organizations, so

we’ll let you imagine how often organizational practices imported from one company

may really be poisonous to another.

THE ORGANIZATIONAL ANALYTICS E-BOOK 6

Of course, you might say criticizing is easy. Relying on industry best practice has its problems

but perhaps there is no better alternative… is there? Yes, there is! Recent developments in

two areas have changed what is possible today in the field of organization design.

● First, developments in algorithms and the availability of “Big Data” have enabled

us to observe much more of the internal workings of companies and analyze them

with far greater sophistication than was possible before. Critically, the data and

associated methodologies allow us insight at a micro-level into how individuals are

affected by the organization designs they live in, as well as how their behavior

aggregates up to produce the outcomes of their organizations.

● Secondly, new theories of organization design have emerged that move the action

away from the “boxes and arrows” of the organization chart and instead help us think

about design at the level where the data and indeed the innovations in organizing

now reside: individuals, teams, online communities, and networks—all of which may

exist within the legal boundaries of a large corporation or government department,

or indeed cut across them. Together these developments enable an organization to

analyze its own data to design itself.

An article and short video on the contrast between old and new approaches to

organization design
EMBEDDED LINK

We use the term “Org2.0” to refer not only to radically new organization designs that are

emerging today in response to developments like the post-millennial workforce,

THE ORGANIZATIONAL ANALYTICS E-BOOK 7

advancements in AI and the COVID19 pandemic, but also to the suite of methodologies

and thinking tools that enable organization designers to make much more sophisticated

decisions than they could in the past (with “Org1.0” if you will). We can use these new

techniques to better perceive the organizational environment of a company, make predictions

about specific events in specific parts of a company, and experiment quickly but rigorously

with new practices to choose the ones that work best. These are tools for organizations to

design themselves, rather than rely on designs copied from elsewhere. With these, we want

to show you how to design, not what design to use.

New designs emerge all the time - “flat firms”, “holacracies”, “all remote”,

decentralized autonomous organizations (DAOs)” and “agile” are the ones in vogue as

we write this- but they often wax and wane (see the last chapter for some of our thoughts

on these). The shift towards data-driven design, on the other hand, won’t be going away.

The goal of this e-book is to give you an intuitive understanding of the analytical tools

which make this possible, provide a framework to link these tools together and show you

THE ORGANIZATIONAL ANALYTICS E-BOOK 8

why (and when) they can lead to better decisions with sizeable benefits. And all that in 7

very short chapters! - and quite a few embedded articles and video links if you’re curious.

Reading this book won’t turn you into an analytics expert, but it should make you fluent

enough to interact with and evaluate the work of analysts whose reports will support

your decisions. The ideal data science-based organization designer would be a bit of a

Unicorn- in this widely shared image from IBM labs, you can see that it’s hard to find people

with all the required skills.

But we can build teams that collectively have these skills, and one of the basic things we

know about teams is that they are able to

work better together when the members

know enough about what others know and

how to use each other’s expertise.1 Our

book is written primarily for the “subject

matter experts” in the picture above, and

the managers whose teams hold these

different skills. Accordingly, our approach

will be to “black-box” the technical detail in

the text: we’ll focus on what analytical

techniques can do for you rather than how

they do it. The embedded video lectures and

links provide extensive details on several analytical methods for curious readers.

1
Moreland, Argote and Krishnan (1996);Reagans, Miron-Spektor, & Argote, (2016)

THE ORGANIZATIONAL ANALYTICS E-BOOK 9

We have prototyped and refined this approach extensively with MBA students and senior

executives we have taught at INSEAD and Stanford. It seems to work well... but in the

spirit of this book, we hope you will experiment with it before deciding. For the same

reason, this book contains no detailed case studies about exemplar companies that are

using Org2.0 approaches in their organizational development work. We know first-hand

many companies that do, and many of them have visited our classrooms, but using them

as examples to argue our case would constitute nothing but the “best practice” fallacy all

over again. So, read on and make up your own mind!

A short video providing an overview of the Org2.0 approach

EMBEDDED LINK

Reflection question: why do you think “people” related decision-making in

organizations has been slower to rely on data and evidence, compared to

decisions about customers, finance, and operations?

THE ORGANIZATIONAL ANALYTICS E-BOOK 10

CHAPTER 2
The core problems of Organization Design

Overview

● It’s useful to think of any group with a goal as an “organization”. That includes

teams, departments, divisions, and entire companies.

● Every organization, regardless of size, location, and industry faces the same

universal problems- how to divide labor and then integrate effort.

● While the problems are universal, the solutions are not. Finding solutions that work

for your context is the essence of organization design.

At the most basic level, Org2.0 is just a particular approach to solving the age-old problem

of organization (re)design: how to create systems of individuals who may differ in what

they know and personally want, that can still accomplish a goal together. That involves a

few fundamental problems which are neither new nor necessarily unique to human

organizations. Ants face the same challenges, as do computer networks, and squads of

THE ORGANIZATIONAL ANALYTICS E-BOOK 11

robots working together- though as far we know, none seem to

rely on management consultants to solve them...

What are these fundamental problems? Division of labor and

integration of effort.2 Roughly, these correspond to “breaking

things down” and “putting them back together again”:

organizations divide their activities into different tasks performed

by different people and groups (often because specialization

yields efficiency), which then need to coordinate to produce a

coherent output.

We can detail it down further. Designing the division of labor requires partitioning the overall

goal of the organization into sub-tasks (task division) and assigning sub-tasks to

individuals and groups (task allocation). Designing the integration of effort also involves

dealing with two related sub-problems: motivating individuals through money, autonomy,

recognition, and other means (reward distribution) and making sure they are sufficiently

informed about what they need to do (information flows). And since no set of solutions to

these problems is likely to be either perfect or permanent, we need some mechanisms for

re-solving these problems when exceptions arise (exception management).

Embedded Link: The universal problems of organizing illustrated

EMBEDDED LINK

2
Mintzberg, (1979); Burton and Obel (1984); Puranam, Alexy and Reitzig (2014)

THE ORGANIZATIONAL ANALYTICS E-BOOK 12

 The Universal Problems of Organizing

When we (re)design an organization (and we stress “re-” because most of the time, we

must change existing designs) we are always trying to answer the same basic questions.

Given the goals of the organization we are trying to design.

● what needs to be done (task division)?

● who does what (task allocation)?

● who gets what to motivate effort (reward distribution)?

● who knows what in order to ensure competent and coordinated effort

(information provision)?

● what to do when exceptions arise and new solutions are needed? (exception

handling).

THE ORGANIZATIONAL ANALYTICS E-BOOK 13

To summarize: designing an organization means searching for solutions to these

universal problems. Improving the effectiveness of an organization (a.k.a. redesign) is

equivalent to finding better solutions to the universal problems of organizing that confront

the particular organization. This will be true in every existing organization you know of,

whether it is an “old economy” manufacturing company, a cool new technology start-up,

a software development team, a football team or a string quartet. It will equally be true

when you are trying to figure out whether the entire company should be reorganized, and

when you are simply interested in how to get a particular team to work more effectively.

That is the beauty of thinking in terms of universal problems! The problems are always

the same, and the differences will be in the solutions deployed by different oganizations:

different solutions work for different organizations, which is why you should always be

very careful in asking yourself whether what other companies are doing will also work for

you.

Reflection question: Think of any group that you have recently been a part

of, at work or home, that had a specific goal or task. Can you explain what

solutions this group had to each of the core problems of organizing? How

effective were these solutions?

THE ORGANIZATIONAL ANALYTICS E-BOOK 14

CHAPTER 3
Organization Design Analytics

Overview

Analytics-driven approaches to organization design can be conducted at three levels of

sophistication:

● Perception- describe what’s happening now; this is where Big Data combined with

traditional statistics have been very useful

● Prediction- forecast what’s going to happen before it does and – this is the domain

of machine learning/AI applied to Big Data

● Prototyping- decide which intervention has the best chance of producing the

results you want - this is the world of A/B testing.

You don’t always have to go for the most sophisticated approach. For some problems, it

might make better business sense to just rely on perception or prediction. You have to

consider the value of the information generated through prototyping.

What does an analytics and data-driven approach to organization design – finding

solutions to the universal problems of organizing that are applicable in your specific

context- look like in practice? The analytics-driven design comes in three levels of

sophistication: perception, prediction, and prototyping.

THE ORGANIZATIONAL ANALYTICS E-BOOK 15

 ORGANIZATION DESIGN & DEVELOPMENT PROBLEMS

We’ll use a typical organization design challenge as our running example: the problem of

“silos” (i.e., absence of coordination) between different units within an organization. Silos

are often blamed for delays, missed opportunities, cost inefficiencies, and lack of

innovation. How would an analytics-based approach tackle this most fundamental of

organization design challenges?

The first step towards better design is perception. Perception is about the present. It

involves forming an accurate picture of how the organization is working currently. A major

THE ORGANIZATIONAL ANALYTICS E-BOOK 16

requirement in our case is to gather data that allow us to answer some basic questions,

such as:

● How frequent are the instances of breakdown in collaboration?

● What is the (estimated) economic impact of these problems?

● Where do they seem to occur most frequently?

Many apparently “burning issue” design challenges may actually turn out not to be so

important after all; as we will show later, “silos” are not per se a problem. In fact, it’s nearly

impossible to organize a large company without some silos. The trick is to have the silos

where they should be, protecting from distraction the things that need such protection,

but allowing interactions between people who need to interact to get things done. To

figure this out, we need data. Archival records, surveys, and even interviews can help us

see if there really is a systematic problem. All too often, we are just listening to one irate

(and vocal) senior manager who feels their project did not get the attention it deserved

from colleagues in other departments. When we have a large enough sample of data, we

can run perception analytics (summary statistics, cross-tabulations, and hypothesis tests

for example) to assess whether the data are showing a pattern or just noise.

A video lecture on the basis statistical tools for Perception

EMBEDDED LINK

THE ORGANIZATIONAL ANALYTICS E-BOOK 17

Good data also enables us to perceive problems before they become widely visible (or

because they are being masked by politics). Are certain interfaces between departments

particularly prone to friction? Are the “missed opportunities” created by silos imposing a

large opportunity cost on the organization? Sometimes, data is useful to simply debunk

myths around any of these types of issues too. Consider “Those folks in marketing live in

their own world!” vs the data showing that most of the breakdown in handoffs are

occurring between design and manufacturing. Data can also help to generate insights and

test hunches. For instance, your data might show (or you might be able to test your hunch)

that silos are particularly damaging when they coincide with geographic or time-zone

boundaries. Put simply, the analytics of perception is first and foremost about making sure

we are perceiving the right problem- it is the equivalent of a thermometer or a blood test

to detect malaise, as opposed to judging from looking at the patient or holding a thumb

up to their forehead, or just asking them how they feel.

The next level of sophistication involves prediction. Here, we move from understanding

what has led the organization to work in a certain way until now to predict what will

happen to it in the future. This is like a blood test or an X-ray – it helps predict what

happens next, not just describe what is happening now. In doing so, we can also move

from simple aggregate patterns (“Legal department is routinely holding up approval on

bids our engineering team is making, and we are losing business”) to forecasts about

specific events (“Tom’s bid that he needs to get to the client in 15 days is going to be held

up at legal cell with a 69% risk of failing to meet their deadline!”). Making accurate

predictions can be very powerful because ultimately all decisions rely on predictions:

accurate predictions lead to better decisions. Consider this: every time you make a

decision, you choose between alternatives, based on your beliefs about which will

THE ORGANIZATIONAL ANALYTICS E-BOOK 18

produce a better outcome for you. That’s a prediction problem, and in many situations

machine learning algorithms (the relevant branch of AI) can help build predictions (for

instance, by looking at historical data about projects and approval timelines).

A video on the basic machine learning tools for predictive analytics

EMBEDDED LINK

Though very useful, machine learning isn’t the most sophisticated technique you can use

for organization design. That’s because prediction isn’t the same as explanation: just

because we can predict that something is likely to happen doesn’t mean we know why it

may happen. For instance, we might be able to predict that contracts written by a certain

sales manager are likely to face more delays at the legal unit than those written by other

managers, but that doesn’t mean we know whether this is because these contracts are

poorly prepared or because the legal unit lacks expertise about these types of contracts.

Org2.0 thinking, therefore, suggests that we use predictive analytics to produce indicators

that something is worth investigating (“let’s keep an eye on the contracts that are forecast

to go slow and put some energy into chasing those”) rather than directly as interventions

(“these contracts are likely to take forever to clear, let's replace the managers!”). Indicators

could be variables that are good predictors of other variables, and the predictions

themselves can be useful as an early warning system that something is likely to happen.

An intervention on the other hand is a manipulation of a variable that causes a change in

THE ORGANIZATIONAL ANALYTICS E-BOOK 19

another variable. Indicators can give us good hunches about interventions but should not

be confused with interventions.

There are three other reasons why we advise caution with machine learning. First,

predictive accuracy often comes at the price of an intuitive explanation. More complex (and

less explainable) algorithms may make better predictions in many situations but leaving

our decisions entirely to “black-boxed” algorithms can be highly problematic when there

are significant human consequences to the decisions taken as a result. Algorithms

optimize what we tell them to optimize (for instance, “find strong correlates of who is likely

to be a high performer in the data”) but not any of the criteria we usually leave implicit in

our decisions (“make sure we give equal chances to employees from different

backgrounds”). They learn from historical data that may reflect past biases in decision-

making, and if used blindly may perpetuate such biases.

Second, human beings are hardwired to try to understand why things happen. People are

much more likely to accept and support decisions when they understand the reasoning

behind them. Telling them that “the random forest algorithm says this is likely” won’t cut

it. This is why predictions are a good start for decision-making processes, but this is not

an end in itself.

Third, just having a lot of data does not guarantee that machine learning can work for you.

There are a number of potential challenges with data, that even in large volume, may make

it useless. In essence, all machine learning techniques work on the premise that data

gathered today are representative of how the underlying processes generating the data

THE ORGANIZATIONAL ANALYTICS E-BOOK 20

will unfold tomorrow. That may be false for a variety of reasons. In general, the stronger

the resemblance between the situations you are gathering data from and applying it

to, the better.

A short article on why lots of data don’t always translate into success with

machine learning EMBEDDED LINK

The last design tool, prototyping, is the only method we know that helps understand

what causes the relationships we observe in data. Here, we don’t rely on past data but

rather create new data by running “experiments” within an organization. Experiments

involve randomly assigning some units (i.e. people, teams, projects, bids) to a treatment

condition (the new policy we’re thinking of implementing) and others to the control group

(where things stay as they were without the new policy).

Randomization is crucial! Imagine that you implement (without randomizing) a new

training policy and find that employees who applied for the training and attended it saw

a rise in their performance evaluations. You have no way of knowing whether this is

because your training was effective or because the people who applied for it are just

generally more motivated high performers whose evaluations were going to rise anyway.

You might object that in this case, you might just make the training mandatory for all

employees and see whether everyone’s performance went up afterward. But then, how

do you know that the increase in performance wasn’t due to something else which

affected all employees, perhaps because demand for your products went up this year for

THE ORGANIZATIONAL ANALYTICS E-BOOK 21

unrelated reasons? Randomization allows you to avoid all of these problems by creating

counterfactuals, that is to understand what would have happened without the

intervention. This is possible because randomized treatment and control groups are

statistical twins: they are similar enough to be treated as identical, so the control group

can serve as the counterfactual. We cannot establish causation without

counterfactuals, and we cannot establish perfect counterfactuals without

randomization (unless we have time machines).

Randomization can help you understand whether any new procedure (not just training!) is

effective. For instance, we could imagine solving silo problems by setting up a new

procedure to process contracts by the legal unit. How do we know it will work? We pick

a random subset of contracts to apply the new process to, let the rest go ahead as usual,

and compare outcomes later.

That’s it! For instance, to solve the silo problem, should we invest in team building?

Change the staffing of boundary-spanning roles? Create a fast track for certain projects

that have to cut across departmental boundaries? Rather than implementing one of these

based on our hunches, we could prototype them and only roll them out widely if the

evidence looks promising.

There is one other advantage that prototyping enjoys. Prediction and perception are both

hungry for data: they require you to collect data not only on the outcome you're interested

in but also on a whole set of variables that correlate with it (i.e., the X’s that associate with

the Y’s).

THE ORGANIZATIONAL ANALYTICS E-BOOK 22

But prototyping need not be. If you randomized the treatment properly, all you need is to

measure differences in outcomes (Y’s) between the treatment and control group. This

avoids a whole lot of concerns about data privacy that often come up when using the Xs

in prediction or perception.

Video lecture on why an experiment with randomization is the gold standard

EMBEDDED LINK

Of course, experiments with randomization are costly. This means that they should be

used only when the cost of “getting it wrong” is large (i.e., when the potential benefits of

your new design are large and should not be missed out, and the potential costs are large

too if it turns out to be a failure). If either the benefit of succeeding is large, but the costs

of failure are small, or vice versa you don’t need an experiment.

THE ORGANIZATIONAL ANALYTICS E-BOOK 23

There are a host of other tricks we can use to create reasonable counterfactuals when we

can’t resort to randomization in the organization. Scenario discussions and computer

simulations, for instance, can help us understand which intervention will be effective to

solve our silo problem. One might gamify the task that people do (i.e., create a toy version

of the kind of project they usually work on) and run a hackathon over a few hours or days,

having different groups try to solve the problem after randomizing them into different

interventions that can help solve the design problem we are tackling. For instance, are

agile teams good for us? We could randomly assign some teams to agile structures while

others stay in the current design and all teams compete in a 24-hour hackathon to

complete a simplified version of a typical problem.

A short article on Gamified Randomized Control Trials

EMBEDDED LINK

Prototyping with randomization involves experiments in vivo (literally “within the

organism”). An alternative is an experimentation “in silico” through computer simulations.

Simulations allow organizations to create representations of how they work in the present

and explore a variety of “what-if” scenarios to understand how to ensure better

coordination between their members, how to convince employees to adopt new

management practices, and many other important processes involving interactions

between many agents.

THE ORGANIZATIONAL ANALYTICS E-BOOK 24

An analogy may be useful: a traffic designer or a bridge designer is unlikely to use

randomized experiments to prototype a new system of traffic lights or construct a new

bridge (we hope) but will instead rely on running the experiments within computer models

of the system. If we had a good computer model of how bids get approved, for instance,

we could see what the impact would be of changing a particular step in the process. This

is no different from trying to figure out what might happen to your projected Net Present

Value (NPV) for an investment if you change the expected growth rate of revenues or the

cost of capital.

Video lecture on alternatives to randomization for prototyping

EMBEDDED LINK

Reflection question: when do you think a Prediction approach (e.g., “Who is

likely to quit next quarter”) is superior to a Perception approach (e.g. “What

factors correlate with people quitting?”)? When is prototyping superior to

both?

Bonus Reflection question: the evidence suggests that many if not most

organizational change initiatives fail. Yet prototyping a change before full-

scale implementation seems rare (how many companies do you know that

use clinical trials to decide on their next re-design?) Why?

THE ORGANIZATIONAL ANALYTICS E-BOOK 25

CHAPTER 4
The first lever of organization design

Overview

● Structure refers to stable patterns of interaction. They can be laid down in rules or

emerge organically

● Interactions come in two types: interdependence – A needs B’s inputs to get her

job done; and influence-A‘s beliefs and actions are shaped by B’s.

● The most basic principle of structure in organization design is to make sure that

“influence matches interdependence”.

● To fix a design problem through structure, therefore, requires identifying and

correcting mismatches between interdependence and influence.

When they think about the levers of organizational (re)-design, most people think

immediately about structure: the “boxes and arrows” of the organizational chart. This view

is doubly narrow. The structure is not the only lever of organization design, and structure

is not only about the boxes and arrows. The research is quite conclusive that individual

THE ORGANIZATIONAL ANALYTICS E-BOOK 26

behaviour in an organization is jointly shaped by three levers of design: structuring, sorting,

and sensemaking.3 Therefore, we suggest that organization design analytics can

potentially operate through each of these levers individually or in combination, though

every approach is not equally feasible in all situations.

“Structure” simply means

“stable pattern of

interactions between

people”. There is a lot that

goes into that! They occur

at all levels and layers.

They involve peer-to-peer

as well as superior-

subordinate relations.

Some interactions are

mandated by somebody

with authority—they are part of the “formal” structure—and others emerge on their own—

they represent the “informal” structure. Interactions between people are the lifeblood of

an organization.

Organizational charts only show interactions that are part of the formal structure (the

figure below is the first documented instance of an organization chart, used by the New

3
For instance, see the recent review by Chapman and O’Reilly (2016). The McKinsey “7S framework” is also broadly consistent with
these ideas.

THE ORGANIZATIONAL ANALYTICS E-BOOK 27

York and Erie Railroad from the 19th century). For instance, a product division in an

organization chart is shorthand for “all employees in any function who contribute to

making and selling this product are expected to prioritize their interactions with each

other (over interactions with those in other divisions) and be subject to a common

source of authority (the director of that division)”.

No doubt, boxes, and arrows are useful

to think about complex organizations in

aggregate, “macro” terms. They hide

underlying complexity and allow

designers to focus their attention at

most a few layers deep into the

organization. However, typical redesign

efforts occur at lower levels in the

hierarchy and in smaller portions of the

organization. They also require more

direct levers to influence individual

behaviour and must deal with the gaps

between the prescribed formal

structure and the actual, realized

structure of interactions. As re-organizations of the boxes and arrows amount to different

ways to shuffle the hierarchy, such an approach comes into the game with its hands tied

when confronting new forms of organization that try to eliminate hierarchy or at least

flatten it dramatically.

THE ORGANIZATIONAL ANALYTICS E-BOOK 28

To give you a sense of how coarse

organization charts are, consider

the following example: given these

four individuals who belong to two

departments (R&D – yellow and

manufacturing- red), how

many organization charts can we

create with them? We will use

arrows to indicate who gets

assigned to work closely with whom. Perhaps you immediately thought of these two:

The one on the left is a

functional grouping, and the

one on the right is a

divisional grouping. Perhaps

you also thought of the

matrix structure (with both

vertical and horizontal sets

of arrows in the same

picture). But how about any

of these?

THE ORGANIZATIONAL ANALYTICS E-BOOK 29

In fact, how many of these alternatives are there? It turns out we know exactly, through the

simple application of the binomial theorem: for N unique individuals, the number of distinct

interaction structures (i.e. organization structures) is 2 [(N)(N-1)/2]. In our example with N=4, that

means 64 different structures. Yes, you read that correctly: the three common organization

charts we are used to seeing represent 3 out of 64 possibilities. This brutal simplification

depends on what we call the “assumption of uniformity”- i.e., every individual within a

function (of the same colour in the picture above) is identical to every other, including in how

they interact with other individuals in the organization. Sometimes that is a reasonable

assumption; sometimes it is not.

So, is the alternative to consider every possible interaction structure between everyone?

The official as well as the unofficial ones? Who listens to whom; who takes orders from

whom; who has a crush on whom?

Fortunately, these are not the only options. New thinking about “micro-structures”

suggests that we should look at large complex organizations as collections of smaller,

THE ORGANIZATIONAL ANALYTICS E-BOOK 30

 simpler, and recurring patterns of

interaction.4 Start with the simplest

possible organization—a pair of

people; let’s call them Ann and Bob.

They can have two types of

interactions: Interdependence and

Influence. Interdependence depends

on how Ann and Bob are incentivized.

If Ann gets paid regardless of what

Bob does and vice versa, there is no

interdependence between them. But if

Ann’s outcomes depend on Bob’s

actions, they are interdependent. Workflows are the most obvious source of

interdependence in organizations; if Ann’s work depends on Bob’s inputs (and Ann’s

outcomes such as evaluation, bonus, and pay depend on getting her job done), then clearly

Ann is dependent on Bob.

Influence implies that Ann can affect Bob’s actions (e.g. “She convinced me I should take

on this role”). Strong influence can arise from a direct peer-to-peer connection, or indirectly

because both Ann and Bob report to a common superior. We will use the same single-

headed vs. doubled-headed arrows to distinguish one-sided vs. mutual influence.

We can use this notation, with single and double-headed arrows to distinguish between

influence and interdependence, whether symmetric or asymmetric:

4
Puranam (2018)

THE ORGANIZATIONAL ANALYTICS E-BOOK 31

On their own, neither influence nor interdependence is good nor bad. What matters is the

match: in fact, one of the most basic principles in organization design, also sometimes called

“the folk theorem” (because many scholars independently converged on it with no clear

first) is that influence must match interdependence. The idea is quite intuitive: if there is

interdependence, you need the influence to help manage it. This is at the heart of ideas

about “fit” in organization design- the influence structure must fit the interdependence

structure, and the latter is dependent on the goals and environment of the organization.

THE ORGANIZATIONAL ANALYTICS E-BOOK 32

In fact, we can go a step further, to ask which type of influence should we use to match

the interdependence: symmetric or asymmetric? A useful heuristic is as follows: if the

interdependence involves a “win-win” situation – both actor A and actor B benefit by

aligning their actions- then a symmetric influence tie is sufficient. If however the

interdependence is of the “win-lose” type (which may still be a net win for the

organization), then you might need some form of asymmetric influence- either A

influences B more than the other way around (A B in our notation above) or both are

subject to a common source of authority (A C B).

Besides changing the patterns of influence to match the pattern of interdependence, one

can also do the opposite- change the interdependence to match the existing influence

THE ORGANIZATIONAL ANALYTICS E-BOOK 33

structure - for instance by changing task division and allocation so that A and B no longer

depend on each other, or by changing the incentives for A and B, so that their

compensation only depends on their individual contribution.

FROM A PAIR TO NETWORKS

So far, we have focused on a pair – Ann and

Bob to understand the folk theorem. But the

N-actor picture of the “folk theorem” is not that

different, just more complex looking:

The basic design principle does not change;

you want to make sure that interdependencies

are matched by realized influence network. In

fact, every pair of links in the second picture

follows the same principle of matching. If we

have data on networks of interdependence

(e.g. based on business processes or work-flows) as well as networks of influence (e.g.

who report to the same boss, or seek advice from each other), then in fact we can use

algorithms to do a check for every pair of actors whether the matching of interdependence

to influence holds or not.

An important aspect to bear in mind is that the influence network can be a consequence

of both the formal structure (e.g. reporting, grouping) as well as the informal

THE ORGANIZATIONAL ANALYTICS E-BOOK 34

structure (e.g. friendship, advise seeking), so a more complete depiction of these various

networks might look like this:

So, here’s how we apply the folk theorem or the principle of fit at large scale:

1. we must gather and aggregate data on the formal and informal structure to create

the influence network and

2. compare the result to the interdependence network.

As we compare the interdependence and influence networks, we will discover broadly

three different kinds of mismatches:

● Commission errors: an influence tie exists even though there is no

interdependence tie

● Omission errors: no influence ties exist even though there is an

interdependence tie

THE ORGANIZATIONAL ANALYTICS E-BOOK 35

 ● Directional mismatch: the direction/symmetry of influence and

interdependence ties do not match: A depends on B, but B may influence A

Remember, we do not have to do this tie-by-tie search manually! We have computer

programs that can do this in the twinkling of an eye (with the twinkle having to last slightly

longer as network sizes grow…. but a twinkle nevertheless!). The detected errors can be

corrected at the individual level, and in fact, we might use algorithms to find the most

high-impact interventions. Patterns that show omission or commission errors that

systematically lie between two departments or geographies may highlight the need for

major re-grouping (e.g. divisional to functional structures).

FROM INDIVIDUALS TO GROUPS

Going straight from the dyadic level (“Ann and Bob”) to the organizational level makes a

network look a lot more complex. But we can simplify that picture thanks to what we call

“micro-structures”. This exploits a remarkable property of organizational structure: it looks

and works in roughly the same way when seen at any level in the hierarchy, or when we

replace individuals with entire organizational units. You might say this is because structure

is “fractal”- the patterns look similar at different levels of aggregation. For instance, the

basic pattern of a supervisor with multiple subordinates is seen both in the relationships

between a CEO and divisional heads, between the HQ of a holding company and multiple

affiliate companies, as well as between a foreman and factory workers. Peer-to-peer

relationships exist between teammates, divisions, and between companies in a strategic

THE ORGANIZATIONAL ANALYTICS E-BOOK 36

alliance. The clustering of informal friendship ties into clusters within clusters has the

same basic structure at any level of aggregation.5

These similarities across levels exist because, in organizations, the key interactions are

often between individuals who represent aggregates of other individuals (over whom

they may have authority) This allows us to exploit the fractal property- that at any level

in an organization, similar patterns of interaction will have similar dynamics – whether the

pattern occurs on the shop-floor or in the boardroom, between individuals who represent

other individuals (e.g. department heads) or just themselves (i.e. workers).

So how do we go about re-designing the structure using the ideas we have introduced so

far?

5
Ravasz and Barabasi, 2003; Puranam, 2018

THE ORGANIZATIONAL ANALYTICS E-BOOK 37

The first step is to identify the key actors, and the interactions among these, that you

think are at the heart of the design problem you are trying to solve. For instance, if

you are trying to correct a tendency for certain decisions to be slowed down because they

all pile up at one place in the organization(a.k.a. the bottleneck problem), identify the key

roles/units involved, and map out the interactions between them. Who needs decisions to

be approved? Who has the authority to approve which parts and kinds of decisions?

Where do the delays typically occur? Or perhaps, you see frequent breakdowns in

collaboration or missed opportunities across departments. Which units or roles should be

collaborating but are not? Again, you should identify the key roles/units involved and draw

how they interact. Or your issue may be the classic problem of the wrong persons in the

role, who doesn’t have the right skills and so creates bottlenecks and breakdowns (a.k.a.

The “bumbler” problem). Who is affected by the poor decisions of the bumbler? How are

they connected to each other? Yet again, you need to identify the key actors and the

pattern of interactions among them.

The second step is to distinguish clearly between two different types of interactions:

interdependence or influence. You can use workflow data and process mapping to

understand the interdependence structure. Organizational charts, surveys, and seating

plans can give us a picture of the influence structure. Increasingly we can also use digital

exhaust from emails, video calls, and collaborative chat tools (like Slack/Teams) for the

same purposes.

The third step is to apply the principle of fit (folk theorem): find ways to improve the

match between interdependence and influence. Remember this can be done by changing

influence structures, interdependence structures, or both.

THE ORGANIZATIONAL ANALYTICS E-BOOK 38

Being able to reduce the complexities of organization design problems to a picture with

the smallest number of actors and their interactions is a skill. But the good news is it is a

skill that you can build with practice. Here is a case to get you started:

Use Case#1: Re-designing to eliminate bottlenecks

Slow decision-making because of bottlenecks (i.e. places in the organization that function

like black holes when it comes to decisions) is a common signal for initiating an

organizational re-design. Slow decisions mean cost overruns, missed opportunities, and

ultimately competitive disadvantage. So how would we solve this problem solving Org2.0

techniques?

STEP 1: Perception

The first step, as always in analytics-based design is Perception. We gather data on the

existence and extent of the problem we are trying to solve. How do we define a

bottleneck? Who are the key individuals involved in the decision process? For the key

individuals, can we identify the interdependence and influence relationships among them?

How much must a decision be delayed for us to consider a bottleneck to exist? What is

the (opportunity) cost of the bottleneck? Is there really a bottleneck or are we just hearing

subordinates complain about superiors who rejected their proposals (perhaps on good

grounds)? If a proposal was not delayed but rejected, what is the evidence that this was

in fact a bad decision?

THE ORGANIZATIONAL ANALYTICS E-BOOK 39

One of the first things we are likely to learn from the perception phase is how widespread

the problem is: is it local or global? If it just involves one individual, then we may be

looking at a sorting problem (i.e. finding better talent for this role). We address sorting

problems in the next section. It could also just be a simple capacity problem, in the sense

that the individual has no bandwidth to deliver timely decisions (perhaps they have too

lager a span of control or do not have clarity of responsibility and delegation), in which

case we can just make the needed tweak to the structure for this localized problem.

But if the same pattern of bottlenecks, with the same or similar set of roles (e.g. managers

from sales, manufacturing, and somebody in the top leadership team) always show up as

the cast of characters, then we know we are possibly looking at something more global

than local, that might require a significant re-design. We proceed here on the assumption

that this is the case- that we are seeing a pattern of

delayed decision-making involving not necessarily the

same specific individuals, but perhaps the same sets of

roles in different parts of the organization.

Figure 1 illustrates the figures we might draw to

understand the bottleneck problem. Decisions that M4

needs from M1 seem to be the ones that get held up

forever. It could be that M1 is overburdened; has

insufficient delegation from M0 or is struggling to

manage the interdependence with M2. There are many possible solutions, which could

involve changing the interdependence structure (e.g. by changing incentives or task

THE ORGANIZATIONAL ANALYTICS E-BOOK 40

allocation) or the influence structure (to create a new reporting or coordinating

relationship). Which one will work will require knowledge of the context. These “micro-

structure” diagrams help with diagnosis and a thorough consideration of alternatives.

STEP 2: Prediction
If the problem is indeed widespread and recurring, this also makes it possible to gather

data at a large scale on the decisions that have been bottlenecked vs. those that have not.

This is where Prediction based analytics comes in useful. Can we forecast which decisions

are most likely to end up in the “black hole”? Are there certain attributes of those decisions

(e.g. the ones originating in a particular department, involving a particular set of issues, or

a level of financial implications) that make it more or less likely to be bottlenecked? With

prediction-based analytics, we can take one of two approaches: we might build the “early

warning system”, that tells us which decisions can be expected to be stuck, so that we can

spend our limited resources and energy just chasing those to make sure they don’t. In such

an approach, we would be agnostic to the “why” question-the reasons that make certain

decisions more or less likely to be stuck. Instead, we are pragmatic and focus on preparing

to deal with the symptoms before they arise, based on prediction models that can make

such forecasts. The use of “black box” (hard to interpret) algorithms may be perfectly OK

for such an application.

But an alternate approach is to go for “root cause detection”- to understand why certain

decisions seem more at risk of getting stuck (and then to do something about that). This

will usually limit the predictive analytics models we can use to those that are relatively

easier to interpret. But the good news is that such models are also precisely the ones that

THE ORGANIZATIONAL ANALYTICS E-BOOK 41

do not need massive amounts of data. In practice, this is the approach we think is most

likely to form the predictive analytics strategy for the bottleneck problem. Using such an

approach we may see that decisions that involve a particular kind of complexity (requires

consideration of the impact on 2 or more departments) that arise in the 3rd financial quarter

of the year are the ones that are at the highest risk of getting delayed (to take an instance).

With this insight, we might now want to think of how to solve the problem by changing

how the organization processes these decisions. But remember, predictive analytics only

generates indicators, not interventions! The results at this stage are correlational so we

still need to prototype before rolling out the changes.

STEP 3: Prototyping

Let’s suppose you believe that adding a coordination meeting between managers M1 and

M2 in Figure 1 would help speed up the decision-making process of the cases that have

an impact on both groups. How do we know this will solve the problem? Prototyping is

all about knowing whether your proposed intervention will cause an improvement. The

way to assess that is to think in terms of counterfactuals –what might happen if all else

stayed the same except your intervention was implemented. As we noted in Chapter 3,

there are a range of techniques one can use to prototype, ranging from scenario-based

discussions (e.g. getting a group of experienced managers to think through the last few

decisions that were bottlenecked, and doing the thought experiment- would your

proposed change have prevented the problem? ), to computer simulation (e.g. modeling

the flow of decisions and time taken for each on average to understand how frequent and

effective the coordination meetings would have to be to ease the bottlenecks), and

culminating perhaps in a field experiment (e.g. try the new system in some parts of the

THE ORGANIZATIONAL ANALYTICS E-BOOK 42

organization, create a control group where it is not implemented and compare the average

speed of decision making in both).

Which one you want to use will depend very much on the urgency of the problem and the

resources at your disposal. But it will also depend on the value of information- it is only

when both the potential upside from the intervention working is huge, as well as the

potential downside of it misfiring is huge that you need elaborate prototyping. If at least

one of these conditions does not hold, you don’t need to prototype- you would either not

bother implementing the intervention or implement it anyway without waiting. We also

reiterate that even something as simple as scenario analysis can already help to stop bad

interventions in their tracks. For instance, experienced managers might be able to point

out to you that M1 and M2 will be very unlikely to agree in their coordination meetings,

because their interests are not aligned, and their first common boss is several layers away.

In this case, we need to go back to the drawing board and think of other interventions.

Reflection question: Compare the approach you undertook the last time

you had to re-design something in your organization, with the Org2.0

approach we prescribed above. What are the barriers to using our approach

instead?

THE ORGANIZATIONAL ANALYTICS E-BOOK 43

CHAPTER 5
The second lever of organization design- Sorting

Overview

● Sorting refers to choices that determine who works in an organization. Two

organizations with the same structure but different people will perform quite

differently.

● Improving hiring and retention through data has become one of the most successful

applications of organizational analytics. However, sample selection bias remains a

key challenge- we often only have data on those we hired, and this can mislead us

about whom we should hire.

● Prototyping through A/B tests or even some degree of “hiring at random” can be a

powerful antidote to biased data.

Expanding our thinking about structure is useful, but we should be aware that structure is

only one of the approaches to solving organization design problems. Sorting is another.6

When using structure as a design lever, we take the composition of individuals in the

6
Some of the seminal work on non-structural aspects of organization design has been reviewed under the broad rubric of “culture” in
Schein 1985; O’Reilly and Chatman, 1996; Chatman and O’Reilly, 2016.

THE ORGANIZATIONAL ANALYTICS E-BOOK 44

organization as a given, but sorting does not. Sorting can occur in both directions- into an

organization as well as out of it. Further, it can occur within an organization in particular

projects. groups and teams.

As organizations are systems with boundaries, the entry (and exit) decision is a critical one.

Individual-level variations in goals, skills, and beliefs can be exploited to create better

matches between individual attributes and organizational requirements in terms of skills

and goals. A particular form of sorting is self-selection- in which individuals sort

themselves into an organization, and even within the organization, into particular tasks.

This has the advantage of generating intrinsic motivation through the freedom to choose

as well as the match between what is chosen and what the individual prefers.

Consider the following: Team A was formed by a group of individuals who self-selected

themselves into it because they were deeply motivated by the goals of the team and the

prospect of working with like-minded individuals. Team B was composed of individuals

who were assigned to it because they were available. In which team do you think

motivation and commitment levels will be higher? In which team will colleagues be more

often willing to make adjustments and resolve issues peer-to-peer? Which team do you

think is more at risk of getting trapped by groupthink? Which team is more likely to be

THE ORGANIZATIONAL ANALYTICS E-BOOK 45

cliquey and have lower diversity? As you can see, even if Team A and B had the same

objectives and the same formal structure, the way they work would be quite different.

As we often like pointing out employees are not randomly assigned to different

organizational structures. They are sorted in by themselves and of course by HR

professionals. No wonder, structures that appear to work with one set of individuals often

do poorly with others, and vice versa. In fact, this is one of the reasons why imitating

industry best practice is dubious – a structure that works for Company A given who works

there is unlikely to work for

Company B if their employees are

different.

The application of data analytics

to improve hiring, retention, and

staffing decisions first emerged as

a widely popular instance of

Org2.0 thinking. Traditionally, HR

professionals have relied on their

past experience and judgement to

make decisions on whom to hire

and whom to try to retain based

on considerations of fit. But increasingly, we can use data and algorithms to possibly do

better. To be clear all decisions are always based on data and algorithms, though

sometimes, the algorithms sit between our ears, and the data is called “experience”.

THE ORGANIZATIONAL ANALYTICS E-BOOK 46

The question simply is whether we can do demonstrably better by using electronic

databases and computer algorithms.

Use Case#2: Improving diversity in the talent pipeline

Sorting is pretty much the “baseball” of organizational analytics: it’s the area in which

analytics are the most developed, mainly because it’s easy to identify the unit of analysis

(people) and we don’t need to consider as many complex interdependencies in

understanding what makes people give their best as we would need in understanding

what makes a whole team or unit work best.

Many applications have been around for some time, but in recent times one issue has been

ruling them all: diversity. Organizations have come under great pressure to employ a

workforce more representative of the broader society. That means not only hiring a diverse

workforce but also making sure that diversity remains intact from the bottom to the top of

the organization by providing equal opportunities for promotion and development to all

populations. The lack of diversity can have different sources and managing opportunities

within an organization that already exhibits low diversity can be difficult.

Few organizations apply sophisticated methods to this problem. The typical case goes

along these lines: top-level managers recognize that the organization has a diversity

problem when an embarrassing statistic or a prominent case of discrimination makes the

press, and they identify areas of the organization that seem most guilty of excluding

minorities by simply looking at which unit present the least demographic diversity. They

might scramble to hire minorities in those areas (or promote minorities if the lack of

THE ORGANIZATIONAL ANALYTICS E-BOOK 47

diversity is only prominent at the top of the organization) without much consideration for

whether diversity issues might show up again in the future and whether this detracts from

the organization’s ability to function well (hiring minority managers on an emergency basis

is a lot less effective than preparing employees for managerial roles far in advance). Can

we do better than this? Applying Org2.0 thinking should help organizations scan for a lack

of inclusion in their workforce over time, identify specific minority employees at risk of

leaving the organization, understand why they might be leaving and identify policies that

can help hire, retain, and promote them.

Let’s look at each step in the organizational design analytics cycle.

STEP 1: Perception

A good analytics cycle starts with good data. People provide information about

themselves when they apply for a job; they go through an application process that leaves

(or should leave) digital traces at its different steps; and legal matters make it necessary

to gather data about when employees start their job, what that job is, and how well the

employee has been performing in that job. Keeping track of that data should be common

sense, but it is definitely worth emphasizing: many companies have trouble even knowing

how many employees work for them. There are always immediate concerns that seem

more urgent than creating a good data infrastructure, but in the long run, it definitely pays

off to gather workforce data systematically. Making the effort to link different databases

together is worth it, too: by merging data about applications, job interviews, work

performance and promotions, you will already start to formulate hypotheses about where

your company’s lack of diversity comes from. It might stem from not attracting a diverse

enough applicant pool, from weeding out minority applicants through a biased interview

THE ORGANIZATIONAL ANALYTICS E-BOOK 48

process, or from reducing progression opportunities for the employees you do hire

(because of a negative environment or a biased performance appraisal and promotion

process). Recognizing which stage is the most problematic (and how this may differ in

different units of your organization) is already valuable.

STEP 2: Prediction

A predictive approach to the talent pipeline is helpful both in order to identify specific

employees at risk of leaving and understanding the factors which put them at risk of doing

so. The easiest application would be an early warning system identifying employees at

risk of leaving the organization in the near future. The approach is simple: collect all the

data you may think of about your employees, generate as many variables as possible

based on that data (about employee demographics, performance, working environment),

apply different machine learning algorithms to the data you collected in the past to see

which algorithm makes better predictions, and finally make predictions about the future

using the algorithm which seems to perform best. The whole process shouldn’t take more

than a few days for a good analyst with clean data and will leave you with precise

predictions about which minority employees seem the most at risk to leave the

organization. You might want to at least talk to these employees to understand whether

and why they might be thinking of leaving, and whether there is something you can do

about it. In the same spirit, you could use your employee data to predict which managers

are the most likely to promote minority employees in the future.

The approach above is completely agnostic to why employees might be leaving: all you

care about is that they are likely to leave and that you know about it early enough to

inquire about them. But knowing why minority employees tend to leave can be very

THE ORGANIZATIONAL ANALYTICS E-BOOK 49

valuable as well and can generate ideas about how you may provide them with a better

working environment. For instance, you might find that minority employees are especially

likely to leave the organization when they rate their managers poorly. Perhaps providing

some training to your managers to sensitize them about diversity can be useful? You might

find that minority employees who went on voluntary team-building trips are much less

likely to quit in the next year. Maybe making these trips more widely possible could help?

Whatever policies you come up with are only a starting point, because machine learning

cannot tell you whether any variable actually causes employees to exit (maybe those

employees who went on team-building trips were simply more committed to the

company, to begin with!). But you can take your policy ideas to the next step: prototyping.

STEP 3: Prototyping

Prototyping is all about generating counterfactuals. By counterfactual, we mean “what

would have happened if X had not been true”. “X” can be anything, for instance, “going on

a team-building trip” in our example above. Maybe most employees who went on team-

building trips didn’t quit in the next year; but would they have quit if they hadn’t gone?

Chapter 3 described several sophisticated ways of generating counterfactuals. But simply

thinking about counterfactuals is already a great start. You may get a few managers to sit

together and re-examine past exit interviews, for example. For each minority employee

who left in the last year(s), does it look like organizing more regular feedback meetings

with their manager could have helped keep them committed to the company? A similar

approach can be applied to promotion processes: for each recent promotion case that did

not yield a minority candidate, could we have found equally qualified minority candidates

THE ORGANIZATIONAL ANALYTICS E-BOOK 50

by involving more managers in the search process? Scenario-based thinking is a very

powerful and inexpensive tool.

Combining insights from predictive analytics and scenario-based thinking, you should be

able to come up with a few ideas for policies that might help alleviate diversity issues. But

how do you know they will work? Do you need to roll them out in the whole organization

to figure that out? Often, you don’t! For instance, you may think that part of the issue comes

from minority candidates not applying for managerial positions. What if, instead of waiting

for applications, you sent targeted emails to candidates whose qualifications look like they

fit a position? You may start by doing this on a random basis for half of the positions you

seek to fill in the next year, and then roll out the policy for all promotion cases if results

convince you that this is good practice.

We close this chapter with a warning: it’s also important to be aware that hiring data has

some unique issues that can make its unthinking use quite dangerous. Here is the core

problem: typically, we observe the performance of people we hired, not those who we did

not. Therefore, using data on current employees to make guesses about whom to hire is

prone to what is known as the sample selection problem. Prototyping your interventions

de-risks this. Surprisingly, so does hiring a small portion of your workforce at random! This

article below explains why.

Article on biases in data and how randomness can help beat them
EMBEDDED LINK

THE ORGANIZATIONAL ANALYTICS E-BOOK 51

 Reflection question: Using data on employees to improve organizational

performance is not necessarily the same as improving employee welfare.

When are these objectives aligned? What ethical safeguards (besides of

course always complying with regulation) should you put in place when these are not

aligned?

THE ORGANIZATIONAL ANALYTICS E-BOOK 52

CHAPTER 6
The third lever of organization design- Sensemaking

Overview

● How people make sense of their organizational context – as reflected in things like

culture, identity, vision, and leadership- affects how the organization performs.

● Two organizations with identical structures and similar workforces may

nonetheless perform differently because of differences in these sensemaking

processes.

● Analytics can help in assessing the similarity of these processes across

organizations (e.g. assessing cultural fit in mergers) as well as the effectiveness of

interventions aimed to change these (e.g. forecasting if organizational change

projects will take off or sputter out).

Sensemaking refers to the process through which shared meaning is constructed within

an organization as its member create a mutual understanding of experiences and

situations.7 Framing an issue, describing a vision, and leadership are processes of one-to-

many sensemaking; a consensus formation is a many-to-many form of sensemaking. In

either case, neither the set of individuals nor their patterns of interaction need to change,

7
Weick, 1995.

THE ORGANIZATIONAL ANALYTICS E-BOOK 53

but there can be a significant change in the individual’s beliefs and goals over time, leading

to changes in behaviour. Org2.0 thinking relies on modeling such processes rigorously

using computational and network tools.

Culture analytics: the Org2.0 approach

How can we measure an organization’s

culture? Without question, culture is

important- it refers to the set of widely and

deeply held mental constructs, such as

values, beliefs, and norms. It shapes what

employees do when managers are not

looking. But how can we measure, let

alone shape it? One of the major advances

in recent years has been the application of

machine learning to text- in the form of

survey responses, chat, emails, or employee descriptions of their company (such as are

found in job posting/employer review websites). One can not only describe an

organization’s culture using such tools but also compare cultures across organizations.

That can be very useful in situations involving collaboration between organizations, such

as alliances, mergers, or even between divisions within a company. One can gauge the

sentiment and sense of engagement of employees, and isolate trending issues and

themes.

An article outlining how text analytics helps to understand organizational culture

EMBEDDED LINK

THE ORGANIZATIONAL ANALYTICS E-BOOK 54

 Use Case#3: Assessing cultural compatibility using Natural Language Processing

It’s a common argument among managers that culture is too complex to apply analytics

to. Sensemaking is about what people think; it’s about which values they hold and how

they go about communicating those values. Surely the only valid way to understand this

is to ask them about it? Not really. Taking the time to talk with your employees and

understand what’s on their minds is a great idea in theory, but you won’t be able to ask

each and every one of them about what they think. And they might not tell you to your

face what they might be willing to say anonymously. What if you could get insights into

what all of your employees are thinking? Recent developments in text analysis, combined

with new sources of information about employees’ perceptions of their companies, allow

you to generate those insights at scale.

STEP 1: Perception

One of the major advances in recent years has been the application of machine learning

to text. The idea behind one of the most widely used approaches known as ‘topic

modelling” is simple. When employees write about their company, they have certain

topics in mind which can be described using specific words (for instance, we typically do

not use the same words to describe work/life balance and compensation). Topic modelling

is about uncovering the topics that are important to writers by looking at the words they

use when describing them. You could collect this data from a variety of sources, from

internal surveys to emails/chat archives and even external job posting/employer review

websites. To tell you the truth, someone probably already has done that work for you:

THE ORGANIZATIONAL ANALYTICS E-BOOK 55

some prominent consultancies now routinely collect data (for free!) from websites such as

LinkedIn and Glassdoor to better understand their clients. Simply looking at Glassdoor

reviews written by your employees tells you a lot about what your employees care about

(for instance, is compensation one of the topics they mention the most? Or do they seem

to care more about doing challenging work?), how they think the company is doing on

those aspects (do they write about compensation using positive or negative terms?), and

how these values are distributed across the company (does everyone seem to care about

the same few things or is there wide variation? Are there systematic differences between

different units in the organization?). You will also be able to track the evolution of your

company’s culture by collecting this kind of data over time.

STEP 2: Prediction

Simply being able to observe what your employees care about is already very valuable; it

allows you to take the pulse of your organization and check whether its culture is evolving

in a direction that aligns with the culture its leaders are hoping to instill. But you can even

go further and use these insights to make predictions about the future of your organization.

Many aspects of an organization’s future are affected by its internal culture: whether it will

successfully integrate new employees, how successful it may be in collaborating with

other organizations, and even how well it may execute a strategy. We will focus on one

example which we think is quite promising: the management of mergers between

companies. Mergers are a very common occurrence in today’s business world, and it’s

widely acknowledged that they also very commonly fail. Culture is a major reason for this:

mergers look like great ideas on paper when they offer opportunities for synergies, but

THE ORGANIZATIONAL ANALYTICS E-BOOK 56

synergies typically require collaboration among the two firms’ employees. Cultural

differences make collaboration harder: people have a much easier time working with

people who share similar values and communicate in similar ways.

Can we predict cultural clashes in mergers? We can, using text analysis8. Glassdoor, for

example, provides reviews about thousands of companies. The odds are pretty high that

you can find reviews from the company you are thinking of acquiring! You can easily

compare the topics most prominent in two companies to better understand their

respective culture, which should help you predict potential conflict areas (for instance,

employees who care mostly about challenging work may experience some frustrations

working with colleagues who mostly care about work/life balance). Making those

predictions will require some judgement on your part (most tools won’t reliably tell you

that your merger has an X% probability of failing because of cultural issues). But text

analysis provides a very reliable starting point for understanding cultural differences:

when you spot significant differences through text analysis, you might invest then in

further information (surveys, interviews) to understand whether they reflect true cultural

differences between companies, and use common sense to predict their consequences. In

the extreme case, this might lead you to renounce a merger deal. In most cases, it will at

least provide you with valuable insights about which cultural differences to watch out for

during the post-merger integration phase.

STEP3: Prototyping

Websites that store employee reviews typically do so for a large number of companies.

So why would you only analyse the culture of your one preferred acquisition target? You

8
Marchetti (2020).

THE ORGANIZATIONAL ANALYTICS E-BOOK 57

can browse through as many possible targets as you would like to see how their

respective cultures differ from yours. At the very least, this will provide you with

benchmarks to understand just how different your preferred target’s culture is relative to

other potential targets. But it might also change your opinion about which target should

be the preferred one in the first place.

Change management in the Org2.0 way

Organizations change things all the time; at least they try to. Entering a new market,

changing an organizational structure, implementing new enterprise software…all are

examples of major organizational change which can be decided by top managers but

require buy-in across an entire organization to be truly successful. Many managers

consider change management as the most difficult aspect of their job. We all know there

is comfort in preserving the status quo, and most change efforts encounter stark resistance

from employees. So, how can you get buy-in? There is a lot of received wisdom on the

topic, typically around the importance of early buy-in by informal opinion leaders: the idea

is that once you convince your most central employees about the value of change, they

will go around and convince others until everyone believes in it. The reality is a lot more

complex, and it’s unlikely you will be able to predict it correctly without a systematic

approach.

THE ORGANIZATIONAL ANALYTICS E-BOOK 58

 Use Case#4: Forecasting change adoption using computer simulation

STEP 1: Perception

Broadly, change management depends on two things: what your employees believe

about the change you are proposing, and how they can influence each other to change

those beliefs. The first step is to generate a reasonably accurate picture of both.

Interviews and surveys can be useful for understanding your employees’ attitudes toward

change. Interviews are especially useful because people might not feel comfortable

sharing their honest opinion in surveys. It can be a good idea to ask your most important

managers what they think about a change initiative, and also ask them what they hear

others around them think about it. As for understanding how they can influence each

other, what you will need is to form an accurate picture of your organization’s social

network: who talks to whom? There are different ways to get data on this: depending on

what your employees are comfortable with, you may either ask them directly about who

they talk to (surveys) or use unobtrusive measures such as email traffic to better

understand who talks to whom. Simply looking at a combined picture of the social

network and the attitudes of employees towards change should already give you a few

hunches about your likelihood of success: if your most influential employees are against

the change, you are probably not standing on very solid ground. But we can do even better

if we could figure out what the most effective interventions might be to increase the

chances of successful diffusion of change.

THE ORGANIZATIONAL ANALYTICS E-BOOK 59

STEP 2: Prediction

The diffusion of opinions in an organization is a complex process. Because it depends on

collective dynamics among many different people who influence each other, it is quite

hard to predict what the outcome will be: will the change progressively convince more

and more members, or will it end up being rejected even by those who started out in favour

of it? To make things even worse, organizational change is a relatively rare event:

organizational transformations don’t happen every week, and even in tumultuous

organizations today’s change is rarely perfectly comparable to yesterday's. This means

that organizational analysts have little prior data to learn from when predicting the

success of the change. Does that mean we have to fall back only on intuition? No! Just

because we can’t perfectly predict the outcome of a complex process doesn’t mean we

shouldn’t try to predict it as well as we can. In the absence of past data, what we rely on

is theory: we try to get an accurate picture of the different pieces of the puzzle (in this case,

people’s opinions and the social network that links them) and try our best to think of how

these pieces interact together to produce the outcome we’re interested in predicting.

Luckily, we don’t need to rely exclusively on our brains to understand these complex

interactions. We can leverage the power of computers to simulate the process of

organizational change through what is called agent-based models. In these models, we

represent an entire organization (for instance, we can represent its different members,

what they believe, and who talks to whom) and set assumptions about how the

organization evolves over time. For instance, we might assume that every week, any two

people who know each other can talk to each other about the change initiative and one

might convince the other. By simulating the model over time, we can get predictions about

the likelihood of successful change implementation. These predictions answer the

THE ORGANIZATIONAL ANALYTICS E-BOOK 60

question “Given what we think we know about our organization, and given how we think

people influence each other over time, what is the likelihood that most of our employees

will buy into the change?”. The approach is actually quite similar to how the spread of

pandemics is modelled. If we know something about the structure of connections between

people and their infectiousness and susceptibility to infection, then we can make a pretty

good guess about how many victims we will have on our hands. The spread of ideas is

similar though not identical: it may require multiple exposures before an idea is accepted

and unlike resistance to disease, opposition to an idea can also be infectious.

STEP 3: Prototyping

The beauty of computer simulations is that they also allow us to experiment with different

scenarios in silico (see Chapter 3). Since the analyst fully controls the simulation

parameters, it is easy to create counterfactual scenarios where only one variable is

changed: “What is the likelihood that the change will diffuse throughout the organization

if we don’t do anything? What about if we convinced person X about the change? What

about if we convinced person X and named her ‘Change Initiative Officer’?” What’s even

more beautiful? You are running these experiments on the computer, practically for free.

This is similar to checking the sensitivity of your financial projections to different scenarios

using a spreadsheet. With good models, we can do the same for organizational processes.

An article outlining how computer simulation helps test our intuition about when
framing and vision may help collaboration
EMBEDDED LINK

THE ORGANIZATIONAL ANALYTICS E-BOOK 61

 Reflection question: How do you currently make decisions about assessing

culture or managing change? Are you intuitively following the Org2.0

approach but without data, or do you do something altogether different?

THE ORGANIZATIONAL ANALYTICS E-BOOK 62

CHAPTER 7
An Eye on the Future

This book is not about the latest organization designs – it is about a new approach to

designing organizations. By this point, we hope you have developed an appreciation for

organization design thinking the Org2.0 way- please don’t adopt a new organization

design without thinking carefully about whether it can work for your context. Data

and algorithms can help with this—that’s the basic message of this book.

In this book, we already gave several use cases for how to apply Org2.0 techniques. But

these are hardly the complete list. For instance, as the world of organizations adjusts to

the post-pandemic era, one thing is clear- remote collaboration will feature prominently.

Here is a sample of things that Org2.0 techniques may help with as companies prepare

to tackle this change:

● Identify how to modularize work as well as prioritize what needs to be done at the

office - to create optimal shifts for remote collaboration

● Identify who emerged as “super-connectors” to coordinate across teams virtually

● Diversity and inclusion- is it getting worse or better with remote collaboration?

● Text analysis to identify cultural fragmentation; and identify “cultural champions”

who strengthen and reinforce the organization's culture.

THE ORGANIZATIONAL ANALYTICS E-BOOK 63

How we design organizations in the age of algorithms is itself likely to change, as

developments in AI algorithms and innovations in organizing continue to unfold at

breakneck pace. In this edition of the book, our focus has been on algorithms as design

tools for organizations. Statistical analysis of data (including inductive inference from

machine learning, traditional deductive inference from experiments, and network analysis)

relies on algorithms for processing (possibly large quantities of) data. These applications

of algorithms are quite concrete—they involve fully specified instructions in the form of

computer programs. We also discussed the use of computational algorithms as

representations or models of organizations. One can for instance simulate how the division

of labour would look under self-selection instead of authority, to understand its relative

advantages and disadvantages. We are now developing a generative AI based

organization design co-pilot- think ChatGPT but with specialist expertise in organization

design. Look out for it!

Another approach that has been gathering steam is to use algorithms as embodied

solutions to organization design problems. For instance, Uber’s ride-sharing service uses

an algorithm to do task allocation (i.e. rides to drivers), GitLab’s “continuous integration”

software coordinates the coding efforts of remotely distributed programmers, and

ODesk’s algorithms compute compensation for tasks based on the nature of those tasks

and worker skills. Previously, each of these organization design solutions would have

been instantiated in the form of managers making and implementing decisions.

It is certainly possible that in the near future, we will see organizations in which at least

some of the workers are algorithms. This goes beyond the use by human agents of

THE ORGANIZATIONAL ANALYTICS E-BOOK 64

algorithms to make decisions—rather we are talking about intelligent agents with some

degree of decision rights. How such agents will work in conjunction with human agents,

and what this means for the design of organizations are the key questions. For instance,

how the division of labor may occur in mixed teams of humans and intelligent algorithms,

how they co-specialize, and how their joint productivity can be enhanced are all topics

within this theme.

That may well bring us to Org3.0!

Video and article on what the future of organization design may look like

EMBEDDED LINK

THE ORGANIZATIONAL ANALYTICS E-BOOK 65

References to academic literature cited in text

● Burton, R. M., & Obel, B. (1984). Designing Efficient Organizations: Modeling and

Experimentation.

● Centola, D., & Baronchelli, A. (2015). The spontaneous emergence of

conventions: An experimental study of cultural evolution. Proceedings of the

National Academy of Sciences of the United States of America, 112(7), 1989-

1994. doi:10.1073/pnas.1418838112

● Chatman, J. A., & O'Reilly, C. A. (2016). Paradigm lost: Reinvigorating the study

of organizational culture. Research in Organizational Behavior, 36, 199-224.

doi:10.1016/[Link].2016.11.004

● Clement, J., Gaba, V. and Puranam P. (2020) Organizational change as a complex

diffusion process, Working Paper.

● Marchetti, A (2020) Organizational Culture and Performance, Ph.D. Dissertation,

INSEAD. Mintzberg, H. (1979). The Structuring of Organizations: A Synthesis of the

Research: Prentice-Hall.

● Moreland, R. L., Argote, L., & Krishnan, R. (1996). Socially shared cognition at

work: Transactive memory and group performance. In What's Social about Social

Cognition? Research on Socially Shared Cognition in Small Groups (pp. 57-84):

Sage.

● Puranam, P., Alexy, O., & Reitzig, M. (2014). What's "new" about new forms of

organizing? Academy of Management Review, 39(2), 162-180.

doi:10.5465/amr.2011.0436

● Puranam, P. (2018). The microstructure of organizations: Oxford University Press.

THE ORGANIZATIONAL ANALYTICS E-BOOK 66

 ● Reagans, R., Miron-Spektor, E., & Argote, L. (2016). Knowledge utilization,

coordination, and team performance. Organization Science, 27(5), 1108-1124.

doi:10.1287/orsc.2016.1078

● Schein, E. H. (1985). Organizational Culture and Leadership: Wiley.

● Weick, K. E. (1995). Sensemaking in Organizations: Foundation for Organizational

Science-Sage Publication.

THE ORGANIZATIONAL ANALYTICS E-BOOK 67