The Decision Intelligence Handbook

What a decision really is

A business is precise about almost everything except the one thing it is made of. Ask a manager to show you last quarter’s revenue and they will produce it to the penny. Ask them to show you a decision, the actual moment of judgement that set that revenue in motion, and they will point instead at its consequence: the signed contract, the completed hire, the approved budget. The decision itself has no file. It happened in a meeting, was confirmed in a message, and is now gone, remembered only as the thing that led to the outcome everyone can see.

This handbook is about closing that gap, and it begins with a definition that sounds small and turns out to be structural. A decision is not an event that happens and passes. It is an object with parts, and it can be held. When a business treats its decisions as events, it keeps the outcomes and loses the reasoning, which is why it can be busy, growing, and quietly unable to say why it did any of it. When it treats them as objects, the reasoning becomes something it owns: searchable, challengeable, and, in time, improvable.

A decision is not an event that happens and is gone. It is an object with structure, and the moment you treat it as one, it stops evaporating.

To make that shift you have to understand what a decision is actually made of, and here the discipline rests on decision science rather than software fashion. The first and most important idea belongs to Herbert Simon, who won the Nobel Prize for it. Simon observed that real people, and real organisations, do not optimise. They cannot. Faced with a choice, they do not evaluate every option against every possible outcome, because they lack the information, the attention and the time to do so. Instead they satisfice: they set an aspiration, a threshold of good enough, and take the first option that meets it. Simon called this bounded rationality, and it is not a description of weak decision makers but of all of them, always.¹ The rational actor of the textbooks, who weighs everything and chooses perfectly, does not exist and never did.

This matters for a practical reason. If deciding were optimisation, you would not need to record the reasoning, because the reasoning would just be the correct answer, reproducible by anyone with the same data. But because deciding is satisficing under limits, the reasoning is the decision: which aspiration was set, what was known at the time, which options were even considered before one was judged good enough. Throw that away and you have thrown away the only thing that would let you tell, later, whether the judgement was sound or merely lucky.

The second idea comes from Daniel Kahneman, and it explains why the reasoning, left uncaptured, tends to go wrong in predictable ways. Kahneman distinguished two modes of thought: a fast, intuitive System 1 that produces an answer effortlessly, and a slow, deliberative System 2 that checks it and mostly does not bother. The trap he named is WYSIATI, what you see is all there is: the mind builds a coherent story out of whatever information happens to be in front of it, and treats that story as complete, regardless of what is missing.² A confident decision, in other words, is often just a decision whose gaps were invisible to the person making it. The remedy is not to be smarter; it is to make the evidence, and its gaps, explicit, so System 2 has something to check.

There is a practical consequence to calling a decision an object rather than an event, and it is the reason this handbook labours the distinction. An event is something you attend; when it is over, all that remains is memory, and memory is precisely where reasoning goes to be lost and quietly rewritten. An object is something you keep. It has a state you can return to, a structure you can inspect, and an identity that survives the people who made it. Treating a decision as an object means that months later, when the outcome is in, there is a specific thing to open and examine, not a disagreement about what everyone thought they had agreed. The apparatus that follows, the labelled evidence, the priced options, the recorded choice, the scored outcome, exists to give the object enough substance to be worth keeping.

Put Simon and Kahneman together and you have the case for the whole discipline. We decide under bounds, and we do not see the bounds we decide under. The answer is not a cleverer decider but a better structure for the decision: one that records what was known, labels how well it was known, names the options, and keeps the choice, so the invisible becomes inspectable. That structure is what the analyst firm Gartner has, in the last few years, given a name. Gartner defines decision intelligence as a practical discipline for engineering how decisions are made and how their outcomes are evaluated and improved through feedback, and it now tracks a whole category of software built to support it.³ Decision intelligence is not a technology; it is the practice of treating a decision as a first-class object, and the technology exists only to make that practice possible at scale.

The rest of this handbook is a working manual for that practice. It sets out the five parts a decision is made of, the hierarchy that labels its evidence, the constraint principle that prices its options, the way its confidence is composed rather than asserted, the record it keeps of the human choice, and the score it takes of the outcome. It is written for people businesses, the outsourcers, consultancies, managed service providers and staffing firms whose product is a chain of decisions, but the discipline is general. If you want the single-page version of the idea before the detail, our guide to the decision object is the place to start, and the full philosophy is set out in the pillar essay on enterprise decision intelligence.

Why organisations throw the reasoning away, and what it costs

Every organisation is two machines running at once: a machine for producing decisions and a machine for forgetting them. The first is loud and visible, all meetings and reviews and approvals. The second is silent, and far more efficient. By the following week the reasoning behind a choice has been overwritten by the next choice, and all that remains is the consequence, sitting in a system that was never designed to explain it.

The reason is not carelessness. It is that the reasoning has nowhere to live. The customer system records the deal, the applicant system records the hire, the finance system records the invoice. Each stores a state, a fact about the world after the decision. None stores the decision: what was known when it was made, which options were weighed, which constraint was expected to bind, who chose, and what they were betting would happen. The reasoning spans those systems and belongs to none of them, so it lands in the two places nothing is ever recovered from: the meeting, where it is spoken and then gone, and the inbox, where it is buried in a thread nobody will read again.

The cost of this is not a matter of opinion, and it is larger than most leaders assume. McKinsey’s research on organisational decision making found that only about 20% of executives believe their organisations are good at it, and, more strikingly, that 61% say most of the time they spend making decisions is used ineffectively.⁴ Scaled across a large company, McKinsey put the drag at roughly 530,000 days of managers’ time a year, on the order of USD 250m in salary spent on decision making that the decision makers themselves consider inefficient.⁴ That is not the cost of bad decisions. It is the cost of the process around them, before a single choice has even turned out well or badly.

The instinct, faced with those numbers, is to decide faster, to cut the meetings and force the call. But the evidence says speed and quality are not opposites. In a companion study, McKinsey found that respondents who described their decisions as fast were nearly twice as likely to also describe them as high quality, and the organisations that combined both were about twice as likely to report top-quartile financial returns from their recent decisions.⁵ Fast and good travel together; it is slow and bad that go hand in hand, and they are usually the symptom of a process with no structure, where every decision is argued from scratch because none of the previous ones were kept.

Bain’s decade-long work on the same question makes the financial stakes explicit. Across a survey of almost 800 companies it found a 95% correlation between how effectively a company makes and executes its decisions and its financial performance, and the underlying research reported that the strongest decision makers, the top quintile, delivered total shareholder returns nearly six percentage points higher than the rest.⁶ Six points of shareholder return is not a rounding error. It is the difference between a good business and an ordinary one, and it is attributable to something most organisations do not even measure: the quality of the decision itself.

A business that cannot say why it did what it did cannot get better at doing it.

Now place that inside a people business, where the finding stops being general and becomes existential. A software company can make a mediocre decision and let a brilliant product carry it; the product absorbs the mistake. A consultancy or an outsourcer has no product to hide behind. Its output is the decision, made visible as delivered work: which deals it took, who it hired, how it staffed the shift, what it agreed to at what price. If those judgements are wrong in aggregate, there is nothing to cushion the loss. The margin simply erodes, quietly, through decisions that were never written down and therefore never improved. We have argued at length that this is exactly why enterprise software should remember decisions, and it is the recurring theme of our sister report, the state of decision intelligence for people businesses. The organisation keeps the outcome and discards the thinking, which is the one asset that would let it decide better next time.

The anatomy of a governed decision

If a decision is an object, it has parts, and the discipline begins with naming them. A governed decision has five, and the reason to insist on all five is that each is the answer to a question a reasonable person, or a regulator, or your future self, will eventually ask. Miss one and you have a decision that cannot be defended, learned from, or trusted.

The objective. Every decision is a choice against something: a goal, a threshold, a deadline, a definition of good enough. Simon’s aspiration, made explicit. Most decisions go wrong here, not because the wrong option was chosen but because nobody agreed what was being optimised. Was the renewal decision about revenue, or margin, or the reference the account provides, or the risk of losing the team on it? Those point at different choices. Writing the objective down first, before the options, is the cheapest discipline in this handbook and the most often skipped.

The evidence. The facts the decision rests on, each carried with a label for how well it is known. This is the part that separates decision intelligence from a well-argued opinion, and it is important enough to have its own section below. The essential point is that evidence is never treated as uniform. A measured number and a hopeful guess may both appear in the same spreadsheet cell, but they do not carry the same weight, and a decision object refuses to let one impersonate the other.

The options. The live alternatives, including the one everybody assumes and the one nobody wants to say out loud. A decision with a single option is not a decision; it is a rationalisation. Each option is priced not against the date someone hopes for but against the requirement it is least able to meet, which is the worst-constraint principle, again its own section below. An option that commits to something the business cannot actually do is marked as such, and the thing it cannot do is named.

The choice. A person decides, and the object records who, when, and against what evidence, including any override of what the analysis recommended. This is not bureaucracy. It is the difference between a decision the organisation made and a decision that merely happened to it. When the human overrules the model, that disagreement is the most valuable thing on the record, because it is where the two sources of judgement can later be checked against reality.

The outcome. What actually happened, scored against what was expected. Without this the object is inert; with it, the object becomes a lesson. The outcome is the part almost every organisation omits, because by the time reality lands the decision is old news and everyone has moved on, which is precisely how a company manages to make the same mistake in slightly different clothes for years.

It helps to see the five parts move together through a single decision, offered as an illustration of the mechanism rather than a real account. Suppose a firm is deciding whether to bid for a large support contract. The objective, written first, is not simply to win it but to win it at a margin above the portfolio average without starving two existing accounts of their best people. The evidenceis assembled with its quality showing: the client’s volume forecast is stated, not measured, because it came from the buyer; the cost to serve is modelled from a similar contract; the availability of cleared staff is measured from the current roster; the competitor’s likely price is inferred from the last three tenders. The options are three, bid to win, bid to qualify at a higher price, or decline, each priced against its binding constraint, which for the aggressive option turns out to be a six-week security clearance rather than anything on the commercial side. The choice is made by a named partner who overrules the analysis, on the record, because they judge the client relationship worth a thinner first term, and that override is kept. And the outcome, whatever it proves to be, has a date in the diary on which it will be scored against every one of those expectations. Nothing in that description required software. All of it required treating the decision as an object with parts.

The uncomfortable truth is that no system a people business already owns holds these five parts, and this is not a failure of any one vendor. The average enterprise runs 897 applications, of which only 29% are integrated with one another,⁷ and yet you could connect every one of them perfectly and still not have the decision, because none of them was ever designed to hold it. They record state, the deal, the hire, the invoice, not the judgement that spanned them. The five-part anatomy is therefore not a report you can assemble from existing systems after the fact. It is a thing that has to be captured at the moment of deciding, or it is lost. Name the five parts and a decision stops being a story you tell afterwards and becomes a record you can inspect. Our guide to the decision object walks the same anatomy with worked examples.

The evidence hierarchy

The most dangerous number in any business is the one nobody has labelled. It sits in the model next to the measured figures, formatted identically, carrying the same visual authority, and it is a guess. When the decision is made, the guess is weighed exactly as heavily as the fact, because nothing on the page distinguishes them. This is not a data problem in the usual sense. The data may be fine. The problem is that the decision has no way of knowing which of its inputs to trust, and so it trusts them all equally, which means it trusts the weakest one too much.

The evidence hierarchy is the fix, and it is deceptively simple: label every fact underneath a decision by how it is known. Five levels are enough, and they run from the most trustworthy to the least.

The value of the hierarchy is not that it makes weak evidence strong. It cannot, and it does not try. Its value is that it makes weak evidence visible, so the decision can account for it. A choice built entirely on stated and inferred evidence is not forbidden; sometimes it is the only choice available. But it should know that it is, and it should carry its fragility on its face, so that when it is revisited nobody is surprised to learn the foundation was soft.

This is, in operational form, the antidote to Kahneman’s WYSIATI. The mind treats the information in front of it as the whole picture precisely because the missing information is invisible; there is no felt difference between a fact you have and a fact you have merely assumed you have.² The evidence hierarchy forces that difference into view. It gives System 2 something to check by making the gaps and the guesses show up as gaps and guesses rather than hiding inside a confident total. An unlabelled assumption is not a small sin. It is the origin of most expensive mistakes, because it is priced by everyone downstream as if it were true.

An assumption wearing the costume of a measurement is the most expensive item on any balance sheet, because it is priced as if it were true.

Applying the hierarchy in an hour

The hierarchy is not a system to install; it is a column to add. Take any spreadsheet on which a real decision currently rests and put one field beside every number: measured, modelled, inferred, stated or unmeasured. The exercise takes about an hour and is uncomfortable in a useful way, because most people discover that figures they were treating as bedrock are stated promises and inferred proxies, and that several of the numbers the decision most depends on are not there at all. The discomfort is the value. It is far cheaper to feel it before the decision than after the outcome.

Two rules keep the labelling honest. The first is that a gap is labelled, never left blank, because a blank cell is read by the eye as a zero or a non-issue when in truth it is the most important thing on the page: a number the decision needs and does not have. Naming it out loud, unmeasured, stops it disappearing into a false sense of completeness. The second is that a label may only be raised by doing the work, never by wishing: a stated figure becomes modelled when someone actually models it, not when a senior voice declares itself sure. The hierarchy earns its keep precisely because it cannot be talked upward.

The cost of unlabelled evidence shows up in two ways, one obvious and one insidious. The obvious one is error: decisions made on numbers that were never as solid as they looked. Gartner has long put the average cost of poor data quality, the downstream effect of numbers that do not hold, at around USD 12.9m a year per organisation.⁸ The insidious one is worse. When people cannot tell measured from assumed, they eventually stop trusting the numbers altogether, and the decision quietly reverts to instinct while the analytics become theatre. KPMG’s research found that only 35% of executives report a high level of trust in their own organisation’s analytics, even as 92% worry about the reputational damage of using them badly.¹¹ When two-thirds of leaders do not trust their own numbers, the expensive data infrastructure is decorative, and the real decision is being made somewhere it cannot be seen. Labelling evidence by its quality is the single cheapest thing a business can do to make its decisions trustworthy, and it can begin in a spreadsheet on Monday. The idea is developed further in our guide to the evidence hierarchy.

The worst-constraint principle

A plan is only as fast as the one thing that cannot go faster. This is obvious when you say it and forgotten the moment a decision is actually made, because the natural way to assess an option is to look at its average: on the whole, can we do this? On the whole is a lie the constraint will expose. The business does not experience the average of its constraints. It experiences the one that fails.

The clearest articulation of this belongs to Eliyahu Goldratt, whose 1984 business novel The Goal introduced the Theory of Constraints to a generation of operators. Goldratt’s claim was deceptively narrow and turned out to be general: the throughput of a system is governed by its single tightest constraint, its bottleneck, and improving anything other than that constraint does not improve the system at all.⁹ Speed up a machine that was never the limiting step and you produce more work-in-progress in front of the machine that is, and the factory ships no faster. The only improvement that counts is the one made at the constraint. Everything else is motion mistaken for progress.

Translate this from the factory floor to the decision and it becomes a pricing rule. An option’s real date, its real cost, its real risk, is set by the requirement it is least able to meet, not the one it meets most comfortably. A commitment to deliver a programme is not ready when the sales team is confident and the client is keen; it is ready on the latest date it clears every binding requirement at once, the hiring, the training, the security clearance, the systems access, the margin. If any one of those cannot be met by the promised date, the commitment cannot either, no matter how green the others are. A decision object prices the option against that worst constraint, and it names which constraint binds, so the conversation moves from a vague nervousness to a specific, addressable fact: this option is blocked, and the thing blocking it is the two-week security clearance nobody costed.

You do not experience the average of your constraints. You experience the one that binds, so that is the one a decision must be priced against.

The failure mode this prevents is the most common in a people business, and it is worth seeing concretely. Consider, as an illustration of the mechanism rather than a specific case, a contact-centre programme being judged healthy because its occupancy, the proportion of logged-in time agents spend handling contacts, is high. High occupancy looks like efficiency: the people are busy, the resource is sweated, the average utilisation is excellent. But occupancy sustained above the mid-eighties behaves like burnout, not efficiency; it is the constraint that binds on quality and attrition, and it is invisible in the average. Across an analysis of more than 160,000 calculations, the average maximum occupancy target that contact centres actually plan to, the point past which the human constraint starts to fail, sits around 83.3%.¹⁰ A staffing decision that optimises the average and ignores that ceiling books a saving today and pays for it in six months, in agents who leave and quality that slips, and the decision that caused it will be long forgotten by the time the consequence arrives.

Computing the date an option actually clears

The principle becomes operational the moment you stop averaging and start taking the maximum. Every option in a real commitment depends on several things coming true, each with its own earliest date: the hires cleared, the training completed, the systems provisioned, the contract signed, the funding released. The temptation is to call the option ready when most of those are ready, or to quote the date the business hopes for. The correct date is the latest of them, the point at which the last binding dependency clears, because until that moment the commitment cannot be met however green everything else looks. An option is not the average of its readiness; it is the date of its slowest binding part.

This has a bracing effect on planning, because it replaces a vague optimism with a specific bottleneck. Instead of “we think we can probably do this by spring,” the decision reads “this clears on the fourteenth of April, and the thing setting that date is the security clearance.” If you want it sooner, that is the only lever worth pulling; every other improvement, a faster contract, an earlier training slot, is motion that does not move the date, exactly as Goldratt warned. Naming the binding constraint does not merely make the estimate honest; it tells you the one place effort will actually change the answer.

The worst-constraint principle is, in the end, a discipline of honesty about where the real limit lies. It resists the comfortable habit of judging a decision by its strongest feature, and insists on judging it by its weakest binding one, which is the only feature that determines whether it works. The full treatment, including how a system computes the binding date across several constraints at once, is in our guide to why the worst constraint decides. An option is only as good as the single requirement it is least able to meet, and a decision that does not know its own binding constraint does not know what it is really choosing.

Composing confidence, not asserting it

Confidence is the most counterfeited currency in a meeting. It is easy to produce, it spends well, and most of it is forged. Someone senior says they are confident, the number goes into the plan carrying their authority, and everyone downstream treats it as though it were derived from something. It was not. It was asserted. And an asserted confidence and a composed one look identical right up until reality tests them, at which point they turn out to have been completely different things.

A composed confidence is built, not declared. It starts from the evidence, each fact weighted by its place in the hierarchy, and it produces a number that means something because you can trace it back to its inputs. If a decision rests mostly on measured evidence, its confidence is high and earned. If it rests on stated promises and inferred proxies, its confidence is low, and it should say so, out loud, on the object, no matter how senior the person who wishes it were higher. The system that composes confidence this way can do something an assertion never can: it can name the dimension dragging the number down. Not merely we are 60% confident, but we are 60% confident, and the reason is that the single largest input, the delivery cost, is unmeasured. That is a confidence you can act on, because it tells you what to go and find out.

The alternative, which is the industry norm, is confidence theatre. A figure is produced, it feels authoritative, and nobody can say where it came from. The evidence that this is the norm is in the trust numbers: KPMG found that only about 35% of executives have a high level of trust in their own organisation’s analytics.¹¹ Read that carefully. It does not say the analytics are wrong. It says the people who own them cannot tell whether to believe them, which is what happens when confidence is asserted rather than composed. A number you cannot trace is a number you cannot trust, and a number you cannot trust is one you will quietly override with your gut while pretending you did not.

Confidence you can trust is composed from evidence you can see, not declared by whoever is most senior in the room.

Composing a confidence figure is not mysterious, and it does not require pretending to a precision the inputs cannot support. It means starting from the evidence, weighting each fact by its level in the hierarchy, and letting the weakest load-bearing inputs pull the number down rather than rounding them away. A decision resting on a measured roster and a modelled cost can carry real confidence; the same decision resting on a stated volume and an inferred competitor price cannot, and the composed figure should show that difference plainly. The output is not a false decimal but an honest band, together with the one sentence that matters most: the largest reason the number is not higher. A confidence that can point at its own weakest input is doing its job; a confidence that is only a feeling with a percentage attached is not.

There is a subtler benefit to composing confidence, which is that it makes disagreement productive. When two people disagree about an asserted number, they are really disagreeing about whose authority wins, and that argument has no resolution except rank. When two people disagree about a composed number, they are disagreeing about an input, and that argument can be settled by going and measuring the input. The composition turns a clash of confidences into a question of evidence, which is the only kind of disagreement that a business can actually resolve rather than merely survive. This is why we treat business confidence as a derived quantity with a stated method rather than a mood, an approach set out in our guide to business confidence. A confidence figure is only worth having if it can name the reason it is not higher.

Recording the human choice, and the override

A system can recommend. Only a person can be accountable. This is not a limitation to be engineered away as the models improve; it is the point. A decision is a commitment the organisation makes about an uncertain future, and commitments are made by people who can be held to them, not by analyses that cannot. The governed decision therefore always records a human choice: who decided, when, on what evidence, and, crucially, whether they followed the recommendation or overruled it.

The reason to insist on the human choice runs straight back to Simon and Kahneman. Simon told us that the person deciding is bounded, working with incomplete information under real constraints of time and attention;¹ Kahneman told us that same person is subject to predictable biases they cannot feel from the inside.² It would be tempting to conclude that the human is the weak link and should be removed. But the model has no less of a problem: it is bounded by its training data and biased by whatever it was optimised on, and, unlike the human, it cannot be held responsible for what it recommends. The answer is not to choose between a fallible human and a fallible model. It is to put them on the record together, so that when they agree you know it, and when they disagree you keep the disagreement.

The override is where this earns its keep. When a person overrules a recommendation, the instinct in most organisations is to treat it as friction, a manager ignoring the analysis, and to lose it. That is exactly backwards. The override is the single most valuable event a decision system can capture, because it is a labelled bet: the human is saying, on the record, that they know something the model does not, or that they distrust an input the model trusted. When the outcome lands, that bet can be scored. If the overrides consistently beat the model, the model is missing something and you can go find it. If the model consistently beats the overrides, the organisation is carrying an expensive habit of second- guessing its own evidence, and now it can see that too. Either way, the learning is only possible because the disagreement was written down instead of smoothed over.

An override is not a failure of the system. It is the most valuable thing the system can record, because it is where human judgement and the model disagree on the record.

It is worth heading off a fear that recording the human choice tends to provoke, which is that it is really about assigning blame. It is not, and a decision culture that treats the record as a hunt for someone to punish will destroy the very honesty the record depends on. The point of keeping who chose and why is not to have a name to hold responsible when an outcome disappoints; it is that a good decision can lead to a bad outcome and a bad decision to a good one, and you cannot tell which you had unless you kept what was known and expected at the time. Recorded well, the choice protects the decider who reasoned soundly and was unlucky, and exposes only the reasoning that is consistently unsound. The record is a memory, not a charge sheet.

Recording the choice also happens to be what accountability increasingly requires. A decision about a person, who was hired, who was let go, whose shift was cut, who was flagged for performance, is exactly the kind of decision that regulators and courts now expect an organisation to be able to reconstruct: what evidence it used, which options it weighed, who made the call, and whether a human was genuinely in the loop or merely rubber-stamping a machine. An organisation that records its choices as a matter of course is not doing this to satisfy a regulator. It is doing it because it is how you learn. That the same record satisfies the auditor is a convenient side effect of a practice you needed anyway. Automate the analysis if you can, but never automate away the recorded human choice, because the choice is the accountable act and the override is the lesson.

Scoring outcomes and closing the loop

A decision no one scores is a decision no one learns from. This is the part of the discipline that organisations find hardest, not because it is technically difficult but because it requires going back. By the time an outcome is knowable, the decision that produced it is months old, the people have moved on, and the appetite to reopen a settled matter is low. So the outcome is never compared to the expectation, the bet is never settled, and the organisation makes its next decision no wiser than its last. It has a memory for consequences and amnesia for causes.

The clearest way to understand why this matters is John Boyd’s OODA loop. Boyd, a military strategist, described decision making in conflict as a cycle of four steps: observe, orient, decide, act. His insight, easy to state and rarely honoured, was that the value is in the loop: acting is not the end of the cycle but the start of the next one, because the result of the action becomes the next thing observed.¹² The whole point of Boyd’s model is the feedback, the arrow from act back to observe, which is what lets a decider adapt faster than the situation changes. Cut that arrow and you do not have a slower loop; you have no loop at all. You have an arc that ends the moment you act, and an arc cannot learn.

Most organisations run an arc and call it a loop. They observe, orient, decide and act, and then they never observe the result of that action against the decision that caused it. The contract renews and nobody checks, eighteen months later, whether the account performed as the renewal assumed. The hire is made and nobody asks, a year on, whether the person turned out as the decision predicted. The staffing level is set and nobody scores it against the attrition it was betting on. Each of these is an act without the return arrow, a decision that produced a consequence that was never fed back into the next decision of the same shape. The information that would make the organisation better is generated, reality produces it for free, and then it is discarded.

Boyd’s insight was that the loop is the point. A decision process that never scores its own outcome is an arc, not a loop, and an arc cannot learn.

A score is more than a verdict of right or wrong, and its richness is what makes it worth keeping. A properly recorded outcome holds four things next to each other: what was expected when the decision was made, what the analysis recommended, what the human actually chose, and what in the end occurred. Laid side by side, those four answer questions no single one can. Did reality match the expectation, meaning the decision was well founded even if the outcome stung? Did the human override beat the model, or the model the override? Was a good result the product of a good decision or merely good luck, a distinction organisations are forever tempted to blur and must learn to hold? Across many decisions these comparisons become calibration: the organisation learns not just whether it was right but whether its confidence was warranted, which is the whole difference between being lucky and being good.

Scoring an outcome is mechanically simple: when reality lands, compare it to what was recommended and to what was chosen, and record the difference. But the effect compounds in a way that is hard to overstate. A single scored decision is a tidy piece of hygiene. A thousand scored decisions, each comparing a bet to its result, is an asset no competitor can buy, because it is built entirely from your own history. If commitments with a certain profile consistently deliver below forecast, the next recommendation with that profile can say so, in advance, with your own evidence behind it. This is the loop that turns a pile of decisions into something that gets measurably better over time, the arc bent back into a circle. The destination is not better dashboards. It is an organisation that decides better this quarter than it did last, because it kept the score. The scored outcome is the institutional memory most organisations throw away the moment reality arrives, and it is the only part of the decision that teaches the next one.

The Decision Room as the venue

A discipline needs a place to happen. You can hold the five parts of a decision in your head, or scatter them across a spreadsheet, an email chain and a meeting, but a practice that lives in fragments will die in fragments. What decision intelligence needs is a venue: a single place where a decision is assembled, argued, chosen and, later, scored, and where it stays, as an object, so it can be revisited when reality arrives. We call that venue a Decision Room.

A Decision Room is not a meeting, though a meeting may happen in it, and it is not a dashboard, though it shows you numbers. It is the place a decision goes to become an object. Into it come the parts this handbook has described: the objective, stated plainly; the evidence, each fact carrying its label from the hierarchy; the options, each priced against its binding constraint; a composed confidence that names the dimension holding it back; and, when the moment comes, the human choice, recorded with whatever it overruled. And then, unlike every meeting that ever dispersed with its reasoning, the room keeps all of it, waiting for the outcome that will be scored against it. The decision does not evaporate when the participants leave, because the room is its home, not their memory.

The reason a people business in particular needs such a venue is that its most important decisions are the ones that cross functions, and functions do not naturally reason together. Whether to take a deal depends on delivery and hiring, which live in other systems and other teams. Whether to renew an account depends on delivered value and true cost, which the service view cannot see. Whether to expand depends on the pipeline and the labour market at once. These are not questions a single department can answer well, because the evidence they need is spread across departments that each hold a slice. The Decision Room is where those slices meet: a shared surface on which each function publishes its facts, with their quality labelled, so the others can reason with them. That capacity, functions reasoning together on shared evidence, is what we call cross-module intelligence, and the room is where it becomes concrete.

A Decision Room also has a life, which is what separates it from a document. It opens when a decision is framed and its objective stated. It stays open while evidence is gathered and options are priced, and it is where the deliberation happens rather than in a thread nobody can later reconstruct. It closes, provisionally, when a human makes the choice on the record. And then, crucially, it does not vanish; it goes dormant, waiting, because the decision is not finished until its outcome is known. When reality arrives the room reopens to be scored, and the comparison it then holds, expectation against result, becomes the seed of the next decision of the same shape. A meeting ends when the people leave. A Decision Room ends only when the thing it decided has finished happening.

It is worth being clear about what the room is not, because the failure to draw this line is where a lot of good intentions go to die. The room is not a new place to have the same unstructured argument. If the evidence is not labelled, the constraints are not named, the confidence is not composed and the outcome is never scored, then you have merely moved the old meeting into new software and kept all its bad habits. The venue is necessary but not sufficient; it is the discipline, practised inside the venue, that does the work. What the room adds is permanence and structure: a guarantee that the parts are present and that the object survives the conversation that produced it. The full description of the venue, and how it differs from the tools it replaces, is in our guide to the Decision Room. The Decision Room is the place a decision goes to become an object, and to stay one, so that when reality arrives there is something to compare it against.

Common failure modes

Decision making fails in a small number of predictable ways, and the useful thing about the anatomy in this handbook is that each classic failure turns out to be a missing part of the object. Once you can name the missing part, the failure stops being a personality problem and becomes a fixable gap. Four failure modes account for most of the damage.

The HiPPO. The highest-paid person’s opinion, substituted for evidence. The most senior person in the room states a view, and because seniority reads as certainty, the view becomes the decision without ever being tested against the facts. This is confidence asserted rather than composed, and it is corrosive not because senior people are usually wrong but because it teaches everyone else to stop bringing evidence, since evidence loses to rank anyway. The remedy is structural: compose the confidence from the evidence, on the record, so that the senior view has to enter as one input among several rather than as the verdict. A HiPPO that has to say which measured fact it is overruling, and why, is a HiPPO doing its job. A HiPPO that simply wins is a missing evidence step.

Analysis paralysis. The opposite failure, and the one Simon diagnosed most precisely. An organisation that tries to optimise, to evaluate every option against every outcome before choosing, never chooses, because the world supplies new information faster than the analysis can consume it. Simon’s answer was not to analyse harder but to satisfice: set an explicit aspiration, a threshold of good enough, and take the first option that clears it.¹ Analysis paralysis is what happens when the objective was never made explicit, so there is no threshold to stop the search. Name the aspiration first and the search has a finish line. The missing part here is the objective.

Confidence theatre. A number that feels authoritative and cannot be traced. Everyone behaves as though it were solid because it is presented as solid, and the fact that only about a third of executives trust their own analytics¹¹ is the aggregate symptom of a thousand such numbers. The missing part is evidence quality: a figure that cannot say what it is made of is a figure that should not be trusted, and the evidence hierarchy is what forces it to declare its ingredients.

The dashboard mistaken for a decision. The most modern failure, and the most seductive. A dashboard reports the state of the world beautifully and stops exactly at the edge of the choice, yet its polish invites the reader to believe a decision has been made simply because the situation has been described. Worse is its automated cousin: a recommendation produced by a model with no evidence exposed beneath it, which is confident, untraceable and impossible to learn from, an arc with no return arrow. The missing parts here are the options, the recorded choice and the scored outcome, everything that comes after the description. A picture of the present is not a decision about the future, however good the picture.

Escalation of commitment. The quiet, expensive habit of sending good resources after a decision that is no longer working, because reversing it would mean admitting it was wrong. A contract is kept because it was hard to win; a hire is defended because letting them go would indict the manager who made the call; a programme runs on because too much has already been spent to stop. Sunk cost is a bias every operator can name and commits anyway, and it thrives in the dark: when the original decision was never recorded with an expectation to test against, there is no moment that says, on the evidence, this is not going as we bet, and no clean place to reverse. The missing parts are the recorded expectation and the scored outcome. A decision that stated what it was betting, and put a date on checking the bet, can be unwound on evidence rather than defended out of pride.

Every classic failure of decision making is a missing part of the object: no objective, no evidence quality, no constraint, no recorded choice, or no scored outcome.

The reason this framing helps is that it makes the failures teachable and preventable rather than merely lamentable. You do not fix a HiPPO by asking senior people to be humbler, or analysis paralysis by asking anxious people to be braver. You fix them by ensuring the decision object has all its parts, because a decision with a stated objective cannot spiral, a decision with composed confidence cannot be won by rank alone, and a decision with a scored outcome cannot pretend forever that it worked. The failures are not flaws of character; they are gaps in the structure, and a complete decision object is what closes them.

A maturity path, and how to start on Monday

No organisation adopts a discipline in a single step, and it is unrealistic to expect a business running on meetings and spreadsheets to wake up governing its decisions as objects. It is more useful to see the journey as a path with stages, so you can locate where you actually are and see what the next move looks like. The path below is our own framework, a synthesis of the patterns we observe rather than a measured benchmark; treat it as a lens for honest self-assessment, not a league table.

1. Instinctive. Decisions live in meetings and inboxes. The reasoning is verbal and gone by the following week. The organisation cannot say why it did what it did, and each decision starts from scratch. Most of the cost McKinsey measured lives here.
2. Reported. Dashboards describe the past well, and each function can produce competent numbers about itself. This feels like progress, and it is, but it stops at the edge of the choice: the reports tell you what happened and leave the decision exactly where it was.
3. Connected. The functions share facts across a common surface, so sell, hire, deliver and pay can at last be seen together. The cross-functional questions become answerable. What is still missing is governance of the choice itself: evidence quality, recorded overrides, scored outcomes.
4. Governed. The important decisions are made as objects. Evidence is labelled, options are priced against the binding constraint, confidence is composed, and the human choice is recorded with its overrides. The decision can now be defended, audited and challenged. What remains is to close the loop at scale.
5. Learning. Outcomes are scored against recommendations as a matter of routine, and the next decision of the same shape is better informed than the last. Nothing structural is missing now; the advantage simply compounds, quarter after quarter, from a history no competitor can copy.

Most businesses sit between Reported and Connected. They have invested heavily in dashboards and are beginning to link their systems, but very few have reached the point where a decision is captured as a governed object, and fewer still have closed the loop so that outcomes teach the next choice. That gap, between reporting the past and governing the future, is where decision quality is won or lost, and it is reachable from where you already are.

The encouraging part is that the practice does not wait for the platform. You can begin on Monday, by hand, on your very next real decision, with nothing but the discipline itself. Five moves are enough to start:

• Write down what you are actually deciding, and against what objective. Make the aspiration explicit before you look at the options.
• Label each fact underneath the decision as measured, modelled, inferred, stated or unmeasured. Notice how much of the foundation is softer than it looked.
• Name the single constraint that has to hold for the choice to work, and check it directly rather than trusting the comfortable average.
• Record who chose and why, including anything they overruled and the reason they gave. Keep the disagreement, do not smooth it away.
• Put a date in the diary to score the outcome against the expectation. This one habit, the return arrow, is what turns your arc into a loop.

None of these requires software, and that is the point. What software adds, once the habit is real, is scale and durability: the ability to do for a thousand decisions a month what you can do by hand for one, and to keep the objects safely rather than trusting them to a spreadsheet that will be overwritten. But the habit comes first, and the habit is available to any organisation willing to stop throwing its reasoning away. The compounding this eventually produces, an organisation that gets measurably better at deciding, is what we mean by enterprise intelligence, and it starts with a single decision recorded as an object. The habit comes before the software: you can begin recording decisions as objects, by hand, on Monday, and the tooling only makes it sustainable at scale.

Methodology & a note on honesty

A handbook about treating evidence with honesty has no business making dishonest claims of its own, so a word on how it was assembled and where it holds back. Every quantitative figure in these pages is drawn from a named, dated, public source, and each is listed in the References. The conceptual scaffolding is attributed to the thinkers who built it: Herbert Simon for bounded rationality and satisficing,¹ Daniel Kahneman for the two-system model and WYSIATI,² Eliyahu Goldratt for the Theory of Constraints,⁹ John Boyd for the OODA loop,¹² and Gartner for naming decision intelligence as a category.³ Where a claim rests on one of their ideas, the citation points at a real, resolvable source for that idea rather than at a paraphrase of it.

One number deserves an explicit caveat rather than a confident restatement. Bain’s finding on decision effectiveness is widely quoted, and Bain itself has expressed it in more than one way across its publications: the survey page we cite states a 95% correlation between decision effectiveness and top-tier financial results across almost 800 companies, while elsewhere Bain describes the same research programme as showing a correlation at a minimum 95% confidence level over a larger sample.⁶The related figure, that top-quintile decision makers delivered total shareholder returns nearly six percentage points higher, comes from the same Bain Decision Insights work. We report the correlation and the shareholder-return differential together because they are Bain’s, but we flag that the precise statistical framing varies between Bain’s own pages, and a careful reader should treat the headline as directional evidence that decision quality and financial performance move together, not as a single coefficient.

A handbook about traceable decisions has no business making untraceable claims of its own.

We have also deliberately left out a statistic that would have suited the argument, because we could not trace it to a sound primary source. The claim that the average adult makes some 35,000 decisions a day is repeated across business writing and is often attached to a university name, but it traces back to informal estimates rather than a published study, and it is the kind of tidy, unverifiable number this handbook exists to argue against. We mention it only to explain its absence. The frequently-quoted figures on the cost of replacing a departed employee are similarly wide-ranging and method-dependent, and we have not leaned on them as load-bearing claims either.

Two elements here are explicitly our own and are labelled as such in the text. The maturity path is a framework, a synthesis of observed patterns rather than a measured benchmark, offered as a lens for self-assessment. The worked examples, the renewal that glows green while it rots and the contact-centre programme judged by its average occupancy, are illustrations of a mechanism, not reports of a specific customer. We hold ourselves to the same evidence hierarchy we recommend: measured facts are cited, modelled or illustrative content is named as such, and no assumption is dressed up as a measurement.

This is the first edition. The conceptual core, what a decision is made of and how to practise treating it as an object, is stable and unlikely to move. The surrounding figures, particularly those touching technology adoption and the economics of the industries this handbook is written for, will change, and we will revise the document as they do. If you find a claim here you believe is wrong, or a source that has been superseded, we want to know. A reference work earns its standing by being corrected, not by being defended.

References

Every figure in this report is sourced below. Third-party links open in a new tab. Where a figure is modelled rather than measured, or where a widely-quoted statistic could not be traced to its primary source, the report says so in the text.

1. Bounded Rationality (Stanford Encyclopedia of Philosophy) · Stanford Encyclopedia of Philosophy, 2018. Attributes “bounded rationality” and “satisficing” to Herbert Simon, citing his 1955 paper “A Behavioral Model of Rational Choice”; real decision-makers satisfice under limits of information, attention and computation rather than optimise across all options.
2. Thinking, Fast and Slow (System 1 / System 2; WYSIATI) · Daniel Kahneman, Farrar, Straus and Giroux, 2011. Kahneman’s account of fast, intuitive (System 1) and slow, deliberative (System 2) thinking, and of WYSIATI (“what you see is all there is”), the tendency to build a coherent story from the information at hand and treat it as the whole picture.
3. Decision Intelligence, Gartner IT Glossary definition · Gartner, 2025. Gartner defines decision intelligence as a practical discipline that advances decision making by explicitly engineering how decisions are made and how their outcomes are evaluated, managed and improved through feedback; Gartner also maintains a Market Guide and Magic Quadrant for Decision Intelligence Platforms.
4. Decision making in the age of urgency · McKinsey & Company, 2019. Only 20% of organisations say they excel at decision making; 61% say most decision-making time is used ineffectively; the inefficiency costs a typical Fortune 500 about 530,000 manager-days (roughly USD 250m) a year.
5. Three keys to faster, better decisions · McKinsey & Company, 2019. Respondents who called their decisions fast were nearly twice as likely to also call them high quality; organisations combining speed and quality were about twice as likely to report top-quartile financial returns.
6. Score your organization to improve decision effectiveness · Bain & Company, 2011. Across a survey of almost 800 companies, Bain reports a 95% correlation between decision effectiveness and top-tier financial results; the underlying Decide & Deliver research found top-quintile decision-makers delivered total shareholder returns nearly 6 percentage points higher than the rest.
7. Connectivity Benchmark Report 2025 · Salesforce MuleSoft, 2025. The average enterprise uses 897 applications, of which only 29% are integrated with one another.
8. How to Improve Your Data Quality · Gartner, 2021. Poor data quality costs organisations an average of USD 12.9m a year (Gartner, attributed to Melody Chien).
9. The Goal and the Theory of Constraints · Eliyahu M. Goldratt (Theory of Constraints Institute), 1984. Goldratt’s 1984 business novel introduced the Theory of Constraints: a system’s throughput is governed by its single tightest constraint, and improving anything other than that constraint does not improve the system.
10. What Is the Ideal Occupancy Rate for a Contact Centre? · Call Centre Helper, 2023. An analysis of more than 160,000 calculations put the average maximum occupancy target at 83.3%.
11. Guardians of Trust (Forrester Consulting study for KPMG International) · KPMG, 2018. 92% of executives worry about the reputational risk of inappropriate analytics use; only 35% report a high level of trust in their own organisation’s analytics.
12. The Essence of Winning and Losing (the OODA loop) · John R. Boyd, 1996. Boyd’s briefing presenting the OODA loop (observe, orient, decide, act) as a continuous cycle in which acting feeds back into observing; archived public-domain edition.