From Records to Decisions

A merchant in a trading city, some five hundred years ago, keeps a book. In it, every transaction appears twice: once as what was given, once as what was received. Cloth out, coin in. Coin out, timber in. The two columns must agree, and when they do not, the merchant knows, before the ship has sailed or the debt has soured, that something is off in the accounts. This is double-entry bookkeeping, and it is one of the most durable pieces of business software ever devised, though for most of its life it ran on paper, ink, and the discipline of clerks. It did one thing with a rigour nothing before it had managed. It recorded what happened, completely and in balance, so that the merchant could trust the book instead of his memory.

Everything that followed in the long history of business software is a variation on that merchant's ambition: to hold in a system something too important to leave to memory. We built machines to do it faster, then networks to share it wider, then screens to show it better, then models to guess what came next. But look closely at what the ledger did, and did not, contain. It recorded that a bolt of cloth was sold at a certain price. It did not record why the merchant chose that price, what else he weighed, who in the counting house argued for more, or whether, a year on, the choice had looked wise or foolish. The decision left no trace. Only its result did.

Five centuries and a great deal of engineering later, we have replaced the book with a thousand systems, and we have still not fixed that.

The turn

Part II made the case: the modern software estate records everything except the decision, forgets what it learns, and dresses the fragments in dashboards that look joined-up and are not.

Is Decision Intelligence a real next thing, or is it the latest coat of paint on a category the industry repaints every few years? Every era of enterprise software has arrived wrapped in the same promise, that this time we will finally see the whole business clearly, and every era has left the promise part-kept. Why should this be different?

Enterprise software did not appear all at once. It grew in layers, each one adding a capability the layer beneath it lacked, each one solving the very problem its predecessor created. If Decision Intelligence is real, it will be visible as the next such layer: not a product you bolt on, but a shift in what a business considers worth managing in the first place.

The era of recording what happened

The first thing a business ever asked of a machine was to remember. Not to advise, not to predict, simply to hold the facts of trade so that no single person's memory was the only copy. The merchant's ledger did this on paper. When computers arrived, they did it faster and at a scale no clerk could match, and the ledger became the database.

The line runs from the account book through the first payroll and inventory systems to the great consolidating idea of the last century: enterprise resource planning. ERP did not spring from nowhere. It grew out of materials planning on factory floors, the discipline of working out how much steel and how many hours a production run would need, which widened into planning the resources of the whole enterprise, and then into a single connected system of record for the transactions of the business. A sale, a purchase, a shipment, a wage, a stock movement: each became a row, and the rows tied together, so that in principle you could stand in one place and see the operational truth of the company as a set of recorded facts.

This was a genuine achievement, and it is worth respecting rather than sneering at from the vantage of a later era. Before ERP, the left hand of a business quite literally did not know what the right hand had booked. After it, a firm could close its books, trust its stock figures, and know what it owed and was owed. The record became reliable, current, and shared. That is the foundational layer, and every layer above it still stands on that reliability. Take the record away and nothing else in the stack means anything, because a prediction from a broken ledger is a confident lie.

But the record layer had a native limitation, and it was the seed of the next era's problem. A system of record is built around nouns. It holds a customer, an invoice, an employee, an order. It is exquisitely good at telling you the state of a thing at a moment in time. It is silent about judgement. The ledger records that the price was set; it does not hold the setting of it. The record captures the outcome of a decision as if it were a fact of nature, stripped of the reasoning that produced it.

You can see this most plainly in a business that runs on judgement. When a consultancy books an engagement into its systems, the record shows a client, a contract value, a start date, a team. What it does not show is the decision that actually mattered: that the firm chose to win this work at a thin price because the logo would open a new sector, and that it chose to staff it lean on the bet that the client would not need as much hand-holding as the delivery lead feared. Those two judgements will decide whether the engagement makes money. The engagement is a noun in the database; the decisions that made it are nowhere, and when, a year later, the project has lost money, the record can tell you that it did but never why it was allowed to.

And as businesses poured more and more of themselves into these systems, a new difficulty appeared, one the ledger's inventors never had to face. There was now so much recorded that no human could read it. The firm had achieved perfect memory and lost the ability to see. Recording had created a reporting problem.

The era of seeing what happened

If the first era's question was "can we hold the facts," the second era's question was "can we see them." A database with ten million rows is not knowledge. It is a haystack that happens to contain the answer, and a busy leader cannot go row by row. So a new layer formed on top of the record, and its job was not to store but to summarise.

This is the era of business intelligence, of the data warehouse and the reporting suite and, eventually, the dashboard on every executive's screen. The mechanism was a real advance. You took the transactional records, which were organised for holding data safely, and you copied them into a second kind of system organised for reading data quickly: the warehouse. Then you built on top of it the reports, the cubes, the charts, the red and green tiles. Now a leader could ask "how did the north region do last quarter" and get an answer in seconds rather than a week, assembled from records that no person could have read by hand.

Business intelligence solved the bottleneck the record era created. It turned an unreadable mass of transactions into a picture a human could take in. For a while it felt like the destination, the single pane of glass, the source of truth at last. Firms built whole functions around it and taught a generation of managers to start the day by looking at their numbers.

Notice, though, what business intelligence is and is not, because this is the crux of the whole book. It is a mirror held up to the record. It shows you, with great skill, what has already happened. Revenue by month, utilisation by team, churn by cohort, margin by client: all of it is the past, aggregated and made legible. A dashboard is a beautifully rendered rear-view mirror. It tells you the shape of what was, and it is genuinely valuable to see that clearly.

But a mirror cannot tell you what to do. Stand a commercial leader in front of the best dashboard ever built and they will look at the falling utilisation line and ask the question the dashboard cannot answer: so what do we do about it? The chart shows the where and the when and the how much. It is mute on the why, and it is completely mute on the what next. It does not know that utilisation fell because a key engagement slipped, that the engagement slipped because the wrong people were staffed, that they were staffed that way because a promise made in the sale could not be honoured. The dashboard shows the symptom as a coloured tile and leaves the diagnosis, the choice, and the consequence entirely to the human standing in front of it.

Picture the Monday leadership meeting at a mid-sized consultancy. The dashboards are up on the screen and they are good ones: pipeline by stage, utilisation by practice, margin by engagement, all current, all accurate. The room reads them fluently. Utilisation in the data practice has slipped three weeks running, and everyone can see it. Then comes the pause, and then the real meeting begins, the one the screen has nothing to say about. Do we move two people onto the struggling account or hold them for the deal we think closes on Thursday? Did we not try exactly this in the spring, and did it work? The chart brought the room to the edge of the decision and then fell silent, because a dashboard has no opinion and, more tellingly, no memory of the last time the room stood in this spot.

Seeing had created a deciding problem, though the industry did not yet name it that. It reached, instead, for the next obvious thing. If the machine could summarise the past so well, perhaps it could be taught to guess the future.

The era of forecasting what might happen

The third layer is the one that has soaked up most of the last decade's attention, money, and noise. If business intelligence answered "what happened," data science and machine learning promised to answer "what will happen." Feed a model enough history and it will learn the patterns in it, and having learned them, it will extend them forwards. Which customers are likely to churn. Which deals are likely to close. Which candidates are likely to succeed. Prediction, at scale, from data.

This too is a real capability, and it sits, like every layer, on the ones beneath it. A model is only as good as the records it learned from and the reporting discipline that cleaned them. Give a firm reliable records and a well-tended history, and a model can find signal in it that no dashboard would ever have surfaced, because the pattern is buried in interactions across thousands of cases that no human eye would assemble. Prediction extended the business's sight from the past into the probable future. That is not nothing. For some problems it is a great deal.

But prediction, for all its power, did something quietly awkward to the fundamental problem, and it is worth being precise about what. A prediction is not a decision. It is an input to one. The model says this client has a seventy per cent chance of not renewing. Very well: what do you do? Call the client, or write them off? Discount to save the renewal, or hold the price and protect the margin? The number sharpens the question. It does not answer it. Between the prediction and the action stands a human being, weighing things the model never saw: the relationship, the strategic value of the logo, the state of the delivery team, the fact that this same client was flagged last quarter and stayed.

And here the pattern of the whole history reaches its sharpest point. The prediction era, by producing better and better inputs to decisions, threw a harder and harder light on the fact that the decision itself was still nobody's system of record. The model produces a score. Someone acts, or declines to act. And then, almost universally, three things happen that ought to trouble us. Nobody records what was decided, as distinct from what the model suggested. Nobody records who owned the choice or what they weighed against the score. And nobody, months later, goes back to ask whether the decision was any good, whether the client renewed, whether holding the price was right, whether the model deserved to be trusted or overruled.

A staffing firm learns this the expensive way. It builds a model that flags contractors likely to drop out mid-placement, and the model is a decent one. It raises a warning on a placement that matters. A manager sees the flag and thinks about the contractor, the client, the notice period, and the cost of pulling someone now against the risk of a gap later, and decides to leave the placement in place and watch it closely. Three weeks on, the contractor walks, the client is furious, and the manager's careful reasoning has vanished without trace. Nobody wrote down that a human looked at the flag and chose to override it, or why. So the firm cannot tell, across a hundred such moments, whether its managers are right to overrule the model or whether the model should more often win.

The prediction era, in other words, created the deepest bottleneck yet, and named it by omission. It surrounded the decision with better and better evidence and left the decision itself unheld, ungoverned, and unlearned from. Forecasting had created a deciding-and-learning problem. That is the problem the fourth layer exists to solve.

The era of deciding well, and remembering that we did

Follow the pattern to its conclusion. Each era added the capability its predecessor lacked and solved the bottleneck its predecessor created.

The lack that prediction exposed is the decision. Not the data behind it, not the report that framed it, not the model that informed it, but the choice itself: the moment a person or a group looked at the evidence and committed the firm to a course. That moment is the most valuable event in any business, and it is the one event the entire stack was never built to hold. This is where the next layer forms, and it has a name. It is called Decision Intelligence, and it is the discipline of treating decisions as first-class objects: things that can be captured as they are made, connected to the evidence that informed them, governed like anything else the business considers important, and, above all, learned from once their consequences are known.

Put the whole arc in a single line and the shape is unmistakable. The ledger recorded what happened. The dashboard showed what happened. The model forecast what might happen. Decision Intelligence captures what we chose to do about it, and then closes the loop by checking, later, whether we were right. Record, report, predict, decide, learn. And the fifth word is the one that turns the line into a wheel: what a firm learns from each decision becomes the evidence for the next, so that deciding and learning feed one another, round after round. That loop, not any single layer, is the engine.

What makes this layer categorically different from the three beneath it is not that it is newer or cleverer. It is that it is oriented around a different kind of thing entirely. The systems beneath it are organised around nouns: the customer, the invoice, the prediction. Decision Intelligence is organised around a verb. It manages the act of deciding. And it is bound to a second thing the lower layers never held together with the choice: the outcome. Not the outcome as an isolated fact in the ledger, but the outcome connected back to the decision that caused it, so that the two can be looked at side by side.

That pairing, the decision joined to its outcome, is the whole game, because it closes a loop that business intelligence leaves permanently open. Think about what a dashboard actually does with the future. Nothing. It reports, and then it forgets, and next quarter it reports again, and it has no memory of what anyone chose to do about last quarter's numbers or whether it worked.

Decision Intelligence closes that loop on purpose. It captures the decision at the moment it is made, while the reasoning is still live and the alternatives are still remembered. It connects that decision to the evidence that informed it and the outcome that followed it. And it makes the whole record reviewable, so that a firm can do the one thing the amnesiac enterprise of Part II never could: look back at a thousand decisions, see which kinds went well and which went badly, and get measurably better at the next thousand. That is not a fancier dashboard. It is a different object of management. Records preserve facts. Decision Intelligence preserves judgement.

Make it concrete. A renewal comes up on a large account, and the account manager, against a nervous instinct to discount, holds the price. In a firm practising this discipline she does not merely send the email; she captures the choice as she makes it: hold price, her name on it, the evidence she weighed (the client's usage is up, the last discount set a bad precedent, delivery has been strong all year), and the alternative she rejected (a ten per cent cut to be safe), all recorded in the moment, in the flow of the work, not as a chore bolted on afterwards. The client renews at full price, and that result attaches itself to the decision. On its own, one such record is a tidy note and little more. But do this across two years and a thousand renewals and the firm can finally see something no dashboard ever showed it: that holding price on high-usage accounts with strong delivery works far more often than the nervous instinct believes, and that the reflex to discount has been quietly costing it margin all along. The firm has learned to decide, not merely recorded that it did. And the next account manager to face a nervous renewal, two years on, does not start from instinct. She starts from what the firm already knows.

This is also why the discipline insists on the second half of its promise: not just deciding well, but remembering that we did. Part II described the amnesiac enterprise, the firm whose memory lives in its people and so walks out of the door with every resignation. It forgets because the decision, the one event worth remembering, is the one event nothing was built to keep. When the account manager who held that price moves on, her reasoning leaves with her, and the next person re-derives it from scratch, or, more likely, does not, and discounts out of the very nervousness she had already learned past. A firm that captures its decisions can remember them after the people who made them have gone. That is not a filing improvement. It is the difference between an organisation that compounds its judgement and one that resets it every time someone hands in their notice.

And crucially, it does this for the decisions that were never near a model at all. The prediction layer only ever touched the small fraction of choices that happened to have a score attached. The vast majority of decisions in any business, and the overwhelming majority in a people business, are made by human beings weighing human things: whether to push a deadline, which partner to put in front of a nervous client, whether to take the loss-making project because the relationship is worth it. Those decisions never met an algorithm and never will, and they are the ones that actually run the firm. Decision Intelligence is the first layer built to hold them.

But isn't this just prescriptive analytics with a new name?

Here is the objection a well-read reader will have been sharpening for several pages, and it is a fair one.

The analytics world already has a tidy ladder. Descriptive analytics tells you what happened. Diagnostic analytics tells you why. Predictive analytics tells you what will happen. And prescriptive analytics, the top rung, tells you what to do about it, recommending an action by running an optimisation over a model. If prescriptive analytics already recommends the action, is Decision Intelligence not simply that same idea with a more fashionable label wrapped around it?

No, and the difference is not cosmetic. It is a difference in what the thing is for and where it lives.

Prescriptive analytics is a way of producing a recommendation. It sits inside the model layer, at the top of it, and its output is an instruction: given this data and these constraints, take this action. It is an excellent capability when a decision is well-structured enough to be modelled, which is to say when the options are enumerable, the objective is quantifiable, and the constraints are known. Route this delivery van. Set this price. Allocate this inventory. For that class of problem, a prescriptive engine can beat a human handily.

But notice three things it does not do. First, it recommends; it does not record what was actually decided. The whole point of a decision is that a human can accept, modify, or overrule the recommendation, and prescriptive analytics has no place to hold the human's actual choice, the override, or the reason for it. Second, it does not govern. It has no notion of who owned the decision, who signed it off, what alternatives were genuinely on the table, or what evidence beyond its own inputs was weighed. Third, and most tellingly, it does not learn from the decision as a decision. It may retrain on outcomes as data, but it does not hold the choice and its result together as a reviewable object that a person can inspect and judge.

Decision Intelligence is the wider discipline, and its unit is not the recommendation but the decision. It captures, governs, and learns from the decision whether or not any model was involved at all. A prescriptive recommendation, when there is one, becomes simply one piece of evidence inside the captured decision, sitting alongside the human judgement, the context, and the alternatives. Prescriptive analytics is a clever engine that lives inside the third layer of the stack. Decision Intelligence is the fourth layer itself.

Put it at its plainest. Prescriptive analytics automates a decision that can be automated. Decision Intelligence takes care of the decision whether or not it can be automated, which is why it is a discipline and not a feature. The two are not rivals. A firm that practises Decision Intelligence will happily use a prescriptive model wherever one earns its place, the way it uses a dashboard or a database, as a tool inside a larger discipline. But the discipline is the point, and the discipline is new.

The bridge

So the history holds. Decision Intelligence falls exactly where the pattern says the next layer should fall: on top of the record, the report, and the model, depending on all three, solving the bottleneck that the model era created by making the decision the thing worth managing. It is not a repaint. It is the next storey of a building that has been rising, one floor at a time, since a merchant first wrote a sale in two columns.

But naming a layer and living in it are different things. It is one thing to say that decisions should be captured, connected, governed, and learned from, and quite another to say precisely what a captured decision is made of, how it is structured, and what has to be true for a firm to actually hold one.

That is the work of the next chapter. It defines the discipline exactly, separates it cleanly from its neighbours in business intelligence and in artificial intelligence, and introduces the smallest and most important idea in the book: the decision object, the atomic unit of Decision Intelligence, and everything it must carry to be worth the name.