The four primitives of decision intelligence

Most organisations are not short of analysis. They have dashboards in every department, a warehouse full of clean tables, and analysts who can answer almost any question you put to them. Yet the same gap appears again and again. A number moves, someone notices, a chart gets shared, and then nothing happens. The analysis was correct. It just never became a decision.

This is the part of the problem that better data does not solve. You can make the chart faster, prettier, and more accurate, and the gap stays exactly where it was. The reason is that a decision is not a fact about the past. It is an act in the present, taken by a named person, for which someone is accountable, and whose effect can later be checked. Analysis produces facts. Decisions require a different set of moving parts.

This guide names those parts. We call them the four primitives because they are the smallest set of capabilities that, working together, carry a finding all the way from "the number changed" to "we acted, and here is what it did." They are deliberately vendor-neutral. You can build them on a spreadsheet, in a decision intelligence platform, or out of meetings and goodwill. The point is not the tooling. The point is that if you remove any one of the four, the loop quietly breaks, and you are back to sharing charts that no one acts on.

The four primitives form a chain. Each one assumes the one before it and feeds the one after it. Causality without ownership produces an unowned insight. Ownership without a trigger produces a name on a metric that no one looks at. A trigger without measurement produces activity that is never checked. The order matters, and so does the completeness.

1. 1

   Significance-tested driver edges

   A model of what actually causes the headline metric to move, where each link between a metric and its driver has been tested for significance rather than assumed. This is the causality primitive. It answers the question every meeting starts with: why did this change? Without it, you have a pile of correlations and a room full of plausible stories, and no way to tell which story is real.

2. 2

   RACI ownership on every metric

   A named owner on every metric in the model, expressed as RACI: who is Responsible for the work, who is Accountable for the outcome, who is Consulted, and who is Informed. This is the ownership primitive. It answers the second question, the one that usually goes unasked: whose job is it to act? Without it, the causal model tells you exactly which lever to pull and leaves no hand on the lever.

3. 3

   Push to the accountable owner

   A trigger that delivers the finding to the accountable owner at the moment the metric moves, rather than waiting for them to open a dashboard. This is the action primitive. It answers the question of timing: when does the decision get made? Without it, the insight sits in a tool that the one person who could act on it has no reason to open today.

4. 4

   A verified impact loop

   A check, after the action is taken, that confirms whether the metric actually moved in the way the action intended. This is the measurement primitive. It answers the final question, the one almost nobody asks: did it work? Without it, the organisation acts and never learns, repeating interventions that do nothing and abandoning ones that quietly succeeded.

The first primitive is a causal model of the business. The most workable form of this model is a metric tree: a headline metric at the top, decomposed into the drivers that cause it to move, those drivers decomposed further, all the way down to the operational inputs a team can actually change. If you are new to the idea, the foundational explainer is what is a metric tree.

What turns a metric tree from a diagram into a decision tool is the quality of its edges. An edge is a claim: this metric moves because that driver moves. It is tempting to draw these edges by intuition, because the org chart or the data model suggests them. The discipline that separates a real causal model from a wishful one is significance testing each edge. Does the driver explain a meaningful, repeatable share of the movement in its parent, or does the relationship vanish once you account for noise and seasonality? An edge that survives that test is a lever you can trust. An edge that does not is a story you have been telling yourself.

Read the tree above as a set of testable claims, not a hierarchy of importance. The edge from Gross Churn to Product Engagement says that when engagement falls, churn rises. That is checkable. If the data shows the relationship holds, the edge stays and Product Engagement becomes a place worth investing attention. If the data shows engagement and churn drift independently, the edge is noise dressed up as insight, and acting on it wastes a quarter. Significance testing is what keeps the tree honest as the business changes around it.

A causal model tells you which lever to pull. It says nothing about whose hand should be on it. This is the most common gap in analytics, and the most quietly expensive. A perfect metric tree, fully tested, beautifully rendered, will change nothing if no one is accountable for any node in it. The deeper treatment lives in why metric trees need ownership, but the short version is that a metric without an owner is a fact, and facts do not act.

The ownership primitive is best expressed as RACI on every metric. RACI distinguishes four roles, and the distinction matters because real organisations are not flat. The Responsible person does the work. The Accountable person owns the outcome and is the one the trigger should reach. The Consulted are pulled in for their expertise before a decision. The Informed are told after the fact. Collapsing these into a single "owner" field loses the information that makes the model operational, because the person who runs the campaign is rarely the same person who answers for the revenue number it feeds.

There can be only one Accountable role per metric. That constraint is the whole point. It removes the diffusion of responsibility that lets a number drift for three quarters while everyone assumes someone else is watching it. When the metric moves, there is exactly one person whose job it is to decide what happens next, and the rest of the chain knows who that person is. This is also what makes the next primitive possible, because a push has to be pushed to someone.

The first two primitives establish what to do and who should do it. The third establishes when, and it is the one that most quietly defeats good analytics. The default model of business intelligence is pull: the insight sits inside a tool, and it is the user's job to go and find it. This works for the small number of people who live in the dashboard. It fails for almost everyone else, because the person accountable for a metric is usually busy doing the work the metric measures, not refreshing a chart about it.

The action primitive inverts this. When a metric moves in a way that matters, the finding is pushed to the accountable owner where they already work, whether that is an inbox, a chat tool, or a task list. They do not have to remember to check. The system remembers for them. This is a small mechanical change with a large behavioural consequence, because it removes the single most common point of failure in the whole loop: the moment when an insight that no one happened to look at simply expires.

The push is only as good as the two primitives beneath it. Push without causality floods owners with alerts about movements that do not matter, and they learn to ignore the channel. Push without ownership has nowhere to go, so it becomes a broadcast that everyone skims and no one acts on. This is why the order of the primitives is not arbitrary. The push is meaningful precisely because the edge that triggered it was tested and the person it reaches is accountable.

The fourth primitive is the one almost no organisation has, and the one that turns the other three from a workflow into a learning system. After the accountable owner acts, something has to check whether the metric actually moved in the way the action intended. This is the verified impact loop. It closes the circle by feeding the outcome of the decision back to the model that prompted it.

Most organisations stop one step short of this. An owner is told a metric dropped, they run an intervention, and then attention moves on to the next fire. Nobody checks, four weeks later, whether the intervention did what it was supposed to. The cost of skipping this step is not a single wasted action. It is the loss of the ability to learn at all. An organisation without a verified impact loop repeats interventions that do nothing, because it never noticed they did nothing, and abandons interventions that worked, because it never noticed they worked.

Held separately, the four primitives are familiar. Causal modelling, ownership frameworks, alerting, and impact measurement each have a literature and a long history. What is rare is to find them connected into a single loop where the output of one becomes the input of the next. That connection is where the value lives, because a chain is only as strong as its weakest link, and a decision loop with a broken link is not a slower loop. It is no loop at all.

Trace a single change through the complete loop. Net revenue retention dips. The significance-tested edges point to gross churn, and within churn to a fall in product engagement, because that edge has held in the data before. The RACI on product engagement names the accountable owner. The push reaches that owner in their chat tool the morning the dip is detected, not in a review six weeks later. The owner runs an intervention. Four weeks on, the verified impact loop checks whether engagement and, in turn, retention recovered, and records the result against the edge that started the chain. The next time the same pattern appears, the model is a little sharper and the response a little faster.

This is also a useful lens for evaluating any tool that claims to bridge the gap between data and decisions, the gap that the wider field of decision intelligence exists to close. Do not ask whether it is powerful. Ask which of the four primitives it provides and which it leaves to you. A tool that nails causality and ownership but leaves action and measurement to meetings and memory has handed you two of the four. That is not nothing, but it is not the loop, and the missing half is where most decisions quietly die.

The four primitives are not a finished destination. They are the floor that the next generation of analytics has to stand on, and the direction of travel is towards each primitive becoming more automatic and more closely joined to the others. The edges get tested continuously instead of in occasional reviews. Ownership is read from how the organisation actually works rather than maintained by hand. The push gets more selective as the model learns which movements an owner truly needs to see. The impact loop runs without anyone remembering to close it.

This is also the foundation that agentic systems need to be trusted with anything that matters. An agent that can read a significance-tested causal model knows which lever to consider. An agent that can read RACI knows whose decision it is and where its own authority ends. A push gives the agent a moment to act, and a verified impact loop gives it, and the people supervising it, a way to know whether the action was sound. The same four primitives that make a human decision loop reliable are what make an automated one safe. The further reading on this sits in agentic analytics.

KPI Tree is one implementation of these four primitives as a single connected loop. The metric tree holds the significance-tested driver edges. RACI ownership sits on every metric, with one accountable owner each. When a metric moves, the finding is pushed to that owner where they already work. After they act, the verified impact loop checks whether the metric moved as intended and feeds the result back into the model. The reason to build the four together, rather than buying four tools and hoping they meet in the middle, is that the loop only pays off when it is closed. Each primitive on its own is an improvement. The four together are the difference between an organisation that analyses and one that decides.