Decision intelligence is one of those phrases that has been collapsed in two directions at once. Vendors have collapsed it down into a product category, calling almost anything with a model and a dashboard a "decision intelligence platform". Analysts and consultants have collapsed it up into a synonym for "AI in business". Both readings miss what the discipline actually is: a working method for connecting data, models, and rules to commitments that bind, under a stated scope, with a complete audit trail. This guide gives the operational definition.
Key takeaways
- **It is a discipline, not a product:** The field connects three pillars (data, models, and rules) to produce binding commitments under a stated scope, with an audit trail to defend them.
- **Three pillars must connect:** Data answers "what happened", models answer "what is likely", and rules answer "what is permitted now". Skip the rules pillar and the discipline collapses into analytics with extra steps.
- **The outcome is commit or refuse:** Unlike business intelligence or decision support, the field's output is a binding outcome inside scope, not a recommendation a human can ignore.
- **Origins are older than the AI wave:** The field's intellectual roots run through operations research, behavioural decision theory, and Lorien Pratt's mid-2010s synthesis. The AI wave amplified it; it did not invent it.
- **Refusal is part of the discipline:** A field that cannot say what it will not decide is not the field at all. Refusal is what makes commitments defensible.
A working definition
Decision intelligence is the discipline of connecting data, models, and rules into a system that ingests a request, verifies the inputs, evaluates a binding rule set, and either commits the outcome or refuses it, with every step logged for review. The discipline is older than the term, which Gartner (Gartner glossary entry) brought into general use, and narrower than the marketing that has since attached to it.
Two clauses inside that definition do the load-bearing work. First, stated scope: the discipline is only applicable to a bounded set of decisions, named in advance. Outside that scope, the practice is to refuse and route, not to extrapolate. Second, binding rule set: the rules that evaluate the commitment are written in code and the same on Friday as on Monday. A practice whose rules drift with the conversation is not yet a discipline.
For the broader system anatomy in which the discipline operates, our pillar on what a decision system actually is gives the five-component view.
Where the field came from
The intellectual lineage runs through three currents. Operations research, beginning in the 1940s and 1950s, gave us the language of constraints, optimisation, and bounded scope. Behavioural decision theory, from Herbert Simon's work on bounded rationality through Kahneman and Tversky, gave us the empirical view of how humans actually make commitments under uncertainty. Modern computer science, particularly Lorien Pratt's synthesis in the mid-2010s, brought those two traditions together with machine learning and named the resulting practice.
Gartner adopted the term, the consulting industry built around it, and by the late 2010s vendors had attached it to anything that surfaced a chart with a recommendation underneath. The marketing collapse is the noise. The discipline underneath is intact.
The three pillars: data, models, rules
The working method connects three pillars. Each pillar answers a different question, and all three must be present for a commitment to be defensible.
Data. The descriptive layer. What happened, how often, under what conditions. Without clean data, neither models nor rules can produce a defensible outcome. The OECD's AI Principles explicitly call out the integrity of the data layer as a precondition for trustworthy automation.
Models. The predictive layer. What is likely, given what happened. Scoring, forecasting, pattern detection, and anomaly surfacing all live here. The model is a sub-decider, never the decider.
Rules. The prescriptive layer. What is permitted now, given the prediction and the policy. The rules layer is the binding part of the discipline. Skip it and the field reduces to analytics with a commit button, which is the failure mode most regulated deployments eventually face.
The three pillars connect through a gate that evaluates inputs from all three and produces a binary outcome: commit or refuse. The practice we describe in our supporting guide on AI for decision making is what happens inside the model pillar specifically; the discipline is what surrounds the model.
Decision intelligence in practice
The field shows up under different names across industries, but the operational shape is the same.
Credit and lending. A loan application enters a known queue. The data pillar provides KYC, income, and exposure history. The model pillar scores affordability and default risk. The rules pillar evaluates regulatory thresholds (debt-to-income limits, affordability rules, sanctions screens) and the bank's own policy. The commitment is approve, refuse, or refer. The audit trail records every step.
Clinical trial operations. A protocol amendment request enters review. Data pillar provides the safety database. Model pillar flags adverse-event clustering. Rules pillar evaluates regulatory and ethics constraints. The commitment is approve the amendment, refuse it, or escalate to the data-monitoring committee.
Enterprise procurement. A purchase order enters the gate. Data pillar provides supplier risk, budget posture, and contract history. Model pillar flags vendor anomalies. Rules pillar evaluates sign-off authority and policy. The commitment is issue the PO, refuse it, or refer for higher authority.
In all three patterns, the visible artefact (a credit decision, a protocol approval, a purchase order) is downstream of the same three-pillar discipline. The vendors that wrap one of those domains and call themselves a "decision intelligence platform" are selling a configuration of the discipline, not the discipline itself.
How the discipline differs from BI, AI, and decision support
The most common confusion is between the field and three adjacent practices. The differences are operational, not philosophical.
Business intelligence surfaces information about what happened. A human reads the chart and decides. AI alone produces a prediction or score. A human (or downstream system) decides what to do with it. Decision support systems combine the two and produce a recommendation. A human evaluates the recommendation and decides.
The discipline differs from these in one operational respect: the commitment is bound by the rules pillar, inside the stated scope. The output is not a recommendation that someone else acts on. It is the action, logged. This is the line that separates the field from every adjacent practice, and the line that most vendor marketing blurs.
A practical closing
The reason this distinction matters is operational rather than rhetorical. A team that builds inside the discipline can defend its commitments to a regulator, an auditor, or an injured party by replaying the decision. A team that has built BI plus AI plus a recommendation engine cannot replay the commitment; the commitment was made by a person whose reasoning is not in the log. The field's promise is that the reasoning is in the log. Whether that promise holds in a given deployment is exactly what the vendor-evaluation criteria in our companion guide on decision intelligence platforms test for.