Enterprise AI strategy often starts in the wrong place. A company sees a compelling model demonstration, buys access to a platform and then begins searching for processes that might justify the investment. Decision intelligence reverses that sequence: start with the decision, define what better means, understand the consequences of error and only then choose the technology that should support it.
That is the operating principle behind Fernando Angulo’s April 3 essay “Decision Intelligence: Match the AI Tool to the Decision”, which is explicitly inspired by Cassie Kozyrkov’s work. Kozyrkov describes decision intelligence as the discipline of turning information into better actions. The formulation is useful because it moves attention away from whether an AI output looks impressive and toward whether it improves what happens next.
The practical consequence is that there is no single “best AI” for business. Some decisions are better handled by ordinary rules. Others require statistical prediction. Generative AI is useful when the problem involves unstructured information, synthesis or option creation. And some consequential decisions should remain human-owned even when AI contributes substantial evidence.
Tool-first AI strategy inverts the sequence
The current AI market makes tool-first planning tempting. New models arrive with longer contexts, better reasoning, stronger multimodal capabilities and increasingly autonomous agents. Each release creates pressure to identify a use case quickly.
Decision intelligence begins with a more basic question: what action should change if the system works? If nobody can answer that clearly, model selection is premature.
A polished analysis that changes no decision has little operational value. Conversely, a simple system that reliably improves a high-frequency decision can create substantial value without using frontier AI at all.
Use deterministic rules when the decision is already known
The first category is the least fashionable and often the most efficient. When the conditions are stable and the approved action is explicit, conventional automation may be the right system.
Eligibility checks, routing logic, threshold-based alerts and standardized approvals frequently fit this pattern. If an invoice below a defined amount with matching purchase-order data follows one approved path, an organization may not need a language model to rediscover that policy on every transaction.
Rules offer properties that generative systems do not naturally provide: predictable execution, straightforward testing and clear explanations of why a branch was taken. The limitation is brittleness. Rules perform poorly when the environment contains too many exceptions or when the decision depends on patterns that cannot be specified manually.
Use predictive models when probability changes the action
A second class of decisions depends on estimating an uncertain outcome from historical data. Demand forecasting, churn prediction, fraud detection, lead scoring and failure prediction are familiar examples.
Here the useful output is not a paragraph. It is an estimate that changes an action: reorder inventory, prioritize an account, inspect a transaction or schedule maintenance.
The central evaluation question is therefore not simply whether the model predicts accurately in aggregate. Teams need to know how errors affect the decision. A false positive and a false negative can have radically different costs, and the threshold that maximizes statistical accuracy may not maximize business value.
Use generative AI when the work involves synthesis, language and option creation
Generative systems are strongest when the input is messy and the desired output is not a single predetermined label. They can summarize research, compare documents, draft communications, produce scenarios, explain technical material and generate several plausible approaches to an ambiguous problem.
This makes generative AI particularly useful upstream of decisions. A model can reduce a hundred-page evidence set into competing arguments, identify gaps in a plan or create candidate responses that a person or another system evaluates.
But fluency creates a specific risk: teams can mistake an articulate output for a validated decision. Generative AI can make uncertain information sound complete. The workflow therefore needs evidence requirements and a clear rule for what happens after generation.
Keep consequential ambiguity human-owned
Angulo’s framework reserves a different category for decisions where ambiguity and consequences remain high. Hiring, unusual pricing exceptions, regulatory interpretation, reputation-sensitive actions and similarly consequential choices can benefit from AI analysis without transferring final ownership to the model.
The important distinction is between assistance and authority. An AI system can gather evidence, expose inconsistencies, model scenarios and challenge assumptions. The accountable decision maker still needs to understand why the final action is justified.
This does not mean that every decision currently reviewed by a person must remain manual forever. As evidence improves and processes become better understood, some decisions can move toward greater automation. Decision intelligence makes that transition explicit rather than treating autonomy as the default destination.
Every AI project needs a decision owner before it needs a model owner
One of the strongest ideas in the source framework is that each AI-supported decision should have an owner. Someone has to be accountable for the objective, the acceptable error rate, the evidence threshold and the response when the system is uncertain.
Without that ownership, organizations tend to evaluate AI projects using proxy measures: output quality, employee enthusiasm, number of prompts, model accuracy or time saved. Those metrics can be useful, but none proves that the underlying decision improved.
A decision owner can instead ask whether the system changed actions in the intended direction and whether the resulting outcomes justified the cost and risk.
Define success before seeing the AI output
Decision quality becomes difficult to evaluate when the success criterion is invented after the model has produced an answer. People naturally rationalize persuasive outputs, especially when those outputs align with what they already wanted to do.
A stronger workflow defines the objective in advance. If the project is meant to reduce customer churn, specify the action the prediction will trigger and the outcome that will be measured. If generative AI is meant to accelerate research, decide whether success means faster analyst throughput, broader evidence coverage, fewer missed risks or some combination of those outcomes.
This makes model evaluation subordinate to decision evaluation. The model is a component of the system rather than the system’s purpose.
The cost of error should determine the autonomy level
Two tasks can use the same model and deserve completely different governance. Generating five internal headline options has a low cost of failure because a person can discard bad suggestions. Automatically changing the price of thousands of products can have immediate financial consequences.
The model’s benchmark score is therefore insufficient for deciding whether it should operate autonomously. Organizations need to consider reversibility, scale, latency, monetary exposure, legal obligations and reputational impact.
This connects directly to the boundary-design problem NetContentSEO examined in “The Skill That AI Can’t Automate”. As AI becomes better at execution, the scarce organizational capability shifts toward deciding where autonomy should stop and accountable authority should begin.
Evidence thresholds prevent polished outputs from becoming automatic decisions
A decision-intelligence workflow should define how much evidence is required before action. That threshold can differ by decision.
A low-risk content suggestion may need no external verification. A competitive intelligence conclusion might require several independent sources. A compliance-sensitive action could require authoritative documentation and human review.
Generative AI is particularly useful here because it can assemble evidence, but it can also obscure evidence quality by synthesizing weak sources into a confident narrative. The workflow should therefore preserve provenance rather than evaluating only the prose produced at the end.
Uncertainty needs an operational response
Many AI implementations acknowledge uncertainty conceptually but do not specify what the system should do when it encounters it. A useful decision architecture turns uncertainty into a routing rule.
The system might request more information, run an additional analysis, compare a second model, escalate to a specialist or decline to act. The correct response depends on the decision and its consequences.
This is where combining technologies can be more useful than selecting one winner. Rules can enforce hard constraints, predictive models can estimate risk, generative AI can interpret unstructured evidence and humans can resolve high-impact ambiguity.
The operating map is a stack, not four isolated boxes
Real workflows rarely fit perfectly into one category. Consider customer retention. A predictive model can estimate churn probability. Deterministic rules can determine which accounts qualify for an offer. Generative AI can summarize the customer history and draft an outreach message. A human account owner can approve an unusual commercial concession.
The quality of the overall system depends on matching each component to the part of the decision it handles well. Asking one general-purpose model to perform every stage may be simpler to demonstrate but harder to validate and govern.
Decision intelligence therefore encourages decomposition. Instead of asking “Which AI should run this workflow?”, ask which parts require prediction, which require generation, which are already deterministic and which carry accountability that should remain explicit.
Better AI does not repair a badly framed decision
Organizations can spend substantial resources improving model accuracy while leaving the objective itself ambiguous. A sales model cannot resolve disagreement about whether the company values short-term revenue, margin, retention or strategic account growth. Those are business choices.
Likewise, a generative model cannot make an undefined approval policy coherent merely by expressing it fluently. If teams disagree about what constitutes an acceptable exception, automation can scale that disagreement rather than solve it.
Angulo’s framework is valuable because it exposes those problems before deployment. Poor data, conflicting incentives and unclear accountability become design issues rather than surprises discovered after automation is already changing behavior.
AI strategy should be measured at the action layer
Many organizations measure AI adoption through usage: active users, prompts sent, copilots deployed or hours saved. Decision intelligence suggests another layer of measurement: did the system cause better actions?
That can require experimentation. Teams can compare decisions made with and without AI support, track downstream outcomes and examine where recommendations were accepted or rejected. They can also identify whether automation changed behavior in unintended ways.
The objective is not to prove that AI is universally beneficial. It is to learn which decisions improve under which combination of tools, data and human oversight.
Tool selection becomes easier after the decision is specified
The current AI landscape encourages endless model comparisons. Teams debate context windows, benchmark scores, reasoning modes and agent capabilities before agreeing on the job the system must perform.
Once the decision is clearly specified, many of those choices become easier. A deterministic policy does not need a frontier reasoning model. A probabilistic ranking problem may be better served by a specialized predictive system. A document-heavy synthesis task benefits from generative AI. A high-consequence exception may require a human decision regardless of how capable the model becomes.
This does not eliminate model evaluation. It gives model evaluation a purpose.
The best AI architecture may deliberately contain less AI
One counterintuitive consequence of decision intelligence is that a mature AI strategy may sometimes remove generative AI from parts of a workflow. If an experimental agent reveals that a decision can be expressed reliably as a deterministic rule, the simpler system may be cheaper, faster and easier to audit.
Likewise, if a predictive score is sufficient to trigger an operational action, adding a language model between the score and the action can introduce unnecessary variability.
AI maturity should therefore not be measured by how many processes contain an LLM. It should be measured by whether the organization uses the appropriate decision mechanism for each part of the work.
Decision intelligence is the missing layer between AI capability and business value
Generative AI has made capability unusually visible. Anyone can watch a model write, code, analyze and reason in real time. Business value is harder to observe because it appears later, after someone acts on the output.
Decision intelligence connects those layers. Start with the action. Define the objective and the owner. Identify the evidence required and the consequences of error. Decide how uncertainty should be handled. Then choose among rules, predictive models, generative AI and human judgment—or combine them deliberately.
Kozyrkov’s definition captures why this ordering matters: the discipline is about turning information into better actions, not generating more information for its own sake. citeturn0search0
A useful first sentence for the next AI project is therefore not “Which model should we use?” It is the question Angulo leaves his readers with: what decision will change if this system works?