There is a lot of noise right now about AI making decisions. For the people who answer for those decisions, the interesting question is quieter. When the audit comes, can you still explain this one, correctly, in three years, to someone who is not impressed by how clever the model was.
Most of the attention goes to the front of the decision: the reading, the finding, the synthesising, the part AI has genuinely transformed. That work matters, and we do it. But if you own a consequential decision, a price, an approval, an eligibility determination, a claim outcome, your problem was never the reading. It was that the call has to be right, has to be the same for the same facts, and has to be explainable long after the person who made it has moved on. That is a different job from finding the answer, and it needs a different kind of machinery.
Two halves, and neither is optional
The capability underneath a dependable operation is two halves that only work together. One half is the decision: the deterministic call that produces the same outcome every time for the same inputs, and can show its reasoning on request. The other half is the process: the workflow, orchestration, and integration that gets a case ready to decide and carries the outcome onward through the systems and people it touches.
Neither is useful alone. A correct decision trapped behind a broken process never reaches anyone. A smooth, automated process making inconsistent calls is not efficiency, it is a way to be wrong faster, and in a regulated setting that is worse than doing nothing. When people say "decisioning" and quietly mean only the rules, they miss half the problem. The process is not packaging around the decision. It is the half that makes the decision count.
"A smooth process making inconsistent calls is not efficiency. It is a way to be wrong faster."
What the accountable owner actually asks
The person who carries a decision does not ask whether it is clever. They ask a short list of unglamorous questions. Will it produce the same answer for the same case next month. Who owns this rule, and how fast can they change it without a release cycle. If a regulator or a customer asks why, can we show the exact logic that was in force at the time, not a reconstruction. When volume triples, does it still hold.
These are the questions we have spent a long time answering, because they are the questions our clients are answerable for. The work is externalising that logic out of the code and spreadsheets it is trapped in, onto a platform the business can read, test, and change, with versioned rules and an immutable record of every decision made. It is not glamorous. It is the thing that is still standing when the audit arrives.
Rating is the clearest example, not the whole story
The sharpest version of this is rating. It is the largest and most durable body of decisioning work we run, because a rating engine is the deterministic requirement in its purest form: high volume, real money, zero tolerance for drift, and a regulator who may ask about any single calculation years later. When we talk about rating, it is because it is the clearest proof that the discipline works at scale, not because it is all we do.
The same pattern sits underneath eligibility, underwriting, credit logic, validation gates, and policy enforcement. Rating is where it is easiest to see. It is not the identity of the capability. It is the anchor that proves it.
Where AI belongs, and where it hands off
None of this is an argument against AI. It is an argument about the interface between two kinds of work. AI is genuinely good at compressing the effort of getting a case ready: reading the document, extracting the fields, classifying the request, summarising the history. Deterministic rules are what should make the call that carries the consequence.
The point where the probabilistic layer hands to the deterministic one is the interface, and drawing it deliberately is what lets you get the productivity of AI without putting a probabilistic guess in the seat where a consistent, explainable decision has to sit. Put the AI in front of the decision, not in it, and both halves do what they are best at. The Decision Pyramid is how we map that boundary.
"Put the AI in front of the decision, not in it."
The quiet part
The reason this matters is not ideological. It is that the cost of an inconsistent consequential decision does not show up in the demo. It shows up later: in the adverse selection you did not price for, the complaint you cannot answer, the remediation programme, the finding. A decision that holds up is boring right up until the moment it is the only thing that saves you.
If you own decisions like that, the useful conversation is not about which model. It is about which decisions have to hold, how fast the rules behind them need to change, and how you prove they were right. That is the work we do, and it is the ground everything else stands on.
Frequently asked questions
They are two halves of one capability. Decision automation makes the consequential call the same way every time and can explain it; process automation is the workflow and orchestration that gets a case ready to decide and carries the outcome onward. Most stalled work needs both fixed, because a correct decision behind a broken process never reaches anyone, and a smooth process making inconsistent calls just scales the error.
Because a consequential decision has to be consistent, explainable, and reproducible, and a probabilistic tool asked to own that decision will be usually right rather than reliably right. AI is well suited to compressing the effort of getting a case ready to decide; the call that carries the consequence belongs with deterministic rules that produce the same answer for the same facts and can show the exact logic in force at the time.
No. Rating is the clearest and largest example of the discipline, because it is the deterministic requirement in its purest form, but the same pattern sits underneath eligibility, underwriting, credit logic, validation gates, and policy enforcement. Rating is the anchor that proves the discipline works at scale, not the identity of the capability.
With versioned rules, an immutable log of every decision made, and the ability to replay a decision against the rules that were in force at the time rather than reconstructing it after the fact. That is what lets you answer a regulator or a customer with the exact logic that applied, not an approximation.