For a decade, most of what we built were decisioning systems. The unglamorous kind. Rating, eligibility, validation, policy enforcement: the work where the same inputs have to produce the same output, every time, and where someone can later be asked to explain exactly why the system did what it did. It is not the part of the industry that gets talked about at conferences. It is also the thing that taught us the most about where AI belongs.
The tool everyone is excited about now is probabilistic. The work we spent a decade on is the opposite. That contrast is the whole lesson.
What makes decisioning work exacting
A decision system does not get to be usually right. If a rating engine is right ninety-eight times out of a hundred, the other two are a mispriced risk, a complaint, or a regulator asking questions. Rating is only one example of the pattern; the same standard holds for an eligibility check, a validation gate, or a policy rule. The requirement is not accuracy in the loose sense. It is that the same case decided today is decided the same way tomorrow, and that the reason can be shown on request.
Building to that standard for years does something to how you think. You stop trusting anything you cannot trace. You develop an instinct for which parts of a process can tolerate a bit of drift and which parts cannot tolerate any.
What that instinct tells you about a probabilistic tool
AI is genuinely good at a wide class of work: reading, finding, drafting, summarising, holding a conversation. That work benefits from a tool that generalises, and being usually right is a fine standard for it. But a language model does not produce the same answer every time, and no serious vendor will warrant that it does. That is not a flaw to be engineered away. It is what the tool is.
So the question is never whether AI is impressive. It is where in a process a tool that drifts can be trusted, and where the work has to sit under a rule that does not. We mapped where that handoff lives in the Decision Pyramid, and the difference between finding an answer and being able to stand behind it in AI is brilliant at finding an answer.
"Years of building systems that must never drift teach you exactly where a tool that drifts belongs."
Why the decisioning layer is what makes AI safe to deploy
This is the part the market is now learning in public. An agent let loose on real work needs something to keep it inside the lines. That something is a deterministic layer: the rules that govern what can actually happen, no matter what the model proposes. The agent can read the case, gather the context, and suggest the outcome. The rule decides. When the boundary is drawn that way, you can give an agent real work, because the thing that carries the consequence is not the model.
In other words, the decade of decisioning work is not something AI leaves behind. It is the ground AI stands on. The deterministic layer is what turns an impressive demo into something a regulated business can actually run.
Does this make us sceptics?
The opposite. We deploy AI across customer service, knowledge, and process work, and we think most organisations will use far more of it, not less. Our approach to AI is to put it where it earns its place and keep it away from the decisions that have to hold. That is a more durable position than either the hype or the fear, because it does not depend on the technology being something it is not.
The quiet advantage
Everyone building with AI is discovering that it needs governing. We had a decade of building the governing layer before that was a fashionable thing to say. We are one company that has evolved into the AI era on top of that foundation, not a new arrival learning the hard lesson in production. The exciting layer sits on top. The layer that makes it trustworthy sits underneath. We spent a long time on the one underneath, and it turns out that is the part that decides whether the whole thing holds.
Frequently asked questions
Decisioning is the deterministic work of turning inputs into a consistent, explainable outcome: rating, eligibility, validation, policy enforcement, and similar. The defining requirement is that the same case produces the same result every time, and that the reason can be shown on request. It is distinct from the probabilistic work AI does well, such as reading, finding, and drafting.
Because a language model is probabilistic and does not produce the same answer every time, and no serious vendor will warrant that it will. For a decision that has to be consistent, explainable, and repeatable, usually right is not good enough. AI can read the case and propose an outcome, but the decision that carries consequence should sit under a deterministic rule that governs what actually happens.
No. It is what makes AI safe to deploy on real work. A rules layer gives an agent a boundary to operate inside, so the model can do the probabilistic work it is good at while the rule governs the outcome. That combination lets you use more AI, not less, because the thing carrying the consequence is the rule rather than the model.
The interface is the point in a process where probabilistic work hands off to deterministic work: where AI finishes reading, finding, and drafting, and a rule takes over to make the decision that has to hold. Placing that interface correctly matters more than model quality, because a probabilistic tool asked to own a deterministic decision will be usually right rather than reliably right.