Conversational AI Is an Omnichannel Problem
A conversational interface promises that the user can describe their problem in their own words and have it solved. Behind that promise sits an architectural temptation. If the model already understands the problem, it could decide what to do about it too. Wire the tools directly to the agent, and the decision layer disappears as a thing you have to build and maintain.
If the decision moves into the model, it establishes eligibility and pricing from its own context instead of calling the official rule engine. It works ninety-nine times out of a hundred, but in a regulated environment that rate does not even reach a pass.
The hundredth case is not a bug you can find. The same question asked twice can produce different answers, and a defect you cannot reproduce is a defect you cannot close. When a customer disputes a price or a refusal, the answer to why has to be a rule, a version, a timestamp. A decision you cannot trace back to one will not hold up in a dispute, an audit, or a court.
The AI Act tightens the same principle: automating a pricing or eligibility decision is what can put a system in the high-risk category, with the traceability and human oversight obligations that follow.
A boundary keeps the layer intact. The model owns understanding and expression: it interprets what the user meant, asks for what is missing, and phrases the answer so a person can act on it. Everything the answer depends on comes from code. Pricing lives in the pricing service, eligibility in the rule engine. The model calls them, receives a result and a reason, and reports it without deriving the number itself. Composite AI means a generative layer for language and a deterministic layer for truth.
A model may also sit deep in the back end, classifying, extracting, summarising. That is a different system with its own boundaries and its own accountability. The conversational channel is where the layer is easiest to lose.
Anyone who ran an omnichannel programme already knows this shape. Web, mobile and the call centre were channels, and none of them held the pricing logic. The channel presented what the system decided, and that separation still holds. Conversational AI occupies the same layer of responsibility, so the architecture you built around presentation and business logic accommodates it without modification. The architectural discipline it calls for is the one you already have.
If you're designing enterprise conversational AI systems, let's talk.