Architecture debates often start with a model. In an operating process, they should start with a boundary: what must be exact, and what can be generated. Getting that boundary right is most of the design.
Two kinds of work in one call#
A customer conversation contains both. Handling an interruption, phrasing a clarification, or recovering from an unexpected answer benefit from generative behaviour. Committing to a date, an amount, an eligibility rule or a legal statement must not be left to a model’s judgement.
- Deterministic core: business rules, eligibility, amounts, dates, statuses and next actions.
- Generative layer: language, clarification, recovery and natural turn-taking.
- Orchestration: deciding which layer handles a given moment, and constraining what the generative layer may say.
- Verification: checking that what was produced is consistent with the deterministic core before it is acted on.
Boundaries are project-dependent; the same process in two markets can need different ones.
Why this matters for measurement#
If the boundary is unclear, results become unattributable. A change in outcome could come from a better conversational model or from a rule that was quietly relaxed. A clear boundary keeps the experiment honest: the generative layer changes how the conversation goes; the deterministic layer keeps what was promised answerable.
Practical implications#
Before selecting models, write the boundary. Then choose components per process — voice, models, telephony and integrations are selected for the job, not fixed in advance. That is the difference between an operating stack and a product promise.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS