A language model does real work in about twenty places in ExaBusiness: field-level writing assistance, multi-turn field chat, business-plan and policy generation, marketing blueprints and economics estimation, content drafting, influencer discovery, web enrichment of accounts, contact enrichment, person insight, LinkedIn company matching, the weekly-plan interview, domain name ideas, grant summarisation, and the site-extractor generator that reads a page and works out how to parse it.
That list is not the point. This is:
One door
Every model call in the product goes through a single interface. A caller asks for a tier — light, standard or high — and gets an answer. It does not name a model, hold a key, or know which provider answered.
Three things follow from that, and they are the whole argument:
Changing model is configuration. When a better or cheaper model appears, it is a settings change, not nineteen refactors.
Tests never call a provider. One interface is one thing to decorate, so the test suite replays recorded conversations — and fails on a call nobody recorded rather than quietly spending money. This is why the AI features have real test coverage instead of being the part of the system nobody dares run.
There is one place to audit. What is sent, and to whom, is answerable.
Gaps are marked, not filled
Generated documents mark what the model could not determine, in place, instead of producing a fluent paragraph with an invented number in it. You get a draft and a list of the questions only you can answer. A document that is honestly incomplete is worth more than one that is confidently wrong, because you can act on the first.
Not everything should be a model
Some of the most useful answers in the product involve no AI at all, deliberately:
- Marketing economics is arithmetic, and shows its working.
- Channel scoring is seven weighted criteria with a written reason per criterion.
- SLA and breach clocks, evidence expiry and the notifiable-breach countdown are dates and rules.
- Job state gates are a state machine.
Support, Projects, the Wiki, Field Works and Certifications contain no model calls whatsoever. Their answers are derived, and derived answers do not need to be approved — which is why the product distinguishes the two on screen.