The challenge
A salesperson in front of a customer needs to know whether two components fit together. A customer service agent is working a question that turns on a clinical protocol buried in a document library. The product specialist both of them escalate to has finite time, and is the only reliable route to an exact answer.
The knowledge existed. It was scattered. Specifications lived in dense technical documentation, compatibility meant cross-referencing matrices across systems and brands, and clinical protocols sat across hundreds of sites and libraries. Anyone who needed a specification first had to know exactly where to look, and anything complex went to a specialist and waited.
General-purpose AI tools had already been ruled out on this data. For a medical device manufacturer, a wrong answer delivered confidently becomes a safety problem before it becomes a service problem.
The solution
A connector sits in front of the company’s own product system and exposes specifications, compatibility relationships, product codes and manuals to the assistant as a single source.
Grounded retrieval handles everything else, running against more than 200,000 documents ingested from sitemaps, internal libraries and published instructions-for-use material, with answers restricted to approved sources.
Scope is enforced. The assistant answers product and technical questions, and routes anything outside that to the company’s own team. It launched internally, to sales and customer service teams worldwide, where the company could see the answers before anyone outside did.
The capability that decided it
The assistant queries the product data itself instead of reading about it.
That is what lets a technical question be answered the way a product specialist would answer it rather than summarized from a document that happens to mention the answer. Compatibility is resolved from the product data. A code can be found from a partial description or a related product rather than an exact-match search. A clinical protocol is answered from the approved source, with the source shown.
Everything the structured data does not hold runs on grounded retrieval across the approved corpus. One assistant, two ways of answering, and neither one loosened to accommodate the other.
The results
Sales and customer service teams worldwide have asked 26,065 questions on 371 separate days. More than nine in ten arrive through live chat, inside the company’s own environment.
Reported on its own window: in the latest quarter the satisfied answer rate was 99.5%, measured across 5,334 questions that the company’s approved sources covered. Of the questions this corpus can speak to, that is the share whose answers landed well.
The figure deliberately excludes questions on subjects the approved sources do not cover. Those 659 questions in the same quarter are not counted against the assistant. They are treated as a content list, because that is what they are. The rate is not a claim that the assistant answers 99.5% of everything asked of it, and it is not a measure of accuracy or correctness.
The deployment has been reviewed at senior executive level, and the company has since renewed.
Why it worked
Grounding that holds in a regulated setting. Answers restricted to approved sources made a deployment possible where general tools had been ruled out on that exact point.
Structured data, read directly. The connector separates an exact technical answer from a plausible one.
A scope narrow enough to defend. Product and technical questions are answered, everything else routes to a person, and nobody has to argue about the edges.
Built by the team that owns the knowledge. A no-code interface meant the owning team could build the assistant themselves rather than hand off a specification.
The company is now broadening adoption across more business units and departments, working toward deployments serving professionals outside the organization, and developing deeper integration with its own proprietary knowledge sources.
| Measure | To date |
| Questions asked since launch | 26,065 |
| Days carrying a question | 371 |
| Approved documents made answerable | More than 200,000 |
| Share of questions arriving through live chat | More than nine in ten |
| Satisfied answer rate, latest quarter | 99.5% |
| Questions the approved sources covered, latest quarter | 5,334 |
| Questions on subjects the corpus does not cover, latest quarter | 659 |
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.