The challenge
A member works out on a Saturday morning whether a payment posted before the weekend, and what happens if it did not. A contact center representative answers that same question nine times before lunch and will answer it again after. A digital banking manager can see in the analytics that the questions arriving by phone and the questions arriving on the website are not the same questions, and has one website to answer both with.
Member ownership raises the stake on a wrong answer. A member who is told the wrong thing about a fee, a hold or an eligibility rule is an owner who was told the wrong thing, and the recourse is a phone call that takes more attention than the original question ever did. Scale makes that harder rather than easier. At tens of thousands of questions, a small rate of confident error reaches a large number of members.
So the requirements were strict before anything was built. Every answer had to trace back to published material. Whatever went live first had to be extendable. Narrow scopes had to stay narrow. Whether a question was answered had to be recorded. And no member data could be in scope.
The solution
One assistant first, a single grounded deployment on the credit union’s published material. It is still the highest-volume of the three, carrying 36,781 questions on its own, and it ran alone long enough to produce a record rather than an impression.
A second assistant was added alongside the first rather than replacing it, with its own grounding and its own actions. A third came later, narrower again and the smallest by volume at 1,480 questions. The assistants reach no account system and hold no member data, so they serve members and prospective members alike as anonymous visitors on the credit union’s website.
Nothing in the core stack came into scope, and no shared corpus had to be untangled. Each assistant is grounded on its own, so a narrow question gets answered from narrow material.
The capability that decided it
A new assistant here is an addition rather than a migration. The credit union put one in production, let it run, then added a second with its own grounding while the first kept answering, and later a third. Neither addition took the first assistant out of service for a day.
Platforms that make you widen one assistant to cover a new scope force a choice between breadth and precision. Separate grounding removes that choice. It also means the three behave differently in the record, because they are being asked different things, and the credit union can see that difference instead of having it blended into one average.
The results
The first assistant has taken 36,781 questions across 391 days carrying a question, with 58 revisions to its instructions. The second has taken 10,111 across 353 days, with 66 revisions. The third has taken 1,480 across 134 days, with 74 revisions. Across the estate, questions arrive on 446 of the 462 days it has existed, which is the shape of a service rather than a pilot.
The second carries roughly a quarter of the volume of the first and has been revised more often, which is the signal you would expect from a narrower front door taking harder questions. The third is the smallest and the most worked on of the three.
Separately, in the latest quarter, 97.8% of the answers to questions the credit union’s own material covers landed well. That rate was measured across 7,019 in-context questions. Questions arriving on subjects the published material does not cover are excluded from the figure rather than counted against it, and they are tracked as a list of content to write.
Why it worked
The estate could grow. Two assistants were added without interrupting the one already in service, so the credit union never had to choose between keeping a working deployment and extending it.
Scopes stayed separate. Each assistant carries its own grounding and its own tuning, 198 revisions between the three of them, and two of the three run advanced reasoning over their own material.
The record is kept per assistant. Volume, tuning and answer quality accrue separately, which is what lets the credit union see which assistant carries the load, which one is taking the harder questions, and where to act next without disturbing the other two.
| Measure | To date |
| Assistants in production, added one at a time | 3 |
| Questions asked, raw count across the three | 48,372 |
| Questions taken by the first assistant | 36,781 |
| Questions taken by the second assistant | 10,111 |
| Questions taken by the third assistant | 1,480 |
| Revisions to how the three answer | 198 |
| Days carrying a question | 446 of 462 |
| Satisfied answer rate, latest quarter | 97.8% |
| In-context questions behind that rate, latest quarter | 7,019 |
Figures from CustomGPT.ai product usage records, scoped to the credit union’s three production assistants. Shared anonymously at the customer’s request.