The challenge
A contractor compares business checking accounts late at night, because that is when the paperwork gets done. A relationship manager answers the same question about minimum balances and transaction limits a dozen times a week, for prospects who will not become customers for another month. A marketing lead owns the website and knows the answers are all on it somewhere, spread across product pages, rate sheets and disclosures.
The bank publishes everything it needs to. The material exists, it is accurate, and it is maintained. What it does not do is answer a question, because a rate sheet is a document rather than a reply, and a visitor with one specific question about one specific account has to find it first. The gap was never a shortage of published material. It was the distance between a document and an answer.
The solution
One assistant on the public website, available to anyone who visits, with no sign-in and no account required, because the visitor with the question is often not yet a customer.
Scoped to what the bank publishes. The corpus is the bank’s own product pages, rate material and disclosures, which makes the boundary of the answer the same as the boundary of the commitment. Not the open web, not a model’s general knowledge of banking, and not an older version of a rate sheet.
Actions on top of answers. Agentic actions are switched on, so a reply can move a visitor to the right application or the right page rather than only describing one.
One assistant, unchanged. The same deployment has served the whole period, through two years of ordinary product and rate revision, with the wording kept current by 52 revisions that each landed on a deployment already in service.
What it did not take is worth listing. No core banking integration, because the assistant answers from published material and touches no account system. No rewrite of the website, because product pages and rate sheets went in as the bank already maintains them. No second team, because the deployment has not needed a rebuild and the count of assistants has stayed at one.
The capability that decided it
Agentic actions run on top of every reply, so the assistant can move a visitor to the right application, the right product page or the right branch rather than leaving them to search again. For a bank this is the difference between a well-written paragraph and a started application late at night.
An assistant that only replies is a better-phrased version of the FAQ page, and the bank already had one of those on the website. Page awareness and lead capture came built in, with no second system and no integration project behind them, which is why the capability arrived with the deployment rather than as a project of its own.
The requirement underneath it is a banking one. What the bank tells a visitor about a rate, a fee or an eligibility rule is a public statement about a regulated product, so the answer has to come from published material and then has to be able to do something useful with the visitor’s attention while it still has it.
The results
Read the 752 first. A launch spike shows up as a busy fortnight and a long tail of silence, and this is the opposite shape. The assistant has been asked something on 95.1% of the days it has existed, across twenty-six months, two rate cycles and the ordinary revision of a bank’s published material.
The volume is modest by design, at 13,932 questions, because this is one assistant on one website rather than an estate. Pilots end. This one is still being asked something on almost every day it has been live, and demand that survives that long has stopped being curiosity and become a habit the bank’s customers have formed.
One limit is worth stating plainly. No answer-quality classification exists on this deployment, so no answer rate appears anywhere on this page. The day counts describe how often someone asked, not how well any particular reply landed, and the two should not be read as the same thing. Deflection, satisfaction and staff hours saved are not measured here either.
Why it worked
Grounding is the product. The answer boundary is the bank’s own published material, which is the only boundary a regulated institution can defend in public.
Two actions came built in. Page awareness and lead capture run on every answer, with no second system and no integration project behind them.
Marketing kept it current. There have been 52 revisions to how the assistant answers across twenty-six months, every one landing on the deployment already in service.
The estate stayed at one. No second assistant has been built, which is the clearest evidence available that the deployment has not needed rescuing.
| Measure | To date |
| Time in production | 26 months, one assistant |
| Questions asked, raw count | 13,932 |
| Days carrying a question | 752 of 791 |
| Share of days in production carrying a question | 95.1% |
| Revisions to how the assistant answers | 52 |
| Assistants built since | None. The count has stayed at one |
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.
