The challenge
A customer sits down on a weekend to work out whether a particular deposit account suits a seasonal business. A branch officer fields the same handful of product questions between appointments and would rather spend that time on the appointments. A marketing lead owns the website, reads what the assistant tells people, and wants it to sound like the bank on every answer rather than on most of them.
Rates move. Products get added and retired. Phrasing that read well against one version of a page starts to read oddly once the page has been rewritten. A bank of this size cannot run a release process to fix a sentence, so the practical question at purchase was a narrow one. When somebody reads the assistant’s answer and knows a better way to say it, how long does that take to change, and who has to be involved?
The solution
One assistant, grounded in the published site. The bank’s own product and account material is crawled as it stands and recrawled as it changes, so the corpus stays current as a by-product of running the website rather than as a separate task on somebody’s week.
Wording the bank owns. The instructions were written in-house rather than accepted as a generic persona, and they have been revised 224 times, each edit landing on the assistant already in service.
Citations on the answer. A visitor can see which page a statement about a product came from, which is what keeps a public statement about a regulated product tied to material the bank has committed to in writing.
Three agentic actions. Page awareness, conversion and lead capture run on top of the reply, so a good answer can end somewhere other than the end of a paragraph. They were added as the case for each one appeared rather than all at launch.
What it did not take is part of the design. No core banking integration, no customer data in scope, no second copy of the bank’s content, and no release process between a weak sentence and a better one.
The capability that decided it
How an assistant answers is copy, and copy gets better by being edited by someone who reads it. This bank has revised its assistant’s instructions 224 times across twenty-eight months, one revision every four days on average, and every one of them landed on the assistant that was already answering visitors.
Where changing an answer means a ticket, a sprint or a vendor, the wording stops improving on the day the project ends. Here it is a text field the marketing lead can open, which is why the answers have kept pace with a bank that has not stood still either. No other bank assistant in our published case studies has been revised as often.
Every one of those revisions is small. A clearer sentence about a fee, a better way to name a product, a phrasing that matches a page the bank has just rewritten. Two hundred and twenty-four of them stack up into an assistant that sounds like this bank in particular, which is not a thing any vendor can hand over on the first day.
The results
The two day counts tell the story between them. Across the whole deployment the assistant has carried a question on 659 of 852 days. In the most recent year of the record it has carried one on 319 of 365. The opening stretch was sparse, at a handful of active days a month, and every month since has run between nineteen and thirty.
Volume is modest by design, at 3,723 questions, because this is one assistant on one website rather than an estate. Set the 224 revisions beside the day counts and the shape is clear. The bank runs this assistant rather than hosting it.
One limit is worth stating plainly. The 224 revisions are evidence of effort, not of improvement. The record shows the bank worked on its answers steadily. It holds nothing about whether any single edit made the next answer better, and nothing here claims that it did. No answer-quality classification exists on this deployment, so no answer rate appears anywhere on this page.
Why it worked
Editing is not deploying. There have been 224 changes to what the assistant says, none of which needed a rebuild, a reload or anyone outside the bank.
The corpus follows the site. Grounding in the live published material means the bank updates the assistant every time it updates a page, with no second copy to keep in step.
Capability could be added later. Three actions now run on the answer, put in as the case for each one appeared rather than all at launch, so the deployment grew without being rebuilt.
The risk surface stayed small. The assistant serves anonymous visitors, reaches no account system and holds no customer data, which keeps governance within reach of a bank of this size.
| Measure | To date |
| Time in production | 28 months, one assistant |
| Questions asked, raw count | 3,723 |
| Days carrying a question | 659 of 852 |
| Days carrying a question, most recent 365 | 319 |
| Revisions to the assistant’s wording | 224 |
| Agentic actions now running | Three, added over time |
| Assistants in production | One, throughout |
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.