At a glance
| Industry | Community banking, US East Coast |
| Size | Around 250 staff across mortgages, construction and business lending, and everyday banking |
| Objective | Answer the rate, product, account and access questions that arrive on the website outside branch hours, without drifting into financial advice |
| Key use cases | After-hours and 24/7 customer self-service, rates and product terms, account comparison, branch and hours information, online banking help, routing customers to the right application |
| Solution | CustomGPT.ai. One live-chat website assistant on the public site, grounded only in the bank’s published content |
| Owner | Digital marketing, with no IT project team and no core banking integration |
| Reliability | 600 days live with customers asking, zero support tickets in two years |
| Shared | Anonymously, at the customer’s request |
The challenge
A customer at eleven at night, working out whether a certificate renews on its own or whether she has to do something. A small-business owner comparing two business checking accounts on a Sunday, deciding which one to open in the morning. Somebody who cannot get into online banking and does not want to wait until Monday to ask why.
The bank has the answer to all three, published and approved, sitting on its own website. The problem is that the website was built to present products rather than to answer questions, so finding the answer means knowing which page it lives on. The questions were arriving, and the site was replying with a navigation menu.
The solution
- It answers, and it does not advise. A bank can tell you what a product does and what its terms say. It cannot tell you what to do with your money. Retrieval limited to published material keeps the assistant on the right side of that line by construction, rather than by hoping a general model stays there.
- Rates and terms are current, or absent. A rate quoted from last quarter is worse than no rate at all, so the corpus tracks what the bank has actually published and the assistant declines when the published answer is not there.
- Safe in front of anyone. A public assistant on a bank’s website takes every question, including ones about specific accounts it must never attempt to answer. The refusal behavior matters as much as the answers.
- Marketing runs it. The persona has been rewritten 58 times by the bank’s digital marketing team since launch. A community bank has no engineers sitting idle, so if changing the wording needs a ticket and a release, the wording stops being changed and the assistant drifts out of date.
The capability that changed it
An assistant that only replies is a better FAQ. The change came when actions were switched on, so a question could end somewhere useful: the right product page, the right application, the right branch.
Roughly one automated action now fires for every question asked. For a bank, that is the difference between deflecting a call and starting an account opening at eleven at night.
The results
| Measure | To Date |
| Consecutive days with a question answered | 330 |
| Questions this year against the year before | 5.6× |
| Questions answered to date | 14,586 |
| Answer rate, last six months | 91.6% |
| Support tickets raised in two years | 0 |
Set the last twelve months against the twelve before and the assistant handled 5.6 times as many questions, climbing from roughly 180 a month to roughly 1,000. Nothing was relaunched and no campaign ran behind it. People used it, came back, and told others it was there.
The streak is what makes that credible. A question has been answered every day for 330 days and counting, weekends and holidays included, with nine in ten resolved from the bank’s own published material.
Then the number nobody asks for: zero support tickets across two years of continuous operation. The bank has renewed on enterprise terms and is into its second enterprise term.
Why it worked
- Grounding that suits a regulated balance sheet. Own-content-only retrieval with honest refusal. A bank needs an assistant that declines cleanly more than one that always has something to say.
- A marketing team can run it alone. Fifty-eight persona rewrites by the people who own the brand voice, with nothing between a decision and the live site.
- The answer leads somewhere. Roughly one automated action per question, routing customers into the right application instead of stopping at a paragraph.
- It keeps working without being looked after. Two years, 600 active days, zero support tickets, no migration, no rebuild.
Frequently asked questions
What is an AI customer service chatbot for a bank website?
It is an assistant embedded in the bank’s public site, grounded only in what the bank has published: rates, product terms, account types, branch information and online banking help. A customer asks in plain language at any hour and gets an answer drawn from that material, with anything outside it refused.
How accurate is it?
This deployment resolved 91.6% of questions from the bank’s own published material over the last six months. The remainder are declined rather than guessed, which is the behavior that makes a public banking assistant deployable at all.
Does it give financial advice?
No. It reports what a product does and what its terms say, and it cannot tell a customer what to do with their money. That boundary is enforced by limiting retrieval to the bank’s published material rather than by instructing a general model to stay in its lane.
Does it connect to core banking or see customer accounts?
No. There is no core banking connection and no customer data flows through it. The assistant answers published product and rate questions and must never attempt an account-specific question, which is part of why the refusal behavior was a first-order requirement.
How does it keep rates current?
The corpus tracks what the bank has actually published, so a rate that is no longer on the site is no longer answerable. A rate quoted from last quarter is worse than no rate at all, so the assistant declines when the published answer is not there.
Who runs it day to day?
The bank’s digital marketing team, with no IT project team behind it. They have hand-written 58 revisions to the assistant’s persona since launch. That ownership model is the reason the deployment is still current two years on.
What happens outside branch hours?
That is the point of it. Questions arrive at eleven at night and on Sundays, and the assistant answers them from published material rather than showing an opening time. It has answered on each of the last 330 consecutive days, weekends and holidays included.
Can it do anything beyond answering?
Yes. Actions route a customer to the right product page, the right application or the right branch, and roughly one action now fires per question asked. For a bank, that turns a deflected call into a started application.
How much support does it need?
Zero support tickets have been raised across two years of continuous operation. For a small team that owns the assistant outright, an assistant that never needs rescuing is worth as much as one that answers well.
Does a small bank need an integration program to deploy this?
No. There was no core banking connection, no custom build and no standing internal headcount against this deployment. One assistant on the public site, owned by the function that owns the brand voice.
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.
