How a National Accountancy Body Grew Its Assistant 50% Year on Year and Reached a Third Term

At a glance

The challenge

A student halfway through the qualification, working out which pathway now applies to her. A practicing member checking what this CPE cycle requires of him, and whether the course he just booked counts. An employer verifying a credential before an offer goes out.

None of them knows the term the body files that answer under, because a national body publishes for the record. Its material is organized the way a regulator organizes things rather than the way a member searches. The answer is published, it is right, and it still generates an email to the team.

The solution

The evaluation was run out of IT rather than marketing, with the digital experience team alongside it. That shaped the criteria, because the people judging the assistant were the ones who would carry the consequences of a wrong answer.

• Answer only from published material. When a question touches CPE obligations or eligibility for a practicing certificate, a confident wrong answer becomes a compliance problem with the member’s name on it. Honest refusal was a stated requirement, tested on the content with consequences before anything easier.
• One corpus, four audiences. A prospective member, a student mid-qualification, an employer checking a credential and a practicing member chasing a technical point all arrive at the same site. One assistant handles all four, with no login between them and a published answer.
• Hold the advice boundary. A professional body can tell you what its requirements say. It cannot tell you what to do about your client. Own-content-only retrieval keeps the assistant on the first side of that line by construction.
• Run it without a project team. No integration program, no standing internal headcount, no rebuild at renewal. A small IT function owns it alongside everything else it does.

The capability that decided it

A national body’s authority is partly a matter of language, a register that took decades to establish. Hand it to a vendor template and the assistant stops sounding like the institution within a sentence.

So the persona is written inside the product by the body’s own staff and has been revised dozens of times since launch: lead with the direct answer, cite the source, say plainly when the published material does not cover the question. When the tone needs to change, it changes that afternoon, with no engineering ticket and no vendor in the loop.

The results

Set the most recent run of months against the same months a year earlier and questions asked are up 50%, with current weekly volume a quarter above the six-month average. Most deployments peak in their first weeks and decay. This one has done the opposite, and nobody campaigned for it. At its busiest stretch a question arrived every day for twelve consecutive weeks.

Behind that sits the number that makes it possible. Four questions in five are answered outright from the body’s own material. Traffic only compounds when the answers hold up.

Two terms are complete and a third is in motion, with no relaunch, no migration and no second procurement.

Why it worked

• Grounding suited to a regulator’s content. Own-content-only retrieval with honest refusal. Where content carries obligations, a model that declines to guess beats one that is usually right.
• A voice the institution controls. The persona is written and revised by the body’s own team, in the product, without an engineering ticket.
• Placed where the questions already arrive. Live chat on the open site, in the moment somebody is stuck, rather than a search box on a page nobody visits.
• It keeps running without being rebuilt. A small IT function owns it outright, which is why there is a third term rather than a fresh evaluation.

Frequently asked questions

What is an accounting standards research AI tool?

It is an assistant grounded only in a professional body’s own published standards, guidance, qualification and CPE material. A member, student, employer or member of the public asks in plain language and gets an answer drawn from that published corpus with the source cited, rather than a list of links or an answer from the open web.

How accurate is it?

This deployment answers 81.9% of questions from the body’s own published material. The remainder are refused rather than improvised, which is the behavior that makes the rate usable: when the corpus does not cover a question, the assistant says so.

Does it give accounting or tax advice?

No. It reports what the body has published and cites the source. A professional body can tell you what its requirements say and cannot tell you what to do about a client, and restricting retrieval to published material keeps the assistant on the correct side of that line structurally.

Can members and the general public use the same assistant?

Yes, and here they do. Members, students, employers and the public are all answered from the same published corpus, with no login between a member of the public and a published answer. Roughly 85% of questions arrive through live chat on the open site.

How does it handle standards that get revised?

The corpus stays indexed and current as standards are revised and courses are re-listed each term, so the assistant answers from today’s guidance rather than last term’s. Four million words are kept live this way.

Who writes the assistant’s voice?

The body’s own staff, inside the product. The persona has been revised dozens of times since launch, and a change takes effect the same afternoon with no engineering ticket and no vendor involved. A national body’s register took decades to establish, and a vendor template loses it within a sentence.

Can it capture inquiries as well as answer questions?

Yes. Most questions arriving on a professional body’s site are one step from a course booking or a membership application, so a resolved question can become a course inquiry or an application step instead of ending the exchange.

Does it need an integration program or a dedicated team?

No. There is no integration program, no standing internal headcount and no rebuild at renewal. A small IT function owns this deployment alongside its other work, which is the reason it reached a third term without a second procurement.

What happens to usage after the first year?

Here it grew. Questions asked are up 50% year on year on matched periods, with current weekly volume a quarter above the six-month average. That is the opposite of the usual curve, and it is worth asking any vendor whether their customers’ year-three volume sits above or below where it started.

How long before it is answering real questions?

The build here started with the questions staff already answered by hand twice a week, tested grounding and refusal on the content that carries obligations, then went live in chat on the open site. There was no relaunch and no campaign afterward.

Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.