HMRC digital assistant: 20% of tax agents rated it good — and all three groups described the same fault

The short version

• Three groups, one tool, a 36-percentage-point spread. 20% of tax agents rated the HMRC digital assistant good, against 45% of small businesses and 56% of mid-sized businesses.
• The heaviest users rate it worst. 37% of agents had used it, against 15% of mid-sized and 10% of small businesses.
• All three groups described the same failure — it sends you back to guidance you have already read and not understood.
• ⚠️ “digital assistant” is a blend of a chatbot and a human adviser in HMRC’s own definition, and the report never splits them. None of these numbers is a bot score.
• 🚩 The report is from 9 July. What is new is the data. Fieldwork ran September to November 2025.
• 🚩 The spreadsheet labelled “Small Businesses” contains the mid-sized data, so the small-business figures cannot currently be checked against it. We verified that four ways.

What the HMRC digital assistant actually is

If you have ever opened the chat window on an HMRC page, that is it. What most people do not realise is that the phrase covers two different things. HMRC’s own glossary is precise about it:

“HMRC’s digital assistant includes customers using the chatbot (an automated natural language processing tool to provide a response) and webchat (an adviser-led response).”

So behind one chat window there is software and there is a person, and every published satisfaction figure for “the digital assistant” is a blend of both. Keep that in your head for the rest of this piece. It is the single most important line in the report and it is in the glossary, where nobody reads.

What its own users said

Every respondent who had contact with HMRC was asked to rate their experience on a scale of 1 to 5, where 5 is very good. Here is how the digital assistant did, by group:

Rows do not always total 100 because of rounding and, where the published data tables show one, a small “Don't know” share — 1% for mid-sized businesses. Usage is the proportion of each group who said they had interacted with HMRC through the digital assistant in the previous 12 months.

The distance between the top and bottom row is 36 percentage points — not 36%, and the difference matters if you are ever tempted to quote it. Same tool, same year, three populations.

Notice which way round it falls. The group that used it most rated it worst by a distance: more than half of tax agents called their experience poor. That is the opposite of the usual pattern, where heavy users are enthusiasts. The likeliest explanation is in the questions they bring. Agents do not open a chat window to ask what the VAT threshold is; they open it when something has gone wrong with a specific client’s specific return, and that is precisely the question a general-purpose assistant cannot take.

For context, the same three groups rated HMRC’s ordinary web pages far higher — 55% of agents and around 70% of businesses called those good — and rated the telephone helplines lower again. Among small businesses the report puts the digital assistant and the phone almost level: “Telephone helplines and the digital assistant service were viewed in a less positive light, with fewer than 1 in 2 users giving a positive rating (43% and 45% respectively).”

The fault all three groups described

The quantitative scores differ. The explanation does not. In the follow-up interviews, agents, small businesses and mid-sized businesses independently described the same loop.

On agents, the report says there were “consistent reports of encountering repeated barriers when trying to obtain the information they needed, directing them to webpages they had already tried without success”. On mid-sized businesses, the chatbot “was seen as too generic and customers reported being redirected to guidance pages they had already tried”. And one respondent said the whole thing in a sentence:

“If [the chatbot] can’t answer your question, it points you to a page on the website that includes [the key word] rather than answering your specific query. It sends you to a guidance page you have already been on and didn’t understand which is why you were asking the question.”

A small business put it more wearily: “When I use [the digital assistant], I find you don’t frame the question the right way, and it takes me around in a circle, so I have to speak to a human anyway.” An agent was blunter still: “It’s just nonsense. You might as well have picked the phone up and waited in a queue.”

There is a structural point hiding in there, and it applies to every chat widget on every website. Nobody opens a chat window first. By the time someone is typing a question, they have already looked at the page and failed. So a reply that returns them to a page is not a wrong answer, it is a non-answer — it is the loop closing. The report’s own headline insight concedes the underlying problem in flat official prose: “difficulty locating clear and relevant information - particularly for complex or niche scenarios - is a persistent barrier affecting customer experience.”

And it has a measurable cost. Asked why they had phoned rather than gone online, 19% of small businesses said they “had first tried to resolve their issue online but were unable to do so”. Roughly one in five calls is a website that did not work.

The number under the number

Go back to the glossary. The digital assistant “includes” the chatbot and webchat with a human adviser. The report presents one satisfaction figure for the pair and never separates them.

That means the honest reading of the 56% figure above is: 56% rated good a mixture of a bot and a person, in unknown proportions. The bot alone might be better than 56%. It might be a great deal worse, with the human adviser carrying the score. Nothing published tells you, and the qualitative sections — which do name the chatbot specifically — are uniformly negative in all three groups.

This is worth noticing because it is the standard shape of chatbot statistics generally. A deflection rate counts conversations that did not reach a human, not conversations that ended well. A satisfaction score usually includes the sessions a person rescued. If you are being sold an AI chat product on a number, the first question is not how big the number is. It is what was inside it. We wrote about the difference between a bot that waits to be asked and an agent that actually does something in AI agent vs chatbot, and this survey is the best public evidence we have seen for why the distinction is not marketing.

The finding that should worry you before you add one

The most useful sentence in the whole report, for a small business, is not about HMRC at all. Explaining why small businesses so rarely used the digital assistant, the report says they “actively avoided them, citing negative experiences with similar services on other websites, which appeared to shape expectations and reduce trust in their usefulness”.

Read that again. People had decided about HMRC’s chat window before they ever opened it, on the evidence of chat windows elsewhere. Your chatbot inherits the reputation of every bad one your customer has met. You do not start at neutral, and no amount of friendly copy in the greeting bubble changes that; only being useful in the first exchange does.

Which gives a blunt test before you install anything. Can it do something a page cannot — look up an order, check a balance, book a slot, take a payment reference? If yes, it has a reason to exist. If the honest answer is that it finds the right page, then remember where your visitor has just come from.

Can you check any of this? Partly

HMRC did the right thing on 17 September and published the crosstabs. The change note on the page reads, in full: “Data tables for the Agents, Small and Mid-Sized Businesses Customer Survey 2025 have now been added.” Three spreadsheets went up — Agents, Small Businesses, Mid-Sized Businesses.

We downloaded all three on the evening of 19 September 2026, taking the links from the page itself rather than guessing at addresses. The file linked as “Small Businesses Customer Survey 2025: data tables” is the mid-sized business file. We checked that four ways:

• Every one of its 57 data tables prints “Mid-Sized business survey 2025” at the top; the 58th sheet is a contents page. The Agents file correctly prints “Agent survey 2025” on its tables.
• Its all-respondents base is 1,599. That is the mid-sized sample. The small-business sample is 2,200.
• Its channel percentages match the report’s mid-sized table (web pages 80%, business tax account 66%, telephone 51%) and not the small-business one (65%, 50%, 28%).
• All 58 sheets, matched by name, are identical cell-for-cell to the file correctly labelled Mid-Sized. The two files differ as files — 339,482 bytes against 340,384 — but only in formatting, metadata and the order in which two of the tables are stored.

That is an observation about three files, not an accusation. We downloaded all three again on the morning of 23 September, the day this was published, and each was byte-for-byte identical to the 19 September copy, so nothing had been corrected by then. It may well be fixed by the time you read this. But while it stands, it has a consequence worth stating plainly. The mid-sized digital-assistant rating can be checked: the question is asked on a 1 to 5 scale, the base is “Base: All those who had contact with HMRC, excluding those who responded 'Not applicable'”, and 227 people answered it — about 14% of that sample, which squares with the 15% who said they had used it. The same check cannot be run on the small-business 45%, because that file was not published.

Which is its own small lesson, and not only for HMRC: publishing the workings is what turns a figure into something a reader can stand on. We only know the mid-sized number is solid because we could open the sheet.

What argues the other way

• The fieldwork is roughly a year old. Interviews ran September to November 2025 and the follow-ups to March 2026. HMRC may have changed the assistant since; this is a photograph, not a live feed.
• 45% and 56% are not disasters. Only the agents’ 20% is genuinely bad, and agents are professionals arriving with the hardest cases. For a general audience asking simple things, roughly half rating it good is a respectable result for any self-service channel.
• The blend cuts both ways. We have said the human adviser may be propping the score up. It is equally possible that the chatbot performs well on simple queries and the low scores come from complex ones that were always going to end on the phone.
• We build AI agents for a living, so treat our reading of a survey about a struggling chatbot with the scepticism it deserves. That is also why we have quoted the report at length and linked it: the figures are checkable without us.

What to do about it on Monday

1. Open your own chat widget and ask it your three most common awkward questions. Not the easy ones. If all three answers are links to pages, you have the HMRC problem.
2. Find out what your score is actually measuring. If your supplier reports satisfaction or deflection, ask whether human conversations are inside the number, and whether a session that ended with the customer phoning counts as deflected.
3. Count how many phone calls begin with the caller saying they tried the website first. HMRC’s equivalent figure was 19%. Yours is knowable by asking for a fortnight.
4. Give it one job it can finish. One lookup, one booking, one status check. A narrow assistant that completes something beats a broad one that redirects.

Sources, and what we checked

• HMRC, Agents, Small and Mid-Sized Businesses Customer Survey 2025 (HMRC Research Report 868, carried out by IFF Research). Published 9 July 2026, last updated 17 September 2026.
• Both the landing page and the full report were fetched raw and tag-stripped on 19 September 2026, twice, several hours apart, and again on 23 September 2026, the morning of publication. The stripped text was identical across all three fetches on both pages.
• Every quotation from HMRC above was tested automatically against that stripped text and the spreadsheets, on 19 September and again on the morning of publication. Every one traced verbatim.
• Dates were confirmed against the page’s own change history and the GOV.UK content API (first_published_at 2026-07-09, public_updated_at 2026-09-17).
• The three data-table workbooks were downloaded twice on 19 September and again on 23 September, from links taken off the page, and compared by unzipping them and reading the cell values directly, sheet by sheet and matched by sheet name.
• Sample: 5,659 telephone interviews between September and November 2025 — 1,860 agents, 2,200 small businesses and 1,599 mid-sized businesses — plus 60 qualitative interviews between January and March 2026.