Your clients are taking financial questions to AI first. Here is what the evidence says, what a machine genuinely does better, and what remains yours alone.
Nobody wins this argument by claiming AI is bad at financial questions. It is good at them, and pretending otherwise costs the profession the room. Today's models are strong reasoning engines, and they are now the first place a great many small business owners go. The problem is that a machine answers the question it was given. When the question is wrong (and the client’s question often is), it returns a confident, well-structured answer to the wrong question. And the client has no way of seeing what was missed.
Part One sets out what the research actually shows, with every figure tied to its source, sample, and date below (or at the back, if you're reading this as a PDF). Part Two concedes what a model does better than a person, because the rest of the argument only earns a hearing once that is on the table. Part Three is the ten things a model still needs an accountant for. Part Four is what a firm can do about it.
Five studies published between March and July 2026 measure the same behaviour from different angles. They may disagree on the size of it, but they agree on the direction. Four of the five are vendor-commissioned, which is addressed openly in Note 9.
Of 500 UK SMEs who all already work with an accounting firm, seven in ten say they always or often act on AI-generated financial, tax or business advice before they consult that firm.1 One in twenty rarely or never does.
The specific things they take to a machine are: answering tax questions, financial planning queries, business strategy and triaging day-to-day accounting issues.1 That list deserves reading slowly, because it is the same list that many in the profession have spent a decade describing as its advisory future (compliance does not appear on it anywhere). The accountant is still in the process. They just arrive afterwards, to validate what has already been decided.
Figure 1. Acting on AI advice before consulting the accountant
A thousand UK sole traders were surveyed in April 2026, as Making Tax Digital for Income Tax arrived for qualifying income above £50,000.2 A quarter had used an AI platform for guidance on the new regime, and one in five now uses AI regularly for tax and accounting support.
Two figures in that study matter more than the adoption rate. The first is that 88% say they trust the answers. That's with tax and cashflow issues. The person asking has the least ability to spot a wrong answer, which is exactly where confidence is running highest. The second is that 26% have used AI to get a second opinion on advice they had already been given.2
A quarter of sole traders are using a chatbot to check their accountant's work. Set that beside the 91% who have considered changing firms in the past year, and the churn mechanism is laid out end to end.1
Figure 2. UK sole traders and AI for tax and accounting
Speed beats cost, 39 to 32. Firms are losing the first question on response time.
Published adoption rates for small business AI use run from 6% to 80% within the same six months, and none of the studies is wrong. They count different things, so any figure quoted without its definition attached will be contradicted by the next person in the room.
The wide numbers count anyone who has used a free chatbot or clicked an AI button inside software they already pay for (which by now is almost everyone). The narrow ones count deliberate deployment, paid commitment, or genuine embedding into how the business runs. For the advice question the wide measures are the ones that matter. A client needs nothing more than a browser tab to take a tax question elsewhere.
Figure 3. Small business AI adoption, by what was actually measured
Two structural findings sit underneath those bars and both point the same way. In Ireland a third of micro-businesses under ten employees use AI for nothing at all, against 7% of firms with 50 to 250 people. Firms trading more than sixteen years are three times more likely to use no AI than firms trading one to five.3 In the United States, paid adoption skews young, with 28% of businesses led by someone aged 18 to 34 paying for AI tools against 13% of those led by someone aged 55 to 64.4
So the established client paying the largest fee is the least AI-native, while the newest client, often paying the least, arrives already holding an answer. Those are two different conversations, and most firms are only having one of them.
Marketing leads AI adoption in every dataset without exception. The signal worth watching is the finance function, and it moved in January.
Intuit's tracking runs on 34,000 survey responses and payment records from 5.3 million businesses, with methodology led by an economist at the University of Chicago. It found bookkeeping entering the top three tasks where US businesses report using AI, alongside marketing and customer service.4 At the bottom of that list sit employee management, product management and legal (the places where a decision cannot easily be reversed).
The pattern is consistent. AI goes where the output is visible and an error can be caught before it does damage. Bookkeeping arriving in that top three matters because it breaks the pattern. Errors in the records do not stay contained.
Depth remains shallow, and that is the reassuring half of the picture. Around one in ten US small businesses has ever paid for a dedicated AI tool, though 86% of those who paid in 2024 were still paying in 2025.4 Wide shallow usage, a small committed core, and near-total retention once someone crosses the line into paying. Anyone who has sold advisory will recognise that shape.
Clients going to a machine first are not doing it to save money. They say so directly, and their stated willingness to pay is remarkable.
Among those 500 UK SMEs, 92% would pay more for a fuller range of advisory services, and the average business puts the figure it would accept at 16% above what it currently spends.6 Just under half, 47%, would go as far as 25% more. Those are three views of one distribution rather than three competing claims. Average annual spend sits at roughly £19,700, with the largest concentration of clients in the £10,001 to £25,000 band.6
Then the operational number, which is the one to act on. 38% said they would take up additional services if their accounting firm matched the speed and responsiveness of the other providers they deal with.6
Put the arithmetic together, because the numbers are more concrete than they first appear. Against average annual fees of £19,700, a 16% uplift is roughly £3,150 per client per year, so a firm with 150 clients of that size (and plenty carry more) is looking at somewhere near £470,000. That lands within touching distance of the £463,000 headroom Ravical modelled by a different route.
One caution before that number goes anywhere near a partner meeting. The uplift is not evenly spread. Clients of firms with two to ten staff spend an average of £12,559 and would accept only 9% more, which is nearer £1,130 a head.6 Run the calculation on your own book, not on the national average.
Figure 4. Willingness to buy more from the firm they already use
The most useful finding in the whole set has nothing to do with AI. Asked whether their own firm offers a given advisory service, most businesses did not say no. They said they were not sure.10
Nobody had told them. Asked why they buy advice elsewhere, the top two answers came in level at 35% each: the firm never offered it, and the firm only handles compliance for them.10 Price barely featured in the responses at all.
The leakage that follows is measurable, and it is worst where firms are smallest. More than a third of UK businesses buy advisory services from a provider other than their main accounting firm. Among sole practitioners that is 42%, and among firms of two to ten staff it reaches 48%, the highest in the survey. At Top 50 firms it falls to 12%.10
Then the number this whole note turns on. Of the businesses already buying elsewhere, 94% would bring that work back if their own firm delivered the same quality.10 That is the invitation, and it is sitting in books that already exist.
There is peer-reviewed support for the mechanism, and it is the only evidence in this document that is not vendor-commissioned. In a study of twenty small firms interviewed at length, owners did not buy advice from their accountant because they did not believe it would be worth the cost, while the few who did buy it found it valuable.11 Twenty interviews is a mechanism rather than a measurement, and it points the same way as the 500. The barrier sits in what the client believes. The service was never the problem.
The offer was never made. That is a model failure, not a failure of the people in it.
The growth conversation is being bought elsewhere as well, and not only from a machine.
Around 45% of global business coaching engagements now come from small and medium businesses, in a market worth roughly $113bn.10 Owners are paying someone else for the conversation about where the business is going, while the person holding the actual numbers sends reports.
Be precise about what that does and does not show. Clients have not decided they prefer a coach to their accountant, and the trust surveys point the other way firmly. The behaviour has moved while the trust has stayed put. That is more uncomfortable than outright rejection, because it means the relationship is still there for the taking.
Lost fees are the visible cost of this shift. The hidden cost is the lost signal.
People ask a machine the things they are embarrassed to ask a person. Whether the business is actually failing. Whether the question they are about to ask is a stupid one. Whether they can pay themselves this month. Whether the numbers mean what they are afraid they mean. There is no judgement on the other side of the screen, no clock running, and no sense of wasting anybody's time (theirs or yours).
Every one of those questions is diagnostic gold. The firm never sees a single one of them. A client who quietly resolves their own worry at eleven at night arrives at the next meeting looking fine.
You are not just losing the work. You are losing the early warning.
Four of the five studies are vendor-commissioned. Ravical sells AI to accounting firms, Starling launched an accounting product alongside its research, and Intuit, Google and Sage all sell into the market they surveyed. None of that makes the findings false, and Intuit's payment-record layer is genuinely strong evidence, but each has an interest in its own headline.
Official statistics are consistently more conservative. The UK Office for National Statistics put business AI use at 23% in late 2025, the Irish CSO above 20% of enterprises for 2025, and the US Census Bureau between 17% and 20% through spring 2026.8 That is a measurement gap rather than a disagreement about what is happening, and a 70% or a 77% only means anything alongside the line that defines it.
Nothing in this evidence supports the idea that clients would choose a business coach over their accountant. It is the most tempting misreading in the set, and Note 7 sets out what the coaching figures do and do not show.
The Ravical 70% appears in two framings across their releases. Once as acting on AI advice before consulting the accountant, and once as acting on it without first checking with the firm, with a related version reported at 71%.6 Those are different claims about different behaviour. This document uses the first throughout, from the Censuswide base of 500 SMEs with fieldwork in May 2026.
What the same research declines to claim matters just as much. The publisher treats the rise in AI use as a trend running alongside the shift of advisory work to other providers, not as its cause.10 This document keeps that separation. Part One sets out two pressures on the same relationship, and neither is offered as the explanation for the other.
This section exists because a document that skips it deserves to be ignored. Any accountant who has actually used these tools knows what follows, and a firm that cannot say it out loud will not be believed on anything else.
Which raises the obvious question. If the answers are that good and that fast, where is the exposure?
Models are at their strongest where the question is well specified and the answer can be checked. They are at their weakest where the question itself is wrong, because a model answers the question it was given, fluently and with complete composure.
The error is made before the model ever sees it, and nothing about the quality of the model catches it. The answer comes back well structured, numerate, and correct on its own terms. The person who asked has no way of seeing what was missed.
The second failure catches the client who asked well. Models affirm the user's decision around 50% more often than a human adviser would, including where the judgement is poor.7 So the owner who frames the question properly, and brings the right numbers to it, still receives an answer weighted towards agreement. Nothing in that exchange carries a stake in the outcome, so nothing in it pushes back. That is a problem with advice given outside any method, and it lands the same way whether the adviser is a machine or a person who wants the meeting to go well.
Put the two together and 88% trust becomes the figure to worry about, though it has nothing to do with the technology being poor. Confidence runs highest in exactly the domain where the person asking is least able to spot an error.
An owner asks whether the business can afford a salesperson at £38,000. Back comes a competent answer: gross margin, employer costs, a breakeven month, a sensible caution about the ramp period.
What nobody asks is whether revenue per employee is already below the level at which the last two hires paid back. Or whether the cash gap between the hire and their first commission survives the next VAT quarter. Or whether growth has stalled because the owner is the ceiling, in which case a salesperson is the wrong answer to a question about the owner.
The model answered what it was asked, competently. Even if it could have known the question was wrong, it had no standing to say so.
One thing to expect before the list. Any accountant who uses these tools every day will read all of that and recognise none of it, because they argue with the answer. They ask again, they say that is not right, they push until the thing gives them something better. The failure never surfaces, so it is easy to conclude it was never there.
The client does not do that. They ask once, late, and take what comes back. The habits that make these tools safe are the same habits that make an adviser good, and nobody has taught them to the people now relying on the answers.
Read as a scorecard against AI, this list is a competition nobody needs. Read properly, it is a specification. It describes what an accountant contributes to get a good outcome out of these tools, which is why it works equally well as a description of good advisory practice.
Behaviour varies between models and is changing quickly, so treat the specifics below as the shape of the current failure classes rather than a permanent account of any particular product.
A model returns the most plausible answer available to it, and it has no dependable way to mark the difference between knowing something and constructing it. It will not tell you which one you have just received.
Asked to finish something it is perfectly capable of finishing, a model will often propose that you do the next part, or offer to come back to it later. An unsupervised client accepts that. The job stops at three-quarters done.
Give a model a defined outcome, and it can judge whether it has arrived. Without one, it stops at a point that reads as complete. Looking finished and being finished are not the same thing, and the client cannot tell them apart.
The question a client brings is frequently the wrong one. Getting to the right one is most of the job.
The first answer is where the work starts, and a good adviser treats it that way. Worth admitting that this one is aspirational for plenty of firms too.
The affirmation bias set out in Part Two is a human failing as well, which is exactly why challenge has to be built into a method rather than left to character.
Not only remembering what was agreed last quarter, but raising it unprompted at the moment it becomes relevant again.
A model sees the fragment it was handed. We see the payroll, the pipeline, the bank, the tax position, the last three years, and the family sitting behind all of it.
Regulated, insured, professionally liable, and reputationally exposed if we get it wrong. Skin in the game changes what advice is worth.
Two accountabilities, and a machine has neither of them. Accountable for the advice given, and holding the client accountable for what they said they would do.
Points two and three double as operating instructions, worth teaching to every person in the firm who touches these tools.
Those three habits are most of the distance between a mediocre output and a good one. They are also precisely what an owner alone with a chatbot will never do, which is the whole point.
Six moves, all of them available now, none of them requiring a technology decision. The evidence in Part One points at each one.
Three announcements in six months tell you where this is heading. Intuit with OpenAI in November 2025, Xero with Anthropic in March 2026, and Intuit with Anthropic again in April 2026. The accounting software layer is embedding AI into the place your client already keeps their numbers. So the question is no longer whether a client will ask a machine. It is whether that machine will have their actual figures in front of it when they do.
Nobody beats four seconds, so stop competing on that and compete on certainty instead. A published standard removes the uncertainty that sends a client elsewhere: same-day acknowledgement, a stated turnaround for straightforward questions, and honesty that the consequential ones take longer because they involve thinking.
38% said they would buy more if their firm matched other providers on responsiveness. That makes this a revenue decision, not a service one.
Your clients are already pasting management accounts, payroll runs and customer lists into free tiers. Privacy and security is the single largest barrier to deeper adoption at 36%, so they are uneasy about it, and nobody has helped them.
One page, deliverable this month, and it opens the conversation in the next move.
"What have you asked AI about the business in the last month?" is the highest-yield question available in a client meeting right now. Asked without a flicker of judgement (and the flicker is the whole risk), it recovers the signal described in Note 8 and tells you what the client is actually worried about, which is rarely what is on the agenda.
The evidence says your long-established client is the least AI-native, while the newer, younger-led business arrives holding an answer. One group needs bringing along. The other needs a better second opinion than the one they already got. Treat both the same way, and you serve neither.
Xero's research found UK accountants using AI deliver results 31% faster.9 If your client is accelerating and your firm is not, the responsiveness gap widens without anyone choosing it. And a partner who uses these tools daily can tell a client exactly where they break. That is a conversation nobody can fake.
A fast answer to an unstructured question is still an unstructured answer. What holds ground against a machine is method: a defined set of numbers reviewed on a rhythm, a target the client has agreed on, and a record of what was decided last time. Sycophancy is a framework problem rather than a technology problem, and the fix is the same whether the adviser is human or not.
That is the Numbers Mentor position. Support through the numbers and challenge on the decisions, in equal measure. The accountant who only ever supports is replaceable by anything fluent. The mentor who supports and challenges is not.
An owner putting a question to a chatbot at eleven at night, with no financial model behind it and no target to check the answer against, is the drift with better grammar. Small businesses do not collapse. They erode. A confident answer that nobody challenged speeds that up.
So the second opinion is the position now, whether a firm chose it or not. What is still open is whether it turns out to be a better one than the answer the client is already holding.
The Second Opinion · Clarity HQ · July 2026
Figures as published at 27 July 2026.
Prepared for accounting firm leaders. clarity-hq.com