Back in January, a UKBF member called @Byzantium put up a complaint that'll sound familiar to a lot of business owners. He does his own books in Excel, properly - a chart of accounts matching what his accountants use for filing, every transaction categorised, accruals and prepayments done, opening and closing stock reconciled. It's not exactly a shoebox of receipts. And still, across a few companies: "the accountancy bill reaches well over £10k and the accountant thinks it should go up."
In August, @GNewland posted a version of the same frustration from the other end. Newer business, smaller numbers, and accountant quotes that "seem to be quite out of whack with the profits that we are generating so far." So he ran an experiment - fed a spreadsheet export of his business bank account into ChatGPT, told it which money was share capital and which was a loan, and asked it to come back with questions. The output looked, by his own admittedly untrained eye, good enough for a first filing.
We've been reading both threads closely at ANNA, and we should be upfront about why. ANNA is an AI company. We build automated bookkeeping and tax filing tools for exactly the businesses posting in that forum, and since joining forces with UKBF in June we've had the unusual privilege of reading small business owners argue about our own market without anyone performing for us. What came out of those two threads is more useful than most of the AI-and-accounting commentary we see, ours included. So here's what members actually concluded, and where we think that leaves us.
The distinction that does the real work
The most useful thing in either thread is a point @martin_shl made and @Ozzy , UKBF's founder, sharpened further. Members had started arguing about whether AI can do accounts. Turns out that's the wrong question - there are two very different things being described.
@martin_shl put the technical case plainly: "Generative AI is not a deterministic system and therefore should not be relied on." He also pointed out that some accounting packages are integrating machine learning now, which is a different thing entirely from feeding your numbers into a general-purpose chatbot.
@Ozzy picked that up:
"If 'Genrative AI' has been used then generating output is its purpose, it will generate output even if it is wrong. That is why I'd advise against using a basic ChatGPT et all to do your accounts, because it is a generalist generative AI tool.
However, Machine Learning (often incorrectly marketed as AI) and AI tools that have been guardrailed and specifically trained and programmed to perform bookkeeping tasks can be useful. Just make sure you use one that has been setup for that purpose."
That's the line. Not "AI good" or "AI bad" - the real difference is between a tool whose job is just to produce plausible-sounding text, and a tool that's actually built and tested for a specific accounting task. @DavidReed made the same point from the user's side, in one line that should be on a poster: "The annoying thing about AI is that a completely wrong answer can still look very convincing."
If you take one thing from either thread, take that. Give a general chatbot your bank statements and it'll produce a set of accounts - whether or not it's actually understood your director's loan.
The bit nobody in the marketing gets to
Two objections came up that you rarely see addressed in AI product copy, ours included, and both deserve more attention than they get.
First up: legal exposure. @fisicx , a moderator with no obvious axe to grind either way, asked @GNewland straight out: "Are you prepared to take on HMRC in court to defend your account if they don't like your filings? At least with an accountant you have some legal redress should they mess things up." @chenmichael made the companion point: "You're still responsible for the accounts as a director even if software or an accountant prepares them." And @UK Contractor Accountant added the enforcement context - Companies House now actively queries and rejects non-compliant or structurally flawed financial statements, and HMRC cross-references public Companies House figures against Corporation Tax returns.
Automation moves the work. It doesn't move the liability. Worth sitting with that if you're choosing between a £15 a month tool and a £1,500 a year accountant on price alone.
The second objection we honestly didn't expect the community to raise at all. @numbersrule , midway through an AI course run by his professional body, described a RAG assessment - red, amber, green - used to classify which tasks a public AI can be given, which need extra safeguards first, and which are off-limits entirely. His example of a red category activity: "Bank Statements being fed into a Public AI." And his reasoning wasn't about accuracy at all:
"It is Red because it is unacceptable from a GDPR perspective and from a Client Confidentiality perspective."
He was careful to say @GNewland is free to make his own choice, it's his own data. But the wider point stands, and @DontAsk agreed it was a good one. When you paste a bank export into a consumer chatbot, you're making a data decision as well as an accounting one. If you've got customer names in your transaction descriptions, and most small businesses do, that decision involves other people's data too.
What members said you can safely hand over
For all the pushback, the threads didn't land on "do it all yourself". They landed somewhere more practical.
@JamesUniform reframed the whole argument back in the January thread: "So for me the question is not really 'can software replace the accountant'. It is which parts are controlled enough to automate, and which bits still need judgement before the numbers are trusted." Then he listed the boring control checks that actually matter for supplier invoices - original invoice attached, supplier name matching the record, invoice number and date captured, net and VAT and gross checked, duplicates checked, approval evidence before payment, and a link back to the source document.
@UK Contractor Accountant agreed automation is transforming data entry and reconciliation, but drew the line at professional judgement - valuation of work in progress at year end, provisions for warranty claims or litigation, restructuring costs, accrued income determinations. @CLM Controller and @chenmichael landed on the same idea independently: use automation for the bookkeeping and analysis, then pay a professional for a review and the filing rather than for full monthly bookkeeping.
And @Daybooks , an accountant who runs client books in Excel and isn't remotely anti-technology, gave the warning that's stuck with us most: "Be careful: automated reconciliations are no reconciliation at all." A tick in a box isn't the same as someone actually checking.
Worth noting too - the alternative to both AI and an accountant already exists, and a few members use it. @DontAsk files his own accounts for roughly £100 a year for VT Accounts, a few pounds a month for Excel, and around £25 for filing software, on the strength of doing his own bookkeeping and reconciling every month. @tony84 pays about £1,500 for accounts and self-assessment and rates it mainly for getting questions answered in thirty seconds. Neither of those is an AI story. Both are legitimate answers.
Where that leaves us
We build automated bookkeeping. ANNA categorises transactions, calculates VAT, Corporation Tax and MTD Self Assessment, and files straight to HMRC. We think that work should be automated, and our own accounting and tax expert, Aftab Hussain, a Chartered Certified Accountant, has put a number on it: the routine work ANNA automates comes to around 90% of what accountants do for small businesses.
We'd point at the second half of what he said too, because it matters more: "It doesn't mean the end of accountants; it just signals a shift in their role, moving away from routine tasks towards more advisory services."
That's the same conclusion the forum reached by a completely different route. @Byzantium never argued for getting rid of tax advice - he argued he was paying thousands for reformatting he'd already done, and said outright he has no problem paying for good advice. @Paul Kelly ICHYB put the other side of it in one line: "a good accountant is not a cost to your business - they are an asset!" Those two positions aren't actually in conflict. They're both arguments for spending less on typing and more on judgement.
So here's the practical conclusion. Use purpose-built tools for the volume work - the categorising, the matching, the VAT arithmetic, the filing mechanics - and pick tools that are actually built and constrained for that job, not a general chatbot. Then spend what you save on the conversations a machine can't have with you: whether the salary and dividend split is right, whether that claim stacks up, what to do when something non-standard happens, and what the numbers mean for next year.
Automation's very good at the part of your accounts that's arithmetic. It's not the part you sign.
In August, @GNewland posted a version of the same frustration from the other end. Newer business, smaller numbers, and accountant quotes that "seem to be quite out of whack with the profits that we are generating so far." So he ran an experiment - fed a spreadsheet export of his business bank account into ChatGPT, told it which money was share capital and which was a loan, and asked it to come back with questions. The output looked, by his own admittedly untrained eye, good enough for a first filing.
We've been reading both threads closely at ANNA, and we should be upfront about why. ANNA is an AI company. We build automated bookkeeping and tax filing tools for exactly the businesses posting in that forum, and since joining forces with UKBF in June we've had the unusual privilege of reading small business owners argue about our own market without anyone performing for us. What came out of those two threads is more useful than most of the AI-and-accounting commentary we see, ours included. So here's what members actually concluded, and where we think that leaves us.
The distinction that does the real work
The most useful thing in either thread is a point @martin_shl made and @Ozzy , UKBF's founder, sharpened further. Members had started arguing about whether AI can do accounts. Turns out that's the wrong question - there are two very different things being described.
@martin_shl put the technical case plainly: "Generative AI is not a deterministic system and therefore should not be relied on." He also pointed out that some accounting packages are integrating machine learning now, which is a different thing entirely from feeding your numbers into a general-purpose chatbot.
@Ozzy picked that up:
"If 'Genrative AI' has been used then generating output is its purpose, it will generate output even if it is wrong. That is why I'd advise against using a basic ChatGPT et all to do your accounts, because it is a generalist generative AI tool.
However, Machine Learning (often incorrectly marketed as AI) and AI tools that have been guardrailed and specifically trained and programmed to perform bookkeeping tasks can be useful. Just make sure you use one that has been setup for that purpose."
That's the line. Not "AI good" or "AI bad" - the real difference is between a tool whose job is just to produce plausible-sounding text, and a tool that's actually built and tested for a specific accounting task. @DavidReed made the same point from the user's side, in one line that should be on a poster: "The annoying thing about AI is that a completely wrong answer can still look very convincing."
If you take one thing from either thread, take that. Give a general chatbot your bank statements and it'll produce a set of accounts - whether or not it's actually understood your director's loan.
The bit nobody in the marketing gets to
Two objections came up that you rarely see addressed in AI product copy, ours included, and both deserve more attention than they get.
First up: legal exposure. @fisicx , a moderator with no obvious axe to grind either way, asked @GNewland straight out: "Are you prepared to take on HMRC in court to defend your account if they don't like your filings? At least with an accountant you have some legal redress should they mess things up." @chenmichael made the companion point: "You're still responsible for the accounts as a director even if software or an accountant prepares them." And @UK Contractor Accountant added the enforcement context - Companies House now actively queries and rejects non-compliant or structurally flawed financial statements, and HMRC cross-references public Companies House figures against Corporation Tax returns.
Automation moves the work. It doesn't move the liability. Worth sitting with that if you're choosing between a £15 a month tool and a £1,500 a year accountant on price alone.
The second objection we honestly didn't expect the community to raise at all. @numbersrule , midway through an AI course run by his professional body, described a RAG assessment - red, amber, green - used to classify which tasks a public AI can be given, which need extra safeguards first, and which are off-limits entirely. His example of a red category activity: "Bank Statements being fed into a Public AI." And his reasoning wasn't about accuracy at all:
"It is Red because it is unacceptable from a GDPR perspective and from a Client Confidentiality perspective."
He was careful to say @GNewland is free to make his own choice, it's his own data. But the wider point stands, and @DontAsk agreed it was a good one. When you paste a bank export into a consumer chatbot, you're making a data decision as well as an accounting one. If you've got customer names in your transaction descriptions, and most small businesses do, that decision involves other people's data too.
What members said you can safely hand over
For all the pushback, the threads didn't land on "do it all yourself". They landed somewhere more practical.
@JamesUniform reframed the whole argument back in the January thread: "So for me the question is not really 'can software replace the accountant'. It is which parts are controlled enough to automate, and which bits still need judgement before the numbers are trusted." Then he listed the boring control checks that actually matter for supplier invoices - original invoice attached, supplier name matching the record, invoice number and date captured, net and VAT and gross checked, duplicates checked, approval evidence before payment, and a link back to the source document.
@UK Contractor Accountant agreed automation is transforming data entry and reconciliation, but drew the line at professional judgement - valuation of work in progress at year end, provisions for warranty claims or litigation, restructuring costs, accrued income determinations. @CLM Controller and @chenmichael landed on the same idea independently: use automation for the bookkeeping and analysis, then pay a professional for a review and the filing rather than for full monthly bookkeeping.
And @Daybooks , an accountant who runs client books in Excel and isn't remotely anti-technology, gave the warning that's stuck with us most: "Be careful: automated reconciliations are no reconciliation at all." A tick in a box isn't the same as someone actually checking.
Worth noting too - the alternative to both AI and an accountant already exists, and a few members use it. @DontAsk files his own accounts for roughly £100 a year for VT Accounts, a few pounds a month for Excel, and around £25 for filing software, on the strength of doing his own bookkeeping and reconciling every month. @tony84 pays about £1,500 for accounts and self-assessment and rates it mainly for getting questions answered in thirty seconds. Neither of those is an AI story. Both are legitimate answers.
Where that leaves us
We build automated bookkeeping. ANNA categorises transactions, calculates VAT, Corporation Tax and MTD Self Assessment, and files straight to HMRC. We think that work should be automated, and our own accounting and tax expert, Aftab Hussain, a Chartered Certified Accountant, has put a number on it: the routine work ANNA automates comes to around 90% of what accountants do for small businesses.
We'd point at the second half of what he said too, because it matters more: "It doesn't mean the end of accountants; it just signals a shift in their role, moving away from routine tasks towards more advisory services."
That's the same conclusion the forum reached by a completely different route. @Byzantium never argued for getting rid of tax advice - he argued he was paying thousands for reformatting he'd already done, and said outright he has no problem paying for good advice. @Paul Kelly ICHYB put the other side of it in one line: "a good accountant is not a cost to your business - they are an asset!" Those two positions aren't actually in conflict. They're both arguments for spending less on typing and more on judgement.
So here's the practical conclusion. Use purpose-built tools for the volume work - the categorising, the matching, the VAT arithmetic, the filing mechanics - and pick tools that are actually built and constrained for that job, not a general chatbot. Then spend what you save on the conversations a machine can't have with you: whether the salary and dividend split is right, whether that claim stacks up, what to do when something non-standard happens, and what the numbers mean for next year.
Automation's very good at the part of your accounts that's arithmetic. It's not the part you sign.
