Seven AI mistakes accounting firms made in 2026, and the fixes
Seven AI mistakes described by named speakers at the AI in Practice Summit in November 2026, from an agent rolled out before it worked to redaction that only hid the data. Each comes with what the speaker changed afterwards, in their own words.
Part of our AI in accounting coverage. See the full AI for accountants guide →
Some of the most useful sessions at the AI in Practice Summit in November 2026 were the ones where practitioners said what went wrong. A founder rolled an agent out to her team before it worked. A software company trusted a confident answer over a practitioner who had read the Act. A firm owner wished he had made one decision six months sooner.
This piece collects seven of those mistakes. Each comes from a named speaker who described it, with what they changed. Where a speaker was describing other firms’ mistakes rather than their own, we say so.
| Mistake | Who described it | What changed |
|---|---|---|
| Rolling an agent out before testing it | Natalie Lennon | Test privately, spot-check, start small |
| Trusting a confident answer over the Act | Andrew Noble | Rules written as code, reviewed by practitioners |
| Putting generative AI at the centre | Dave Sellick | Use AI to build tools that run without it |
| Putting AI over an unmapped process | Amy Holdsworth | Map and connect first |
| Letting everyone do their own thing | Andy Teece | Shared skills on one firm plan |
| Redaction that only hides the data | James Rose, Travis Bacon | A non-AI tool that removes it first |
| Booking AI as a flat subscription | David Boyar | Track it as a cost per client and job |
1. Rolling an agent out to the team before it worked
Natalie Lennon, founder of Two Sides Accounting, runs an AI agent called Viktor inside her firm’s Slack. One of the first things she built was a Karbon checker, and she put it straight into a team channel.
“I set up this Karbon checker and it was wrong, and the team was like, Nat, this is not working,” she told the summit. Her fix is now her rule: test every new instruction in a direct message with the agent before the team sees it, spot-check the output (“Don’t just take the figures as gospel”), and start with a couple of tools and one channel. See also designing the role for your first AI assistant. Natalie Lennon’s session notes.
2. Trusting a confident answer over a practitioner who had read the Act
Andrew Noble, founder of LodgeiT Labs, opened his session with his own company’s error. A LodgeiT worksheet was calculating tax losses in a way the team believed was right, and they used AI to help decide whether it was. A user, Jenny Haines, a chartered accountant who runs Natural Numbers Accounting, kept telling them the figure was wrong. The software was carrying forward too much tax loss. “Jenny was right, and we were wrong,” Noble said.
Haines described the other side: she had gone back to the legislation, and the answer she was arguing against “quite confidently said it’s right.” What LodgeiT changed was to write the relevant rules (starting with Division 36) as code that gives the same answer every time, asks for missing facts such as residency, and refuses cases it does not cover. Practitioners are now paid to review its rules, and one reviewer found that a CGT rule relied on a subsection that had been repealed. More on the limits of AI in accounting. When AI gets tax wrong, session notes.
3. Putting generative AI at the centre of mission-critical work
Dave Sellick, founder of UK firm Sidgrove, runs about 50 clients on his own. The mistake he warned against is the one he sees the profession being sold: agents completing accounting work end to end, with no human leading them.
They are probabilistic technology by design, which means they’re optimised to be plausible, not correct.
His own fix came from a different problem. His connected spreadsheets worked, but “When I made an update in one client, it wouldn’t update all of them.” So he used AI coding tools to turn them into one app that syncs with Xero through its API, using fixed rules wherever he can and generative AI only for review and exceptions. He also warned against building dozens of separate apps, each one another way for client data to be insecure. For anyone not ready to build, his advice was to start in chat and stay in chat. More in our earlier piece on Dave Sellick’s approach and what AI native actually means for a small firm. Dave Sellick’s session notes.
4. Putting AI over a process nobody had written down
Amy Holdsworth, director of Clarity Street, described a conversation she has with firms over and over. They want AI over their practice management, document and engagement software to streamline a process. She asks if the process is written down. The answer is usually no.
“If you’re expecting it to remove a manual process, it just hides it, and it speeds up the mistakes,” she said. Her order is: map the process click by click, connect the apps you already pay for (her example was Ignition, Xero Practice Manager and Xero used separately rather than linked), clean the client data, and only then add AI where the process is clean. Why you’re losing the AI battle, session notes.
5. Letting everyone in the firm do their own thing
Andy Teece, director and co-founder of Atomic Business Advisers in Sydney, has used AI in the firm since early 2023. Around 18 months ago he moved the team off everyone doing their own thing: short AI training sessions twice a month, and shared skills and instructions so emails, reports and point-form client summaries come out the firm’s way.
His regret was timing. At first he asked staff to load new skills into their own accounts; the firm now has a business plan that rolls them out centrally. “Looking back on it, I probably would have done that 6 months earlier,” he said. On the same panel, Gus McLennan of AI Dojo described a firm where adoption of a new AI platform plateaued until half-hour in-person training sessions began, after which he said usage in the firm doubled within the month. Where to start with AI, panel notes.
6. Redaction that only hides the data
Client data safety was the question the summit audience asked most. James Rose, co-founder of Content Snare, said “there’s a lot of dodgy redaction out there where it just puts a black box over the top of something.” The text underneath can still be there.
His fix is a small tool, built with AI but containing no AI, that strips tax file numbers and other identifiers out of the document before any AI sees it, followed by a separate check that the data is gone. Travis Bacon, founder of TaxDigital, does the same in the tools his firm built, starting with BAS, sending data to the Claude API only after it has been de-identified: “build the tool outside of the AI to get to that result.” Your firm’s AI policy should say which data never goes in at all; our AI for accounting firms hub collects the guides. Teach your team, panel notes.
7. Booking AI as a flat subscription and forgetting it
David Boyar of The Access Group described a mistake he believes most firms are making now rather than one they have already paid for. “Right now, I’ll bet you are tracking whatever AI you’re using to your subscriptions account,” he said. His view is that flat-rate plans are subsidised and heavier work will move to usage pricing. “Good luck to the firm who lets staff run rampant with no token budget.”
His fix: put AI on the books as a cost of delivering work, try to track it per client and per job the way some firms track time, and audit what the usage is paying for. Sellick made a related point from the other side, estimating that one month of his own usage would have cost tens of thousands at published API rates. Both are forecasts, not settled facts, and no vendor has published how it will price heavy accounting use. The great AI con, session notes.
The pattern
None of these mistakes was about picking the wrong model. Each was about trust arriving before checking. The agent went to the team before it was tested. The worksheet answer was trusted before anyone read the section. The AI went on top of a process nobody had written down, and the redaction was trusted without looking underneath. The fixes look alike too: a human leads, a deterministic step sits between client data and the AI, and the firm decides once, in writing, how the tool is used.
What to do this week
- Pick one boring, repeatable job you can check in five minutes, and test it privately before anyone else uses it.
- Write down, click by click, the one process you most want AI to speed up. If you cannot, that is the first job.
- Check how your firm removes client identifiers. Open a redacted file and confirm the data is gone, not covered.
- Set every connected AI tool to draft only, with any write access to Xero or your practice management software needing approval.
- Give AI spend its own line in your chart of accounts, and note which clients and jobs use it most.
- Put your shared prompts and instructions in one place the whole team uses.
Frequently asked questions
Were these mistakes made by the speakers themselves?
Three were first-hand: Natalie Lennon’s Karbon checker, LodgeiT’s tax loss worksheet and Andy Teece’s timing on a firm-wide plan. Dave Sellick’s spreadsheet problem was his own too. Amy Holdsworth, James Rose and David Boyar were describing mistakes they see in other firms or across the market.
Is it safe to let an AI agent write to Xero?
The speakers who do it limit it. Lennon keeps every Xero action behind her approval, and Sellick uses the Xero API with fixed rules rather than letting generative AI act on the ledger. Start read-only, then approve each write.
Do I need to learn to code to avoid these mistakes?
No. Sellick’s advice for anyone not ready to build was to start in chat and stay in chat. Most of the fixes here are process changes: test first, write it down, check the redaction, track the cost.
Can AI do our GST and BAS reconciliation yet?
The speakers were split. Lennon said she had not found an AI that does it without a lot of errors, while Andy Teece described one that found the two periods where a GST error occurred. Both kept a qualified person reviewing the result.
Where can I watch the sessions?
Each section above links to the speaker’s session notes page, which carries the summary and the recording where one is available.
AI in AccountingAI in Practice SummitData securityXeroKarbon
Get the editable AI policy template Download the AI policy template
The Word document (.docx), ready to fill in. Your profile unlocks every template on The Firm.
Or build it with AI
Open Claude or ChatGPT with the prompt filled in. It writes the finished document for your firm straight away, then asks a few questions to tailor it. No client data needed.
Works without client data. Don't paste client names, tax file numbers or ledgers into an AI tool your firm hasn't approved for them. AI data safety playbook.