Before buying tax software with AI features, check five things: its figures match your last filed return on the same data, your accountant will sign off its output, its handling of your records is stated in writing, it files for your entity type, and each AI claim survives a test. Skip any tool that fails the first two.
Most owners judge these products on the demo, where the sample data is clean and every receipt reads perfectly. Your own shoebox of crumpled fuel slips, part-paid invoices and a personal card used once for a business lunch is the real test. A week of that real data, run in a trial account next to your current system, tells you more than any feature page, and the checks below fit into that week.
Three different things sold as "AI" in tax software
Vendors use the same two letters for jobs that carry very different risk, so separate them before you compare anything.
- Capture and categorisation. The software reads a receipt or bank line and proposes a category, a supplier and an amount. Errors here are small and visible, and you can fix them one by one. Xero's Smart Document Capture, which it announced in July 2026 and offers on all plans, is an example of this layer.
- Assistants that answer questions. A chat box that explains a figure, summarises the quarter or answers "can I claim this?". Intuit now groups its QuickBooks features under Intuit Intelligence; its pricing page lists Accounting AI from the Essentials plan upwards, not on Simple Start, and labels the chat feature beta with limited capacity. A chat answer sounds certain even when it is wrong, which makes this layer the riskiest to trust.
- Automatic preparation and filing. The software drafts a return or a set of accounts and sends it to the authority. Here a wrong number becomes a wrong filing, so this is the layer where human review matters most.
Write down which of the three each product on your shortlist actually offers. A tool with excellent capture and no filing is a bookkeeping tool, and judging it as tax software will confuse the comparison. For the bookkeeping side on its own, what AI can and can't do in bookkeeping covers the basics so this page can stay on the buying decision.
Run last year's file through it before you pay
The most reliable accuracy test needs no technical skill: give the new software data where you already know the right answer. You have one such dataset, the period your accountant already finished and filed.
- Pick a closed quarter or a full closed year with a filed result.
- Export the bank lines and gather the receipts and invoices for it. Import them into a trial account, not your live one.
- Let the AI categorise everything without touching it. Note how long this takes.
- Compare the software's totals with the filed figures, category by category.
- For every difference, log whether it was the AI's mistake, a rule your accountant applied that the software did not know, or a genuine disagreement to take to the adviser.
Here is an illustrative sum for an electrical contractor with nine staff. The filed quarter had 412 bank lines. After the trial import, the software categorised 389 lines without prompting and left 23 uncategorised. Of the 389, a manual check of every line found 31 in the wrong category, mostly van fuel booked as general travel and a supplier's trade counter purchases split between materials and tools. That is 31 of 389, about 8 per cent wrong, plus 23 left blank, so 54 of 412 lines, about 13 per cent, needed a human. That is a measurable result. If the figure for your current method is 5 per cent corrections, you know the new tool is not an improvement.
Set your threshold before you start. Something like "no more than 5 per cent of lines need correcting, and none of the corrections may be in a tax-sensitive category" is easy to judge. Tax-sensitive categories are the ones where a wrong label changes what you claim: vehicle costs, entertainment, home-office use, capital equipment, owner drawings and loan repayments.
A mistake the test catches: the personal card
A realistic failure looks like this. A small law firm with four fee earners tests a tool that proudly auto-categorises 96 per cent of lines. In the trial, a partner's personal card appears once in the feed for a conference ticket, and the AI books it as a business expense with confidence. It is arguably correct for that ticket, but the next line, a family restaurant bill on the same card, goes into client entertainment too. The software never asked, because its confidence score was high. Nobody would have found this in a demo. In a week of real data it shows up as the single most expensive error, since mixed personal and business spending is exactly what an enquiry from the authority looks at.
So add a column to your test log called "would have asked me?". A good tool flags ambiguous lines and a poor one guesses quietly. You want the first kind, even if it asks more questions.
What to ask your accountant before you sign anything
Your accountant is the person who signs or reviews the return, and they have to accept the software's output. Ask before buying, not after, because some practices work only with certain packages and will charge more to rework exports from others.
A short message does the job. Fill in the brackets with your own details.
Hi [first name],
I'm considering moving our bookkeeping to [software name], which has AI
features for receipt capture and categorisation. Before I commit:
1. Can you work from its reports and exports, or do you need access
to the live file?
2. Which of its outputs will you rely on (trial balance, ledger,
fixed-asset register) and which will you ignore?
3. Does it handle our [entity type] filings, or would you file from
your own software?
4. Is there anything you'd want switched off, such as automatic
categorisation of [vehicle costs / owner transactions]?
5. Would you review the first period's categorisation before we rely
on it?
I can run it alongside our current system for [one quarter] first.
Thanks,
[your name]The reply tells you more than any feature page. An accountant who says "I can't work from that" has told you the true cost: either you change adviser or you carry duplicate work. An accountant who asks for the first period's categorisation to review has given you the safest route into the software. For what advisers themselves worry about when clients adopt AI, seven AI mistakes that put an accounting firm's client trust at risk reads the question from their side.
Sign-off is also a practical point: the person responsible for the numbers needs to be able to explain how each one was reached. If the AI's reasoning is hidden behind a score, make sure the software keeps an audit trail of who approved each categorisation and when. Ask to see where that trail lives in the product.
Data handling questions that have exact answers
Tax records hold bank details, payroll, customers and sometimes health or family information. Vendors answer vague questions vaguely, so ask closed questions and ask for the answer in writing.
| Question | An answer you can accept | A reason to pause |
|---|---|---|
| Is my data used to train models? | No, or only if I opt in, stated in the terms | "We use data to improve our services" with no detail |
| Which third-party AI provider sees my records? | Named provider, with a contract that bars its own training use | No named provider |
| Can I switch the AI features off? | Yes, per feature | All or nothing, or not at all |
| Where can I export everything? | Full ledger, attachments and audit trail in open formats | Reports only, attachments locked in |
| What happens if I cancel? | Data available for a stated period, then deleted | "Contact support" |
Per-feature switches matter more than owners expect. Intuit states that QuickBooks Online's AI features cannot currently be switched off one by one, so if your accountant prefers manual categorisation on certain accounts, you would have to review rather than disable. Another product may behave differently, which is why the question belongs on your list. Check the vendor's current help pages rather than relying on this paragraph, because these settings change.
If you plan to use a general chat assistant alongside the software, the same questions apply to it. Whether it is safe to use ChatGPT with client data sets out the plan-level differences, and the short version is that business plans exclude training on your content by default while consumer plans need a setting changed. Remove names and account numbers either way.
Filing support: read the list, not the headline
"Tax-ready" and "files directly" are the most abused phrases on pricing pages. The list that counts is narrow and specific, and you should have it in front of you before you pay.
- Your entity type. Sole trader, partnership, limited company and landlord are often treated differently. FreeAgent, for instance, lists a different set of filings for each of those entity types on its pricing page, with company accounts available only to limited companies. A tool that files for a sole trader may not file for your company.
- Your authority's channel. Filing means the authority accepts the submission through its approved digital route. Ask whether the software submits directly, or whether it produces a file you upload yourself. Both can be fine, but the second means more work and more room for slips.
- Periodic returns. Sales-tax or consumption-tax returns, payroll filings and year-end accounts are separate jobs. Tick each one you need.
- Amendments. Ask how you correct a filed return. A tool that cannot amend leaves you re-entering figures by hand.
- Deadline reminders. Calendar alerts need to match your own filing dates, including any extension you have agreed.
- Add-on costs. Check whether filing, payroll or receipt capture sit behind a higher plan. FreeAgent, for example, limits Smart Capture to 10 uses a month unless you pay for the add-on. Capture quotas are common, and a quota that runs out in week two is a hidden cost.
Prices change often, so use the vendor page on the day you buy. As of October 2026, QuickBooks Online lists Simple Start at $38, Essentials at $85, Plus at $140 and Advanced at $340 a month, while Xero lists Early at $27, Growing at $59 and Established at $97, with its automatic reconciliation (JAX) on Growing and above. Those are list prices for the accounting plans, not your filing costs. Ask for the filing cost separately. Xero against QuickBooks on AI compares the two in more depth.
Turning vendor claims into things you can test
Marketing copy is written so that it cannot be tested. Your job is to rewrite each claim until it can. Paste the vendor's feature page into a chat assistant on a plan that does not train on your data, and use a prompt like this one.
Below is the feature page for an accounting product. List every
claim that mentions AI, automation, accuracy, time saved or
compliance. For each claim, write one test a small-business owner
could run in a free trial using their own data, and say what result
would count as a pass. Do not add claims that are not on the page.
[paste page text]An illustrative output for an insurance broker's shortlist might include: claim "saves up to 80 per cent of data entry", test "time 50 receipts entered manually and 50 captured, same person, same day", pass "capture plus corrections takes under half the manual time". Another: claim "audit-ready records", test "export the audit trail for one transaction and ask your accountant whether it answers what an enquiry would ask", pass "accountant confirms". The AI is not judging the claims, only turning them into homework. You still do the homework.
Watch for numbers with no baseline. "Saves 10 hours a month" depends on the 10 hours you started from. Ask the vendor how the figure was measured and for what kind of business. If they cannot say, discount it to zero for your decision.
Scoring a shortlist: an electrical contractor's version
Use a simple weighted score so the choice rests on evidence rather than the last demo. Say the electrical contractor with nine staff is weighing two tools, called Tool A and Tool B here. Weights add up to 100 and reflect what hurts most if it goes wrong.
| Check | Weight | Tool A (0-5) | Tool B (0-5) |
|---|---|---|---|
| Accuracy on last quarter's file | 30 | 3 | 4 |
| Accountant will work from it | 25 | 5 | 2 |
| Data handling and export | 15 | 4 | 4 |
| Files what we need directly | 15 | 4 | 5 |
| Flags ambiguous lines instead of guessing | 10 | 3 | 3 |
| Total price for filing and capture, yearly | 5 | 3 | 5 |
To score, multiply each weight by the score and divide by 5. Tool A: 30 x 3 + 25 x 5 + 15 x 4 + 15 x 4 + 10 x 3 + 5 x 3 = 90 + 125 + 60 + 60 + 30 + 15 = 380, divided by 5 gives 76. Tool B: 120 + 50 + 60 + 75 + 30 + 25 = 360, divided by 5 gives 72. Tool B looks better on price and filing, but the accountant's reluctance drags it below Tool A. If the accountant would change their mind for a fee, rescore with that fee added to Tool B's cost. Putting numbers on the table keeps the argument to the point.
Set two knock-out rules before scoring: any tool whose output your accountant will not work from is out, and any tool that cannot file for your entity type is out. Scores should only rank the survivors.
Reading the trial results without fooling yourself
After the parallel period, resist the urge to judge on speed alone. Make three comparisons in a short table.
- Hours. Time spent on entry, categorisation and reconciliation under the old method against the new, measured for the same period. Include time spent fixing AI errors.
- Corrections. Count of lines changed after the AI had processed them, and how many of those were in tax-sensitive categories.
- Questions from your accountant. If the adviser sends fewer queries about the new file than the old one, that is the strongest sign of better records. If they send more, the tool is creating work somewhere else.
A realistic result for a four-fee-earner law firm might be: bookkeeping hours down from 14 to 9 a month, corrections at 6 per cent of lines, accountant queries unchanged. That is a modest win, worth keeping if the subscription costs less than the five hours saved. The same firm should not expect the AI to produce tax advice. Matters such as how to treat a client disbursement or a partner's capital contribution remain adviser decisions, and software or accountant for tax compliance explains where that line sits.
Who should switch off the AI or skip the upgrade
Some businesses gain very little from AI features, and paying more for them is a poor trade.
- Fewer than about 30 transactions a month. Manual categorisation takes minutes. The AI saves almost nothing, and a mistake costs the same as it always did.
- Highly irregular books. A broker with commission clawbacks, client money or multi-currency items will find the AI's pattern-matching trips over exactly the odd lines that matter. Use rules your accountant sets rather than AI guesses.
- A practice that prefers its own software. If your adviser files from their own system, the pay-off from your software's filing feature is zero, so keep to capture and reporting.
- Anyone with an enquiry open. Do not change systems mid-enquiry. Migrate after it closes.
For a business in the first group, the better spend is usually a clean chart of accounts and a monthly habit of reconciling. The AI features can wait until volume justifies them.
Switching on one feature at a time
If you buy, introduce the AI in stages so you can tell which part helps. Start with receipt capture for a month, since it is low risk and its errors are visible. Add bank-line categorisation next, with a rule that you review everything in the tax-sensitive categories weekly. Leave assistants that answer tax questions until last, and treat their answers as prompts for your adviser, not as decisions. Keep a one-page log: date switched on, what changed, corrections in the first two weeks. If a feature adds corrections without saving time, switch it off or ask for a refund while you still can.
Finally, put a date in the diary three months after the switch to repeat the accuracy test on a fresh quarter. Models and features update without notice, and a tool that passed in week one can drift. If you would like help weighing the trade-offs against your own accounts, that is the kind of decision my AI implementation consultation covers; see the contact page.
Questions owners ask before switching tax software
Can AI tax software replace my accountant?
Not safely. The software can capture receipts, sort transactions and draft figures, but a qualified person still has to decide treatments, sign off judgement calls and answer your tax authority if it asks questions. Use the software to cut the data-entry hours you pay for, and keep the adviser for the decisions.
Is a free AI tax tool good enough for a limited company?
Only if it files in the format your tax authority accepts for your entity type and your accountant is happy to work from its output. Check the filing list line by line, because many free or cheap tools cover sole traders and stop short of company accounts.
Should I paste my accounts into ChatGPT to check them?
Only on a plan whose terms exclude training on your data, and with names and account numbers removed. Treat any answer as a lead to check against the ledger, not a finding. Rows of real client or supplier data belong inside your accounting software, not a chat window.
How long should a trial of tax software run?
Long enough to process one full period of real transactions, usually a month or a quarter, alongside your current system. A one-hour demo on sample data tells you how the screens look, not whether the figures agree with yours.
Further reads
- How Much Does AI Bookkeeping Software Cost per Month? — Puts monthly prices on the bookkeeping side of the same decision.
- AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — Shows how receipt capture works before you rely on it for tax.
- AI Bookkeeping Software vs a Human Bookkeeper: Accuracy and Cost — Compares accuracy and cost when a person checks the books.
- How Accountants Use AI to Explain Tax to Clients in Plain English — Useful once your accountant sends figures you need explained.
- AI Engagement Letters for Accountants: Faster Drafts, Clear Scope — Shows what an adviser's scope document should say about AI.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Intuit and Xero product and pricing pages, FreeAgent's pricing page, vendor help centres, and OpenAI and Anthropic data-use pages checked in October 2026.