Getting Ready for Making Tax Digital With AI Tools

If your farm turned over more than £50,000 in the 2024/25 tax year and you trade as a sole trader, Making Tax Digital for Income Tax already applies to you. It started on 6 April 2026. The first quarterly update was due on 7 August, and the next one is due on 7 November 2026. Many farmers are caught out because the threshold is based on turnover, not profit. A suckler herd that sells £140,000 of stores and makes £18,000 profit is in scope. A sheep enterprise grossing £55,000 is in too.

This post doesn’t cover the whole scheme. It covers three jobs where AI tools save real hours: sorting bank transactions, chasing missing receipts, and explaining the deadlines in plain English. It also covers the points where you should stop using AI and pay your accountant to sign things off.

What MTD for Income Tax actually asks of you

Under the old system you sent one return a year. Under MTD you keep digital records and send HMRC a summary of income and expenses by category every quarter, using compatible software. After the year ends, you send a final declaration that replaces the old Self Assessment return.

For the 2026/27 tax year, the dates on standard quarters are:

The threshold drops to £30,000 from April 2027 and to £20,000 from April 2028, so most family farms not caught yet will be within two seasons. Partnerships aren’t in scope yet, and limited companies aren’t covered at all. If you farm as a partnership with your spouse, check which side of that line you’re on before you spend money.

The quarterly updates are not tax bills. You pay tax on the same dates as before (31 January and 31 July). The updates are running totals, and you can correct mistakes in a later quarter. That means a slightly messy Q2 isn’t a disaster, as long as the final declaration is right.

Job one: sorting bank transactions

Farm bank accounts are messy. Mart cheques arrive with commission and levies already taken off. Fuel, feed and fertiliser get paid by card, by direct debit and by BACS. Subsidy money comes in as one large sum. Some months, half the transactions are transfers between the business account and the house account.

Accounting software like Xero, QuickBooks and Sage already suggests categories using bank rules. AI is useful for the transactions those rules can’t handle, and for building the rules in the first place. Export a quarter’s bank statement as a CSV and remove the account number and sort code. Then paste it into ChatGPT (Plus is about £20 a month) or Claude with a prompt like this:

You are helping a UK sole-trader farmer (beef and sheep, 400 ewes, 60 sucklers)
categorise bank transactions for Making Tax Digital quarterly updates.
Use these categories: Sales - livestock, Sales - other, Subsidy/grant income,
Feed, Vet and medicines, Fertiliser and seed, Fuel, Contractors, Repairs,
Rent, Insurance, Bank charges, Drawings, Capital item - ask accountant,
Transfer - not income.
Return a table with date, description, amount, category, and a confidence
of High/Medium/Low. Flag anything Low for me to check.

This is what came back for a short sample:

Date        Description                    Amount     Category                      Conf
12/07/2026  BORDERWAY MART SETTLEMENT      +8,412.60  Sales - livestock             High
14/07/2026  MOLE VALLEY FARMERS DD         -1,236.00  Feed                          Medium
18/07/2026  RPA SFI PAYMENT                +6,950.00  Subsidy/grant income          High
22/07/2026  NWF AGRICULTURE                  -980.40  Feed                          High
02/08/2026  J SMITH CONTRACTING              -2,160.00 Contractors                  High
09/08/2026  TFR TO 40XXXX                  -1,500.00  Drawings                      Medium
21/08/2026  RYETEC MACHINERY               -11,400.00 Capital item - ask accountant Low
03/09/2026  AMAZON MKTPLACE                   -64.99  ??? (could be repairs/household) Low

The Mole Valley line shows why the confidence column matters. That direct debit might be feed, or it might be fencing, wellies and a bag of minerals on one statement. The model can’t see the invoice, so it shouldn’t pretend to. Anything marked Low or Medium goes on a short list for you to check by hand.

Once a quarter’s sorting looks right, ask the same chat to “write these as Xero bank rules”. It will give you text such as “if description contains BORDERWAY MART, code to Livestock Sales”. You then copy those rules into the software yourself. The AI drafts the rules, and your accounting software stays the record HMRC sees. That keeps you inside the digital-links requirement, because the figures that reach HMRC come straight from compatible software and are never retyped.

On a 900-line year for a mixed farm, expect the first pass to take about 40 minutes, against the best part of a day doing it by hand. In later quarters most lines will be caught by rules before AI sees them.

Job two: chasing missing receipts

Receipts are where MTD hurts. A quarterly update with £4,000 of uncategorised spend isn’t technically late, but every unsupported cost is one your accountant will query in January.

The low-cost setup most farmers settle on is a receipt-capture app linked to the books. Hubdoc comes free with Xero. Dext and AutoEntry cost roughly £20 to £30 a month at small-farm volumes. You photograph the receipt in the tractor cab, and the app reads the supplier, date, VAT and total, then matches it to the bank line.

AI helps with what’s left. Most software can export a list of unreconciled transactions or ones with no attachment. Give that list to ChatGPT and ask:

Here are 14 bank payments from Q2 with no receipt attached. Group them by
supplier. For each supplier, draft a short polite email asking for copy
invoices for these dates and amounts, to be sent from [farm name].
For card payments under £20 with no obvious supplier, list them separately
so I can decide whether to write them off as drawings.

You get back maybe five supplier emails ready to paste: one to the agronomist for two missing invoices, one to the vet practice for a monthly statement, one to the merchant for a copy statement, and so on. It takes ten minutes instead of an evening, and the small unexplained card spends are listed in one place for you to decide.

Don’t let the AI decide what to do with an unexplained £300 cash withdrawal at the mart. That goes to you or your accountant.

Job three: getting the deadlines explained

General AI chat tools are surprisingly good at turning HMRC guidance into something you can follow, as long as you give them your actual dates and check the answer against GOV.UK. A useful prompt:

I'm a sole-trader dairy farmer. My 2024/25 turnover was £310,000.
I use standard quarters. Today is 9 October 2026.
Explain what I must submit to HMRC in the next six months, with dates,
and what I do NOT have to do quarterly. Plain English, no jargon.

A good answer will tell you that the 6 July – 5 October quarter is due by 7 November and the next by 7 February. It will say that payments on account still fall on 31 January, and that you can choose calendar quarters if your accountant prefers them. It may also mention that HMRC has said it won’t give penalty points for late quarterly updates during 2026/27. Check that last point on GOV.UK rather than relying on it. AI tools sometimes state an old rule with complete confidence, and anything dated is where they go wrong. Treat the chat as a patient explainer, and treat GOV.UK as the authority.

The more general farm-office prompts (letters, tenancy summaries, scheme paperwork) are covered in our guide to ChatGPT and AI for farm admin.

Where your accountant still has to sign off

AI handles sorting and drafting well. It doesn’t hold professional indemnity insurance, and it doesn’t know your history. These are the farming items to keep away from the chatbot:

• Capital items. An £85,000 tractor bought in September might get the Annual Investment Allowance, or might be better spread, depending on profits and averaging. The Ryetec line above is a decision for your accountant, not a category.
• Farmers’ averaging. Averaging profits over two or five years can save thousands in a bad-then-good run. It’s a claim made on the final declaration, and it needs someone who can see the whole picture.
• Herd basis elections. These are irrevocable and only worth it for some production herds. Never let AI “assume” one.
• Livestock and crop valuations. HMRC’s deemed cost rules (60% of market value for cattle, 75% for sheep and pigs) affect your year-end stock figure. That’s a judgement made with your records, not a calculation for a chatbot.
• Private-use splits. The farmhouse proportion, the pickup used for school runs and the electricity meter shared with the dairy all need agreed percentages that will stand up to an enquiry.
• Subsidy treatment. SFI, Countryside Stewardship and delinked payments are all taxable income. Grants towards capital items are treated differently. Coding a grant wrongly can shift profit between years.
• The final declaration itself. You’re declaring it’s correct and complete. Have a qualified person review it.

A sensible division of work is that you (with AI help) keep the quarterly updates tidy and your accountant handles year-end adjustments and the final declaration. Many practices quote a few hundred pounds a year extra for MTD support. That is likely to cost less than having them unpick twelve months of uncategorised transactions in January.

A rough monthly bill

Farmplan and Figured are farm-specific alternatives worth pricing if you want enterprise costings (per-hectare or per-litre figures) as well as tax returns. Ask for a quote, because pricing depends on modules.

A note on data

Don’t paste full bank statements, with account numbers, National Insurance numbers or your UTR, into a free chatbot. Remove those columns first. In ChatGPT, also switch off “Improve the model for everyone” in settings. A CSV of supplier names and amounts is low risk. Your full financial identity is not.

If 7 November is your next deadline, start this week. Export July to September from the bank, run the categorisation prompt, send the receipt-chasing emails, and give your accountant a clean list of the Low-confidence lines before the end of October.