Short answer
Automate extraction and matching, not classification. Reading documents into structured data and reconciling them against bank lines is high-volume work with a checkable answer; deciding what an unusual transaction is remains judgement with a tax consequence. In the UK, Making Tax Digital for Income Tax has applied to qualifying income over £50,000 since 6 April 2026, with £30,000 and £20,000 bands to follow. In the US no format is mandated, but the IRS can test your electronic storage system itself.
This summarizes published guidance as at 10 August 2026 and is not tax advice. Thresholds and dates move; check them against the source before acting.
UK: the deadline that already passed
A great deal of UK content still discusses Making Tax Digital for Income Tax as something coming. For the first band it arrived. HMRC’s guidance, last updated on 26 March 2026, sets it out by qualifying income.
| Qualifying income | For the tax year | You must use MTD from |
|---|---|---|
| Over £50,000 | 2024 to 2025 | 6 April 2026 — already in force |
| Over £30,000 | 2025 to 2026 | 6 April 2027 |
| Over £20,000 | 2026 to 2027 | 6 April 2028 |
Two details are easy to miss. It applies to self-employment and property income, so a landlord with a modest portfolio can be caught while a company director is not. And HMRC works from the return you already filed: it says it reviews your Self Assessment return each year and writes to confirm that you need to start by the beginning of the following tax year. The letter is a consequence of a return submitted two years earlier, which is why so many people are surprised by it.
There is a specific claim we are not making here. The “digital links” rules that prohibit manual re-keying between systems are frequently quoted in this context, but the material we could open concerned MTD for VAT rather than Income Tax. Rather than transpose one regime onto the other, read the relevant HMRC notice for the tax you are actually filing.
What the regime asks for operationally is set out in HMRC’s own guide, whose chapter titles are the shortest honest summary available: create digital records, send quarterly updates, and submit your tax return using compatible software. That is a change of rhythm as much as of tooling: a process that produced one return a year now produces something several times a year, which is precisely the kind of repetition that justifies automating the capture rather than the filing.
US: no format, but an examinable system
US small business owners often assume the IRS prescribes a system. Publication 583 says the opposite: “Except in a few cases, the law does not require any specific kind of records. You can choose any recordkeeping system suited to your business that clearly shows your income and expenses.”
What replaces format is capability. On electronic records the publication is precise: the system “must index, store, preserve, retrieve, and reproduce the electronically stored books and records in legible format”, and “all electronic storage systems must provide a complete and accurate record of your data that is accessible to the IRS”. It then adds the sentence that should shape how you build: “The IRS may test your electronic storage system, including the equipment used, indexing methodology, software and retrieval capabilities.”
The system itself is examinable. That reframes the automation question from “can it categorize accurately” to “can it produce, on demand, the document behind any given number”. On retention there is no single figure — records must be kept “as long as they may be needed for the administration of any provision of the Internal Revenue Code”, generally to the end of the period of limitations, with employment tax records kept for at least four years after the tax becomes due or is paid, whichever is later.
What to automate, and what not to
Bookkeeping is not one task. It is a volume task wearing a judgement task’s clothing, and the two should be separated before any tool is chosen.
| Task | Automate? | Why |
|---|---|---|
| Reading invoices and receipts into structured fields | Yes | High volume, verifiable against the source document, and the failure is visible. |
| Matching bank lines to documents | Yes | Mechanical, and the unmatched remainder is exactly the list a human should see. |
| Categorizing routine, repeating transactions | Mostly | Same supplier, same category, every month. Review the rules quarterly rather than each entry. |
| Categorizing anything unusual | No | A plausible wrong answer costs more than an empty field, because nobody comes back to check it. |
| Anything with a tax treatment decision | No | Capital versus revenue, deductibility, personal use. The liability stays with you regardless of what produced the entry. |
The exceptions queue is the product
In LYVIA’s experience the difference between a bookkeeping automation that survives its first year and one that quietly corrupts a ledger is whether it is allowed to say “I don’t know”.
A system that assigns a category to everything looks better on a dashboard and is worse in reality, because the errors are indistinguishable from the correct entries until someone reconciles at year end. A system with a confidence threshold pushes the ambiguous cases into a short queue that a person clears weekly, and the ledger behind it stays trustworthy.
The diagnostic is unusual but reliable: if the exceptions queue is permanently empty, the threshold is set too loose, not the model too good.
Building it on what you already pay for
Most of this capability already exists inside mainstream accounting software, and buying a second tool to do what the first one does is a common and avoidable expense. Before adding anything, check what receipt capture, bank feed matching and rule-based categorization your current package already includes and whether they are switched on.
Custom work earns its place where your process is genuinely unusual — an unusual document format from a dominant supplier, a job-costing structure the package does not model, a volume of paper that defeats the built-in capture. That is a build decision like any other, and worth testing against the questions in custom internal tools versus off the shelf.
Two adjacent processes are usually worth more than the bookkeeping itself: invoicing and payment chasing, which affects cash rather than admin time, and cash flow forecasting, which is what clean books make possible.
What this does to your accountant’s bill
Less than the marketing suggests, and in a different place than expected. The fee attached to processing transactions does fall when the data arrives clean and categorized. The fee attached to judgement, filing and dealing with the tax authority does not, and in a year with a new filing regime it can rise.
The better framing is that clean books buy you a different conversation. An accountant who spends the engagement correcting entries is not advising you; one who receives a reconciled ledger can. Whether that is worth paying for is a separate question from whether the software saved you an afternoon.
A sequence that works
- Establish which regime you are actually in. UK sole traders and landlords: check your qualifying income against the HMRC bands and the year it is measured. US: confirm what your period of limitations means for retention before you plan any disposal of records.
- Turn on what you already have before buying anything. Bank feeds, receipt capture, supplier rules.
- Fix document capture first. Everything downstream depends on the document existing, being readable, and being attached to the transaction it explains.
- Set a confidence threshold and build the exceptions queue. Decide who clears it and how often, before switching anything on.
- Test retrieval, not accuracy. Pick five transactions at random and produce the underlying document for each in under a minute. That is the capability the US system explicitly rewards, and a sound habit under any record-keeping duty.
- Keep the human on judgement. Unusual items and anything with a tax treatment stay with a person. That is not a limitation of the tooling; it is where the liability sits.
Where bookkeeping sits against everything else worth automating is in our process audit method, and the wider picture is in our guide to AI automation for small business.
Frequently asked questions
Has Making Tax Digital for Income Tax already started?
Yes, for the first band. HMRC states that if your qualifying income was over "£50,000 for the 2024 to 2025 tax year, you should've started using Making Tax Digital for Income Tax from 6 April 2026". The next bands follow: over £30,000 for 2025 to 2026 means starting from 6 April 2027, and over £20,000 for 2026 to 2027 means starting from 6 April 2028. It covers self-employment and property income.
Does the IRS require a particular bookkeeping system?
No. IRS Publication 583 is explicit: "Except in a few cases, the law does not require any specific kind of records. You can choose any recordkeeping system suited to your business that clearly shows your income and expenses." The obligation is about what the records show and whether they can be produced, not about which software produced them.
Can we throw away paper once it is scanned?
In the US, conditionally. Publication 583 says original hard copy books and records may be destroyed "provided that the electronic storage system has been tested to establish that the hard copy books and records are being reproduced in compliance with IRS requirements" and that procedures exist to keep it that way. Note also that "the IRS may test your electronic storage system, including the equipment used, indexing methodology, software and retrieval capabilities" — the system itself is examinable, not just its output.
How long do US records have to be kept?
There is no single number. The IRS position is that "you must keep your records as long as they may be needed for the administration of any provision of the Internal Revenue Code", generally until the period of limitations for that return runs out. Employment tax records have their own floor: at least four years after the date the tax becomes due or is paid, whichever is later.
Where does AI genuinely reduce bookkeeping work?
Extraction and matching, in LYVIA's own automation work. Reading a supplier invoice into structured fields, matching a bank line to the document that explains it, and flagging the handful of transactions that do not fit a pattern. Those are volume tasks with a checkable answer. The classification of an unusual item, and anything with a tax consequence, is judgement — and judgement is where the accountant earns the fee.
What is the failure mode to design against?
Confident miscategorization at scale, which is the failure LYVIA designs against first. A person who does not understand a transaction leaves it flagged; an automated system assigns it a plausible category and moves on, and the error is discovered at year end across hundreds of entries. Any setup worth having produces an exceptions queue — and if that queue is always empty, that is a warning rather than a success.
If your books are consuming a day a week and your software is already paid for, the useful first step is finding out which part of that day is extraction and which part is judgement. Book a call.
