Making Tax Digital and AI for UK Finance Teams

Making Tax Digital (MTD) is the UK's single most consequential constraint on AI use inside a finance team, and it is consistently underweighted in 2026 discussions of AI in UK finance. The digital-link rule, the functional-compatible-software rule, and HMRC's published record-keeping expectations together define the shape of what a UK finance team can and cannot do with AI-assisted workings on MTD-scope work. Getting the architecture wrong is not a governance abstraction; it is a compliance exposure that sits on the Finance Director.

This article sets out, in plain terms, what that means for a UK finance function: what MTD actually requires of the record-keeping chain, why AI workings do not meet that requirement on their own, where AI can legitimately sit inside an MTD-compliant workflow, how ICAEW and HMRC have signalled expectations, and a clear next step.

The framing is descriptive. Nothing in this article claims ICAEW, CIMA, ACCA, or HMRC alignment, endorsement, or accreditation on behalf of Learn AI. Any policy or procedure that would bind the firm should be reviewed by the Finance Director, the external accountant, or an ICAEW-accredited advisor.

What does Making Tax Digital actually require of the record-keeping chain?

MTD is a single compliance posture that sits on top of two separate obligations: MTD for VAT (now the default for all VAT-registered businesses) and MTD for Income Tax Self Assessment (rolling in from April 2026 for self-employed and landlord income over £50,000, and April 2027 for the £30,000 to £50,000 band). The two regimes have different scope and different go-live dates, but they share the architecture that matters for AI use.

Three requirements define the shape.

Digital records in functional compatible software. HMRC requires that the primary records for an MTD-scope activity be held in digital form, inside software that can communicate with HMRC's systems via an Application Programming Interface (API). A paper ledger, an unstructured Word document, or a PDF export of a calculation is not a digital record in the MTD sense. A spreadsheet can be, but only if it is connected to bridging software or a full-stack MTD product that carries the digital link through to submission.

Digital links between records and submission. The digital-link rule (originally phased in for MTD for VAT after a soft-landing period through 2021 and now fully enforced) requires that any transfer of data inside the MTD record-keeping chain be done by digital means, not by re-keying. A copy-paste between two spreadsheets is not a digital link; a formula, a cell reference, an import, a linked workbook, or an API transfer is. HMRC's published guidance is specific about what does and does not qualify.

Records retained for the statutory period. Businesses must retain the underlying digital records for the MTD-scope tax for the statutory retention period (six years for VAT, varying for Income Tax depending on filing posture). The records are retained in the format in which they were kept during the return period. Reconstructed records produced after the fact do not satisfy the retention obligation.

Read together, the three requirements define the MTD-compliant record-keeping chain: digital records, digitally linked through every stage of working, held in functional-compatible software, retained for the statutory period. A UK finance team deploying AI anywhere in that chain has to place the AI such that the chain remains intact.

Why do AI workings not meet the MTD requirement on their own?

AI workings do not meet the MTD requirement on their own for four reasons. All four are mechanical, not philosophical. Each is fixable by putting the AI in the right place in the workflow.

AI workings sit outside functional-compatible software. A prompt history inside ChatGPT, a conversation inside Claude, a Copilot session inside Microsoft 365, or a draft produced by a general-purpose model are not records held in MTD functional-compatible software. The conversations can be exported, but the export is not itself the MTD record; it is a document about the record. Treating the prompt history as the record fails the functional-compatible-software test.

Re-keying breaks the digital link. A Finance Assistant who asks an AI tool to draft a VAT calculation and then types the result back into the bookkeeping system has broken the digital link. The transfer of the figure from the AI output into the MTD record is a re-keying operation. HMRC's guidance on digital links is unambiguous: a figure that has been manually re-entered into the chain is a new figure, not a linked one. The audit trail no longer holds.

Non-persisted workspaces do not satisfy retention. An AI workspace that is cleared, archived, or rotated out after a period does not satisfy the statutory retention obligation. An AI prompt session that is not exported to a persistent store is, for MTD purposes, not retained at all. The finance team that relies on AI working in a temporary session has, in MTD terms, failed to keep records.

API-submitted returns require traceability. MTD returns are submitted via API from functional-compatible software to HMRC. The HMRC-facing submission has to be traceable back to the underlying digital records through unbroken digital links. An AI working that sits off-platform creates a gap in the traceability chain that, if queried on a compliance check, is difficult to close after the fact.

None of the four points bans AI use in a UK finance team. All four require the finance team to architect the workflow so the AI sits alongside the MTD chain, not inside it, unless the AI feature is embedded inside a piece of functional-compatible software.

Where can AI legitimately sit inside an MTD-compliant workflow?

AI can sit in four places inside an MTD-compliant workflow without breaking the chain. The pattern is consistent: AI is an assistant on the workings, with the MTD-compliant software holding the record of what is actually submitted.

AI embedded inside functional-compatible software. Xero, QuickBooks, Sage, FreeAgent, and the MTD-compatible product set now routinely include AI features (bank-feed categorisation, invoice capture, suggested reconciliations, VAT-flag review). Because the AI is inside the functional-compatible software, the output lands directly into the MTD record chain. The digital link is preserved because the data never leaves the software. A finance team that uses the vendor's AI features is inside the MTD posture by construction.

AI used on outputs that sit outside the MTD chain. Narrative commentary, management reports, board-pack text, and variance explanations are not MTD records. A finance team that uses a general-purpose AI tool to draft those outputs is not compromising the MTD chain, because the MTD record in the bookkeeping system is independent of the narrative commentary. This is the category where unconstrained AI use has the most value in a UK finance team.

AI used on a review pass, not a drafting pass. A finance team can use AI to review an MTD submission for anomalies, missing lines, or unexpected variances, and then adjust the MTD record inside the functional-compatible software based on the review. The AI output is an input into human judgement; the MTD record is adjusted inside the compliant software. The digital link is preserved because the adjustment is made in the software, not re-keyed from the AI output.

AI used on pre-submission analysis. A Finance Director who wants to stress-test a VAT return against prior-period patterns, a CT600 provision against the prior-year comparator, or a P&L against forecast can use AI to produce the analysis. The analysis is a management artefact, not an MTD record. The return is still prepared and submitted from the functional-compatible software; the AI analysis informs the human review, not the figures on the return itself.

The question the Finance Director should ask on any proposed AI use case is simple: does the AI output become a figure on an HMRC-facing return? If yes, the AI has to be inside functional-compatible software; if no, the AI is in the free zone.

What have ICAEW and HMRC signalled about AI inside MTD?

ICAEW and HMRC have signalled three positions across 2024 and 2025. None of the positions has shifted in the months since.

ICAEW: professional competence applies to AI-assisted workings in full. ICAEW's published material through 2024 and 2025 has consistently framed AI-assisted tax and accounting work as a member-competence question, not a tooling question. A member who relies on an AI-assisted calculation for an MTD-scope return is expected to understand the tool's limitations, evidence the review step, and stand behind the output as their own. The framing is consistent with the broader ICAEW signal on AI in professional work covered in AI for UK Finance Teams. MTD does not relax that position; it tightens the evidence requirement.

HMRC: the record-keeping obligation sits on the business. HMRC's published guidance on MTD record-keeping, supplemented by compliance-check guidance through 2024 and 2025, is unambiguous that the record-keeping obligation sits on the VAT-registered business or the Income Tax taxpayer. A compliance officer who opens a check on an MTD-scope return will ask for the digital records and the digital links, not for the AI prompt history. A finance team that cannot produce the linked records because the figure was reconstructed from an AI output is in a compliance-check exposure.

HMRC: consumer AI tools are not functional-compatible software. HMRC has not (as of April 2026) approved any general-purpose AI chat product as functional-compatible software for MTD purposes. The approved list is the set of MTD-compatible bookkeeping and tax products. A finance team that assumed "the AI can produce the VAT calculation" has misread the approved-software register. The AI can sit alongside the workflow; the submission has to come from approved software.

Read together, the three positions do not prevent a UK finance team from using AI on MTD-scope work. They constrain the architecture. The constraint is simpler than it first looks: keep the MTD record inside functional-compatible software, and use AI on the workings and the review.

What a UK finance team does next

The posture is available. It is a three-step sequence a Finance Director can complete inside the current return cycle.

First, map the MTD chain. Document, end to end, every piece of software and every manual step that currently sits between the firm's primary records and the HMRC-facing submission for each MTD-scope tax. Mark every point where an AI tool is currently being used.

Second, categorise each AI touch-point. For each AI use on the map, confirm whether the AI is (a) embedded inside functional-compatible software, (b) on an output that sits outside the MTD chain, (c) on a review pass with the adjustment made inside the compliant software, or (d) on pre-submission analysis that does not touch the return. Any AI use that does not fit one of the four categories is a compliance exposure and needs to be moved or stopped.

Third, document the position. A one-page note signed off by the Finance Director sets out where AI is used, in which category, and how the MTD chain is preserved. The note becomes the answer to an HMRC compliance-check query if one ever lands. The firm's internal audit function benefits from the same note. The AI Readiness Assessment produces the initial map for a firm running this exercise for the first time, and the AI for Finance Teams workshop walks the Finance Director through the MTD categorisation against the firm's actual stack.

MTD did not anticipate generative AI. What MTD did anticipate was that the record-keeping chain would move to software and stay in software. A UK finance team that keeps that principle in mind while it builds its AI posture will find the two frameworks compatible; a team that does not will find itself reconstructing records after the fact, which is the one thing MTD does not allow.