R&D tax relief and grants for AI projects: what a UK SME can actually claim, and what it cannot.

For accounting periods beginning on or after 1 April 2024 the old SME scheme and the old RDEC were replaced by a single merged R&D expenditure credit. It is an above-the-line credit of 20 per cent of qualifying expenditure, taxable as income, so a company paying corporation tax at 25 per cent keeps about 15 per cent net; a loss-making company can surrender it for a payable credit subject to the scheme’s caps. Alongside it, enhanced R&D intensive support applies to loss-making SMEs whose qualifying R&D expenditure is at least 30 per cent of total expenditure, giving a payable credit worth 27 per cent. HMRC’s guidance and the CIRD manual are the authoritative sources, and the rules on subcontracted and overseas expenditure changed at the same time, so the accountant who makes the claim should be involved before the build, not after.

HMRC’s test is whether the project seeks an advance in science or technology by resolving scientific or technological uncertainty that a competent professional could not readily resolve. Using AI does not qualify by itself. Buying seats, writing prompts, configuring an off-the-shelf chatbot or connecting a product to your CRM is deployment, not R&D. What can qualify is the part of a build where the outcome was genuinely uncertain: whether a model could reach an agreed accuracy threshold on your kind of documents, how to make a voice agent handle interruptions reliably on a UK telephony network, how to route work between models to hit a cost and quality target that had not been achieved for that workload. The qualifying activity is the experimentation that settled the uncertainty, and the evidence is the record of what was tried and what was measured.

Your accountant or adviser applies HMRC’s tests to the facts. Records of experiments, thresholds and measurements are what make a claim defensible.

Three things, none of them costly. Write the uncertainty down before the build: the threshold or target, why it was not obviously achievable, what the competent professional would have expected. Keep the experiment record: what was tried, what was measured, what failed, which is also what a good build produces anyway. And separate the qualifying work from the deployment in the scope and the invoices, so the accountant can identify the eligible cost. We structure scopes this way when a client asks, and we are not tax advisers; the claim is made by yours.

Innovate UK’s BridgeAI programme, launched in 2023 as a £100 million initiative to accelerate AI adoption in sectors with high potential and low AI maturity, originally agriculture and food, construction, creative industries and transport and logistics, was expanded in 2026 from four to eight sectors, and runs competitions through the Innovation Funding Service alongside advice and skills support. In 2026 UKRI announced £7 million across two BridgeAI competitions for SMEs adopting AI to address a business challenge. Grant intervention rates for SMEs in Innovate UK competitions are typically up to 70 per cent for industrial research and 45 per cent for experimental development. Separately, the Made Smarter adoption programme supports manufacturers, and the government’s AI Opportunities Action Plan tracks the wider set of schemes. Competitions open and close on fixed dates and are oversubscribed, so the practical route is to check the Innovation Funding Service quarterly for your sector and to have the measured business case ready before one opens.

Grant-funded expenditure is generally treated differently for R&D relief, and the merged scheme changed the position from the old SME rules, so a project that takes a grant needs the accountant to model both before accepting either. The honest summary for most SMEs: a first AI implementation at £4,000 to £15,000 that removes a measured cost pays for itself without relief or grants, and should be judged on that. Relief is a possible reduction on the part of a build that was genuinely uncertain; a grant is a possible reduction if a competition in your sector is open and your case is ready. Neither should decide whether to build; the measured baseline should.