AI in Architecture: What Actually Qualifies for R&D Tax Relief?

I had a conversation this morning with one of the most forward-thinking architects I work with. Mid-way through our annual R&D review, he paused and asked me something I’ve been expecting more practices to raise: “We’re spending a huge amount of time using AI to generate visual designs. Can we claim that?”

It’s the right question. The answer is “it depends” but not in the waffly, hedge-everything way that answer usually implies. The legislation is quite specific about where the line sits.

How Architects Are Using AI Right Now

Architectural practices are adopting AI tools quickly. From what I’m seeing across our client base, the main uses cluster into a few areas. Tools like Midjourney, DALL-E, Adobe Firefly and specialist platforms like Maket.ai are being used to generate massing studies, material options and elevation permutations at a pace that was impossible two years ago. My client this morning mentioned Banana Pro 2 in the same breath as his CGI workflow: what used to take days of manual rendering now takes minutes per iteration.

Beyond visualisation, AI-assisted tools are being used alongside Grasshopper and Rhino to run structural and spatial configurations, optimising simultaneously for daylight, thermal performance and planning massing constraints. With the post-Grenfell fire safety regime still unsettled and the Greater London Authority’s carbon requirements now imposing complex PSI and thermal modelling obligations, practices are using AI to explore compliance solutions on constrained sites where standard details simply will not work. And increasingly, AI is being used to pull together the 50-page sustainability spreadsheets that planning authorities now demand before they will accept an application.

Where the Eligibility Line Sits

The starting point is technological uncertainty under CTA 2009 s.1042 and the CIRD guidance at CIRD81900. Only activities that directly contribute to resolving a technological uncertainty qualify for relief.

There is a specific exclusion that applies directly to design work: purely aesthetic activity does not qualify unless it directly contributes to resolving the technological uncertainty. This is an explicit boundary in the legislation and HMRC guidance, not a grey area.

“You’re only allowed to claim for aesthetic work if it’s directly contributing towards resolving technological uncertainty.”

So the practical test is simple: why are you generating these images or running these models?

If the answer is “to show the client different looks for the back elevation” or “to produce a CGI for the planning application”, that is aesthetic work and it does not qualify. Full stop, regardless of how sophisticated the tool is or how novel it was to your practice at the time.

If the answer is “to work out whether we can achieve a compliant massing solution on a constrained site given the planning authority’s specific objection to bulk and scale”, you are using the visual output as a tool to resolve a technical problem that has a direct bearing on whether the building can be built at all. That is a different situation.

The Massing Example

My client this morning described a scenario that makes this concrete. The planning officer says: go away and come up with different permutations of how the roof sits in relation to the bulk and scale of the building. Without a solution, planning permission is refused and the project dies.

Running AI-generated massing iterations in that scenario is not aesthetic exploration. It is the mechanism by which a genuine constraint is being resolved. At the outset of that exercise, it is not readily deducible whether a compliant solution exists, let alone what it looks like. That is the technological uncertainty test in CTA 2009 s.1042, and this activity meets it.

Translation: Using AI to iterate through massing options because you need to find something the planning authority will accept = eligible (subject to the constraint being genuinely non-obvious). Using AI to produce five different looks for the back elevation because the client wants to see options = not eligible. The question is always: does this directly contribute to resolving technological uncertainty, or is it purely aesthetic?

Being an Early Adopter Is Not Enough

I hear this regularly: “we’re the only practice doing this, we can’t find a consultant who can do it, we’re learning it as we go, surely that’s R&D?”

Being an early adopter does not, by itself, create technological uncertainty within the meaning of CTA 2009. That the AI tool was new to your practice, that you had to learn it, that no-one else locally was doing it: none of that is the point. The novelty that matters under the legislation is in the problem, not in the tool.

Where practices are using AI as part of a wider R&D workflow (to iterate faster and land on a solution to a genuine technical problem) the AI-related costs can be included in the claim. The tool does not disqualify the work. It also does not automatically qualify it.

Software Costs vs. Staff Costs

A practical point worth noting. Under both the ERIS (Enhanced R&D Intensive Support) scheme and the standard merged scheme, software costs qualify where the software is directly used in qualifying R&D activity. Subscriptions to AI tools used in work that genuinely meets the technological uncertainty test should be captured in the claim.

Staff time spent operating those tools, iterating through models and interpreting outputs in the context of resolving a defined technical problem is claimable as a staffing cost. For most practices, that is where the bulk of the value in an AI-related claim will sit.

The subscription cost of a tool used exclusively for client visualisations or marketing CGIs does not qualify. The software cost follows the same eligibility logic as the staff cost.

Record It Now

If your practice is spending meaningful time on AI-assisted design work, the single most useful thing you can do is record the purpose of that work at the project level, not just the time spent. “Ran 40 AI massing iterations to find a compliant solution to the GLA carbon constraint on a constrained conservation area site” is a very different contemporaneous note to “produced visualisations for client review.” One of those survives an HMRC enquiry. The other does not.

Do not assume aesthetic work is excluded across the board. Where design directly constrains or shapes the technical solution, as it genuinely does in some architectural R&D scenarios, there is an apportionment argument to be made. Make it with contemporaneous records behind it.

And do not overclaim. HMRC has significantly increased enquiry volumes across the sector over the past two years. The practices that will have problems are those where consultants claimed everything and justified nothing. Properly documented, technologically-grounded claims come out of enquiries intact.

If you want to understand what your AI spend qualifies for, I’m happy to have that conversation. We’ve been working with architectural practices on R&D claims since 2009.

Legislative references:
Corporation Tax Act 2009 (CTA 2009) s.1042 — definition of R&D for tax purposes
HMRC Corporate Intangibles Research and Development Manual, CIRD81900 — qualifying activities, aesthetic exclusion
HMRC CIRD82200 — directly contributing activities
Finance Act 2023 — ERIS scheme provisions for R&D-intensive SMEs