If you run an AI startup in the UK, you have probably heard that R&D tax credits are for labs and pharmaceutical companies. That is not true. Software development, including AI and machine learning projects, has qualified for R&D relief for years. The question is whether your specific project meets HMRC's definition of research and development.

The short answer is yes: an AI startup can claim R&D tax credits if its work involves resolving technological uncertainty. That means you are trying to achieve an advance in science or technology that is not readily deducible by a competent professional working in your field. Simply integrating an existing API or fine-tuning a pre-trained model using standard techniques does not qualify. Building a novel neural network architecture from first principles because no existing solution solves your problem almost certainly does.

As ICAEW qualified accountants, we have submitted dozens of claims for AI and software startups. The most common mistake we see is assuming all AI work qualifies, or conversely, assuming none of it does. The truth sits somewhere in the middle. Let us walk through the rules, the qualifying costs, and the traps to avoid.

What HMRC Actually Means by R&D for AI

HMRC defines R&D for tax purposes using the Guidelines on the Meaning of Research and Development. These are not the same as accounting standards (IAS 38) or grant definitions. For tax relief, your project must meet two conditions:

  • It seeks an advance in science or technology.
  • It involves resolving technological uncertainty that a competent professional could not resolve using standard practice.

For AI startups, the advance is usually in the field of computer science or machine learning. It could be a new algorithm, a novel application of a technique to a domain where it has never been applied before, or a way to achieve performance metrics that current approaches cannot reach. The key word is "advance". If you are doing something that has been done before, even if it is new to your company, it does not qualify.

Technological uncertainty means you did not know at the start whether the project was feasible. You had to experiment, test hypotheses, iterate, and fail before finding a solution. If you knew exactly how to build it from day one, it is not R&D.

Real Examples of Qualifying AI Work

Let us make this concrete. A startup building a computer vision system to detect defects in manufactured goods on a production line might qualify if the lighting conditions, speed of the line, and variety of defects mean no off-the-shelf model works. The team has to develop a custom architecture, generate synthetic training data, and experiment with loss functions to achieve acceptable accuracy. That involves technological uncertainty.

A startup using GPT-4 to generate marketing copy, with some prompt engineering and fine-tuning on their own data, almost certainly does not qualify. The uncertainty is commercial, not technological. You know the model works. You just need to apply it. That is not R&D.

The line can be fine. A startup building a recommendation engine for a niche industry with sparse data might qualify if they have to develop a novel approach to handle cold-start problems or incorporate domain-specific constraints. A startup using a standard collaborative filtering library with default settings does not qualify.

Qualifying Costs for AI Startups

If your project qualifies, the next question is what costs you can include. For a typical AI startup, the main categories are:

  • Staff costs: salaries, employer NI, and pension contributions for employees directly working on the R&D project. This includes data scientists, machine learning engineers, and software developers. You can apportion their time if they work on both R&D and non-R&D tasks.
  • Externally provided workers: contractors and agency staff working on the project. You can claim 65% of the payments to the agency or intermediary, not the gross pay to the individual.
  • Consumables: cloud computing costs for training models, GPU time, data storage, and datasets purchased specifically for the R&D project. HMRC accepts cloud costs as consumables if they are directly used up in the R&D work.
  • Software licenses: if the software is consumed in the R&D process (e.g., a specialised simulation tool or a dataset license), it may qualify. Standard development tools like IDEs or version control systems do not.
  • Subcontractors: if you subcontract R&D work to a third party, you can claim 65% of the payment. This is common for AI startups that outsource parts of the model development.

One trap we see repeatedly: startups try to claim the entire salary of a founder who spends 20% of their time on R&D and 80% on running the business. You can only claim the proportion of time actually spent on the qualifying project. Keep timesheets or contemporaneous records to support your apportionment.

Cloud Computing and GPU Costs

Cloud computing is a major cost for AI startups. Training large models on AWS, Azure, or Google Cloud can run into tens of thousands of pounds. HMRC accepts these costs as consumables, but only for the portion used directly in R&D activities. If you use the same cloud account for hosting your production application, you need to separate the costs.

We recommend setting up a separate cloud account or project for R&D work. Tag your resources, track usage by instance, and keep invoices that show the specific compute hours used for training runs. HMRC has pushed back on claims where cloud costs were not clearly attributable to R&D.

The Two R&D Schemes: Which Applies to Your Startup?

For accounting periods starting on or after 1 April 2024, the old SME and RDEC schemes were merged into a single scheme for most companies. However, there is a special provision for R&D intensive loss-making SMEs.

Merged R&D scheme (standard): For most companies, the rate is 20% payable credit on qualifying R&D expenditure. This applies to both profit-making and loss-making companies. If you are profit-making, you reduce your corporation tax liability. If you are loss-making, you can surrender the loss for a cash payment.

Enhanced R&D Intensive Scheme (ERIS): If your company is loss-making and at least 30% of your total expenditure is on R&D, you may qualify for a higher rate of payable credit. The rate is 27% on qualifying R&D expenditure for the intensive scheme. This is particularly relevant for early-stage AI startups that are burning cash on R&D and have little or no revenue.

To qualify for ERIS, you must meet the intensity condition in the relevant accounting period. That means your R&D expenditure (including the enhanced element) must be at least 30% of your total expenditure. HMRC publishes guidance on how to calculate total expenditure. It includes cost of sales, administrative expenses, and other operating costs, but excludes certain items like depreciation and amortisation.

Example: An AI Startup Claiming Under ERIS

Consider a London-based AI startup developing a novel natural language processing tool for legal document review. The company has three employees, all working on R&D. Total qualifying R&D expenditure for the year is £120,000. Total expenditure (excluding certain items) is £350,000. R&D intensity is 34%, so the company qualifies for ERIS.

The payable credit is 27% of £120,000, which equals £32,400. That is a cash payment from HMRC, received approximately 6 to 8 weeks after filing the corporation tax return and R&D claim. For a cash-burning startup, that is significant.

If the same company did not meet the 30% threshold, the credit would be 20% of £120,000, or £24,000. The difference is £8,400. It pays to check whether you qualify for the intensive scheme.

Common Pitfalls in AI R&D Claims

We have seen the same mistakes repeated across dozens of AI startup claims. Here are the ones to avoid:

  • Claiming for routine development. Adding a new feature to an existing product using standard techniques is not R&D. The advance must be in science or technology, not in your product roadmap.
  • No supporting evidence. HMRC can and does open enquiries into R&D claims. You need technical reports, project records, and timesheets to justify the claim. We recommend writing a technical narrative for each project as you go, not retrospectively.
  • Including non-qualifying costs. Marketing, sales, and general overheads do not qualify. Neither does the cost of acquiring training data that is already publicly available, unless you had to clean or structure it in a novel way.
  • Overclaiming cloud costs. Only the portion of cloud spend used directly in R&D qualifies. Production hosting, monitoring, and CI/CD pipelines do not qualify.
  • Ignoring the associated company rules. If you control other companies, their R&D spend may be aggregated with yours for the purpose of the intensity test and the marginal relief thresholds. This can push you out of the intensive scheme or into the main rate of corporation tax.

What HMRC Asks When They Open an Enquiry

If HMRC selects your claim for an enquiry, they will typically ask for:

  • A description of the scientific or technological advance you sought.
  • The technological uncertainties you faced and how you resolved them.
  • The qualifications of the competent professionals involved.
  • Project records, including design documents, experiment logs, and version control history.
  • Timesheets or other evidence of staff time spent on R&D.
  • Invoices for consumables and cloud costs, clearly linked to the R&D projects.

We have handled enquiries where HMRC accepted the claim after reviewing detailed technical documentation. We have also seen claims rejected because the startup could not articulate the advance or the uncertainty. The difference is preparation.

How to Structure Your Claim

If you decide to claim, the process is straightforward but requires care:

  1. Identify qualifying projects. Go through your development work for the accounting period. Separate projects that involved technological uncertainty from routine development.
  2. Quantify qualifying costs. Calculate staff time, contractor payments, cloud costs, and other consumables attributable to each project.
  3. Prepare a technical report. For each project, write a narrative explaining the advance, the uncertainty, and the work done to resolve it. Keep this in your records even if you do not submit it with the claim.
  4. Complete the claim on your CT600. The R&D claim is made in the corporation tax return. You will need to complete the relevant boxes and, for claims from April 2024 onwards, submit the Additional Information Form (AIF) digitally.
  5. Submit the AIF before the CT600. HMRC requires the AIF to be submitted before or at the same time as the CT600. The AIF includes project-level details and cost breakdowns.

If you are unsure whether your projects qualify, we recommend a free initial review. We can look at your development roadmap and tell you which projects are likely to meet HMRC's definition. You can contact our team to arrange that.

When to Start Thinking About R&D

The best time to think about R&D tax credits is before you start the project. Set up your record keeping from day one. Use a project management tool to log experiments, failures, and decisions. Track time spent by project. Keep cloud cost reports segmented by R&D and non-R&D work.

If you are already partway through a project, it is not too late. You can reconstruct the work from version control history, design documents, and team recollections. But it is harder, and HMRC will scrutinise retrospective claims more closely.

For more detail on the mechanics of the claim process, read our guide to R&D tax credits for UK companies. It covers the full process from eligibility to submission.

Final Thoughts

AI startups are well placed to claim R&D tax credits, provided they are doing genuine research and development rather than routine application of existing techniques. The relief can be worth tens or hundreds of thousands of pounds in cash or reduced tax, which is significant for an early-stage company.

The key is to be honest about what qualifies and what does not. Overclaiming invites an enquiry and potential penalties. Underclaiming leaves money on the table. A balanced, well-documented claim is the right approach.

If your turnover is approaching the VAT threshold or you are considering incorporating, those are separate decisions. But if you are spending significant time and money on developing novel AI solutions, you should almost certainly be claiming R&D tax credits.