Unlock Rewards from your AI Exploration:
Claim Your R&D Tax Credits

26th February 2024

Imagine getting paid by the government to explore the cutting-edge realm of Artificial Intelligence. Sounds intriguing, doesn't it?

If your business has been experimenting with AI, you may be able to claim cashback right now. Specifically, Research & Development (R&D) Tax Credits can pay you back 35% of whatever you have already spent on R&D.

My company, Claridian claimed back R&D Tax Credits for the creation of our AI robot. It was a straightforward exercise. So, don't dismiss this blog as just another mundane tax topic. Instead, let me explain how we claimed.

Our R&D Tax Credit claim

Claridian uses Artificial Intelligence to help companies:

understand R&D tax credits,
verify they meet the requirements, and
then complete the application itself.

We started using AI in our own business in January 2023, licensing the OpenAI API in April 2023. This sets the start of our R&D project.

Taxes are separately calculated for each financial year. Our financial year starts on 1st April and thus our first claim is for Financial Year (FY) 2023/24 … which has not finished yet.

Requirement 1: The R&D Project

Ideally, your R&D takes the form of a separate project with a specific goal or desired outcome. Maybe you never referred to it as a separate project, but it was above and beyond “business as usual”.

You were trying to create something new or improved:

a new product or service; or
an improved process:
cheaper,
faster,
higher quality,
more efficient,
less wasteful,
whatever.

If you use tools to provide your service, you can improve the tools or processes. If you use materials, you can try different materials or a different way of using existing materials.

In our case, our exploration of AI is an R&D project. We had a specific goal: to replace what was previously an entirely human process with an AI-led approach. If we succeeded, there would be an “appreciable improvement” to our company’s “capabilities”.

Cosmetic or aesthetic improvements are not the sort of innovation you can claim for because they are not an improvement in your capabilities. Indeed, routine or minor adaptation, upgrading, or bug fixing are not the sort of innovation that the government is interested in.

In our case, our goal – a fundamentally different way of translating the details of business projects into tax claims – is certainly ambitious. For us, we were not looking to use AI as a marketing gimmick. Rather, we were looking to change our business.

PRO TIP: Running your R&D as a separate project helps emphasise that the work was not "business as usual". It also helps when it comes time to detail the specific expenses associated with the R&D.

Requirement 2: Uncertainty

No matter how remarkable the improvement to your business, if the way to get there was obvious or "readily deducible" then it was not innovation.

Imagine if you summited Mt Everest, that would be an amazing personal achievement. However, it would not be remarkable in the wider sense because the path to the top is known; the ropes are in place; experienced Sherpas, who have done it before, are on hand to help you every step of the way.

In a nutshell, there has to be “uncertainty”. The word “uncertainty” appears 54 times in the “43 conditions to be satisfied” to successfully claim R&D tax credits.

Ideally, you were uncertain in 2 different ways:

First, success itself should have been uncertain – because you didn’t know whether it was possible or feasible (i.e. it could be done cost-effectively).

Note: This is not a requirement. HMRC specifically says that if a competitor has already achieved the goal, but you don't know how they did it, then this is still R&D.

Second, the approach you need to take should have been uncertain because it was not easy to work out what needed to be done. Importantly, not everything needed to be uncertain; but, at least one important aspect needed to be.

A good example of this is “system uncertainty” – how each of the individual components work is understood, but it is uncertain as to the best way to combine those elements to achieve the effect.

In our case, generative AI was an entirely new tool. No one had attempted before to apply it providing tax advice and draft tax submissions. We had no idea whether the result would be workable.

Generative AI also was evolving rapidly. Prompt engineering was a new discipline. Social media was flooded with random hints and tips, but we did not know where to start, how to approach the problem, or even what makes a good prompt.

One challenge we faced with early LLMs was that our clients would describe the problems and complexity of their project. Rather than assessing whether that complexity met HMRC's requirements, the LLM would offer suggestions for how to advance the project.

A later challenged related to affirmation bias within LLMs. If you ask an LLM if the requirements are met, it will say, "Yes". However, if you ask it to list how the requirements are not met, it will happily generate such a list.

Both of these examples demonstrate technical uncertainty (around generative AI) and systemic uncertainty (how to integrate generative AI into the user journey). With that said, as newer versions of LLMs became available, we were not uncertain that our goal was unachievable but there was still uncertainty around exactly how to do it.

PRO TIP: Uncertainty often manifests as false starts and wasted effort. If you don't make any mistakes then the approach was probably "readily apparent" and thus you were not doing R&D.

Requirement 3: Competent Professional

To qualify, the uncertainty you faced should be wider than just your company. The answer could not have been in a widely available book (or a blog post) that no one in your company had bothered to read. The answer could not have been “readily available or deducible by a competent professional working in the field”.

This hurdle is not as high as it might initially seem. The hypothetical professional certainly need not be a Nobel prize winner. Rather, it should be an average employee who is up-to-date on all the normal training and has the normal level of experience.

Similarly, this competent professional need not have trade secrets from a competitor. It does not matter if a competitor has already achieved what you are seeking to achieve, as long as how they do it is not generally known.

In our case, there were no competent professionals. No one (not even the inventors of Chat GPT) fully understood how it worked, and there were no definitive guides to prompt engineering. Interesting, I have a computer science degree from MIT and took classes on Artificial Intelligence but the theory and the intricacies of use are fundamentally different things.

There is an important distinction here. We did not license someone else’s implementation of AI – for example, a customer service chatbot with an AI built-in.

In our case, we wrote the prompts ourselves … and when that proved insufficient, we created our own knowledge base to support Retrieval-Augmented Generation (RAG). We have posited that any competent programmer in a relevant language is a “competent professional” for prompt engineering.

However, our case is not the bar you need to straddle. It qualifies if you took an existing tool (e.g. a customer services chatbot with an AI built-in) and then adapted it to your situation in a non-trivial way … in a way that involved choices, and trial and error … where there was uncertainty.

PRO TIP: Arguably, by the end of the project you were a competent professional. The issue here is: How much of the false starts and wasted effort was because you were not up-to-date on current best practice, and how much involved genuinely pushing beyond.

In the last example given above, the time spent figuring out how to use a customer service chatbot is not R&D. However, bending that chatbot to the issues of your clients very well might be.

Final Requirements

If the above sounds like your situation, then it is very likely that you qualify for R&D tax credits. There are a few more points:

You need to be incorporated;
You need to be an SME – so, less than 250 employees; and
The work could not have been subsidised with other government funds (e.g. a Knowledge Transfer Partnership).

What was Claimed

The main expenses you can reclaim are:

staff costs (whether salaried employees, agency staff, or subcontractors);
computer costs (hardware, software, or data licenses … the amount “consumed”). Cloud hardware costs are easy to account for. Hardware (or software) purchased for R&D that is then used afterwards is allowed but a bit more complicated; and
anything else consumed by the R&D – fuel, materials, power, water

The most notable expense that cannot be claimed is real estate – rent, rates, and the cost of land.

In our case, we claimed 100% of relevant staff salaries from the start of the project until “the uncertainty was resolved”. This is an important concept. The uncertainty is resolved once all the substantive questions are answered and those answers are “codified in a form usable by a competent professional working in the field”.

With that said, a given project might have multiple uncertainties. In this case, the R&D project continues as long as you are working on those uncertainties and until they are resolved.

Interestingly, creating and testing a prototype qualifies, but the R&D ends once the solution has been tested sufficiently to prove that the uncertainty has been resolved. Any cosmetic changes after that do not qualify as R&D … unless some sort of technological advance is required to achieve that cosmetic improvement.

HMRC gives a specific example of a company creating a prototype, making copies, distributing it to consumers, eliciting feedback, and then acting on that feedback. The work to create the first prototype was R&D. The work to fix the technical problems uncovered was R&D. However, the work to elicit feedback to uncover problems is not R&D because there is no attempt to advance the technology during this work.

In our case, we claimed 100% of our computing costs (AWS cloud hardware, Open AI licenses) up until the point that the prototype was operational and testing was complete. Work after that point – reskinning the web page to make it more user-friendly – was not R&D.

PRO TIP: Think of each technical challenge as a separate R&D project. Doing so helps you identify the specific false starts and wasted effort, but this also gets you away from thinking that R&D has to be as grand as putting man on the moon.

Preparing the Claim

As you begin to consider your own R&D tax credit claim, be aware that there are a few additional complexities. While our discussion has provided insight into the fundamentals, navigating the intricacies of HMRC's requirements may require additional expertise.

For example, business owners typically start with a business problem and then discuss how they resolved it. An R&D tax claim comes at this situation in reverse. Your claim starts by naming a field of science or technology and defining the baseline before your R&D project.

The reality is that this is a matter of presentation rather than substance, and ...

This is precisely where Claridian shines – equipped with generative AI capabilities, it can streamline the translation of your innovations into language that HMRC readily comprehends and approves.

Ready to unlock the potential of your AI exploration? Reach out to us today and let's embark on this rewarding journey together.

https://www.claridian.co.uk/r-and-d-tax-credit-assessment

Disclaimer: Each R&D claim is assessed on its own merits. The above are guidelines. Many of the terms are not well defined and their meaning is best deduced from relevant case law.

Unlock Rewards from your AI Exploration: Claim Your R&D Tax Credits