R&D Tax Relief: Why Your Machine Learning/AI Project Might Not Qualify

As an R&D Tax Specialist with extensive HMRC experience, I've witnessed numerous companies invest heavily in AI and ML projects, confident they'll secure R&D tax relief, only to be bewildered when they don't qualify.

Here's the catch: while AI and ML are indeed groundbreaking, their foundations aren't as new as many believe. Many of today's algorithms build on ideas developed decades ago. Yes, we've dramatically improved computing power and data availability, but the core principles often aren't as novel as the hype suggests. This fact frequently trips up R&D tax relief claims.

So, what exactly is HMRC looking for when it comes to AI and ML projects, and how can companies ensure their innovative work meets the stringent criteria for R&D tax relief?

At first glance, it seems reasonable to assume that any project using ML or AI would qualify for R&D Tax Credits. After all, these technologies are complex and potentially groundbreaking. However, recent conversations with HMRC reveal a more nuanced perspective.

During a recent HMRC enquiry, a case officer posed a thought-provoking question:

If companies are simply training algorithms with their own data, where's the technological advancement?

This query cuts to the heart of a crucial disconnect between the perceived innovation of AI/ML and HMRC's practical assessment criteria.

To qualify for R&D tax relief, projects must demonstrate a genuine advancement in science or technology that goes beyond the company's existing knowledge or capabilities. It's not enough to merely adopt AI and ML or do fine-tuning exercises. Companies must show they're contributing to the advancement of AI and Machine Learning as a whole. This high bar ensures that R&D tax relief rewards truly transformative development efforts, not just the application of existing technologies in new contexts.

In practice, most machine learning projects don't involve creating novel algorithms from scratch. Instead, they typically integrate proprietary data into existing ML frameworks like TensorFlow or PyTorch. Similarly, AI projects often leverage ready-made frameworks such as OpenAI, using them as a foundation for applications like ChatGPT-powered chatbots.

This reliance on pre-built solutions shifts the innovation focus from the underlying algorithms to their application and integration with proprietary data. However, this presents a challenge: the knowledge of how to apply ML is well-established and widely available. HMRC, inundated with thousands of such projects monthly, doesn't view parameter fine-tuning or the application of existing algorithms as a significant technological advancement.

This stance is reinforced by HMRC guidelines. CIRD 81900 para 8 states:

A process, material, device, product, service or source of knowledge does not become an advance in science or technology simply because science or technology is used in its creation. Work which uses science or technology but which does not advance scientific or technological capability as a whole is not an advance in science or technology.

Furthermore, CIRD 81900 para 14 emphasises:

Uncertainties that can be readily resolved by a competent professional working in the field are not scientific or technological uncertainties. Similarly, improvements, optimisations and fine-tuning which do not materially affect the underlying science or technology do not constitute work to resolve scientific or technological uncertainty.

In addition, as stipulated in CIRD 81900 para 22:

...the routine analysis, copying or adaptation of an existing process, material, device, product or service will not advance overall knowledge or capability, even though it may be completely new to the company or the company’s trade.

Therefore, true R&D in the context of AI and ML involves advancing the field through novel contributions, such as creating new algorithms, developing novel and robust data processing techniques, or developing new methods for data collection that specifically improve AI/ML capabilities as a whole, thus surpassing current limitations within the tech industry.

It is crucial to consider the broader picture when evaluating projects:

• Am I tackling problems that have not been adequately addressed by existing approaches, or am I building upon established methodologies?
  ‍
• How am I pushing the boundaries of traditional AI/ML approaches and exploring alternative methodologies?
  ‍
• Are the algorithms I have developed advancing the field of technology as a whole, or am I merely ingesting data and training the algorithms to reach a high level of accuracy?
  ‍
• To what extent does the methodology I've employed demonstrate innovation, and what level of uncertainty is present within the systems I’ve used?

While HMRC may not be experts in AI and ML, they review countless such projects daily. This means they've developed a keen eye for distinguishing truly innovative work from routine applications. To stand out, it's crucial to adhere strictly to CIRD guidelines and conduct a thorough project assessment with R&D tax specialists from the outset.

Companies claiming R&D tax relief should focus on demonstrating how their work significantly advances industry standards, not just their own capabilities. Be prepared to robustly defend your project's innovative aspects in the face of potential HMRC scrutiny. Remember, it's not just about using cutting-edge technology—it's about pushing its boundaries in ways that contribute to the field as a whole.