'Human driver' still needed for AI R&D claims

Artificial intelligence is not yet ready to replace humans when it comes to writing R&D tax claims, but it is surprisingly clever. Three R&D specialists discuss the pros and cons

Following an article on the risks of uncontrolled use of AI in Business & Accountancy Daily, Brian Williamson, chief strategy officer at AI driven R&D firm, Kreoh, and his boss, co-founder Garry Tiscovschi, met AI sceptic Ele Theochari, partner at Blick Rothenberg, to exchange views on the subject.

They flew to London from their Dublin base for an exchange of views, prepared to agree to disagree. As their discussion developed, however, it became clear that they agreed on fundamental principles: ensuring accuracy, expert-control, and the responsible use of AI in R&D tax claims.

There was strong agreement that the expert writer should remain in the ‘driver’s seat’ when producing reports. They also shared concerns about R&D consultants using AI as a ‘black box’ that generates what Theochari described as ‘a lot of words with very little substance' and Kreoh called ‘AI gruel’.

Instead, the three advocated for AI as an ‘engine for the brain’, enhancing human expertise rather than becoming a substitute for it.

In Kreoh’s view, an engine is a system that includes AI as a component, built to be an extension of the consultant, not a crutch. But what’s the difference? How can AI be used responsibly in this field?

Take two approaches to an AI assisted writer investigating a technology baseline for their report:

Approach one: hit ‘generate’ and without substantive context, citations, audibility or expert guidance, the AI appends an estimated ‘technology baseline’ section to your report. This creates what HMRC scrutinises most, unverifiable, generic content lacking specific project details, ‘AI gruel’.

Approach two: the human driver is assisted by a controllable, user-knowledge aware AI agent. To avoid producing low substance ‘AI gruel’ the agent must be given the user’s knowledge, aka substance.

The consultant chooses what project data the AI uses: human notes, briefs, recordings, etc. Next the user informs the AI ‘research agent’ which external sources and date ranges they would like the agent to web-scrape for further context.

The AI agent conducts multiple parallel web searches on behalf of the user. This is done with more thoroughness and speed than a human could, but the final decision making remains with the human expert.

The AI agent submits a selection of relevant background research web-links (citation source) and accompanying summaries for the user to select from.

This is new AI-discovered information useful to the project, that the human audits and decides upon. The report is augmented by human selected data sources to produce a quality technology baseline.

Both approaches use AI. But one produces improvements to quality instead of ‘AI gruel’.

More common ground was agreed by both parties: accountability and integrity are crucial for the sustainability of the R&D tax credit system. If we want to preserve the benefits of the scheme, we must all play our part, not just by adhering to the rules but by actively promoting responsible practices.

‘This includes educating clients on the spirit of the law, not simply the letter, and maintaining rigorous internal checks to avoid questionable claims,’ Williamson stressed.

‘It is about creating a culture of honesty and respect for the intent behind the incentives. Only then can we ensure the scheme continues to serve its purpose of driving genuine innovation and growth.

‘If we are to maintain a robust and fair R&D tax system, it requires a joint effort from HMRC, providers, and companies alike to act responsibly and ethically.’

Williamson said the discussion had reinforced the principles of Kreoh’s system, highlighting that AI tools ‘must enhance, not replace, expert judgment in R&D claims’. The technology’s role is to augment the consultant’s capabilities, making them more thorough, efficient and consistent, while the final validation remains firmly in human hands.

Visit BrAInbox today where you can find answers to questions like Can I check a person's social media before hiring them?

• National insurance hike to 15% kicks in from 6 April
• Martyn's Law receives Royal Assent: what does this mean for your business?
• Court of Appeal rules whistleblowing protection does not apply to external job applicants
• NHS patients at risk as hospital urgent repair costs triple in a decade