This paper is relevant to anyone in the UK with a stake in how innovation is funded, valued or evidenced, whether you're working on an R&D claim, signing off on one, modelling it into a valuation, or shaping the policy framework around it. That means CFOs, Heads of Tax, CTOs, PE portfolio teams, startup founders and the industry bodies lobbying for reform. The implications touch everyone involved, from claim owners and advisors to investors and policymakers.This paper is relevant to anyone in the UK with a stake in how innovation is funded, valued or evidenced, whether you're working on an R&D claim, signing off on one, modelling it into a valuation, or shaping the policy framework around it. That means CFOs, Heads of Tax, CTOs, PE portfolio teams, startup founders and the industry bodies lobbying for reform. The implications touch everyone involved, from claim owners and advisors to investors and policymakers.

The Environment Is ShiftingThe Environment Is Shifting

Over the past eighteen months, artificial intelligence has moved from productivity aid to primary author of production code. According to Stack Overflow's 2025 Developer Survey, 84% of professional developers now use or plan to use AI coding tools, with 51% using them daily. GitHub reports that its Copilot tool now generates 46% of code written by active users, up from 27% at launch. Both Microsoft CEO Satya Nadella and Google CEO Sundar Pichai have claimed that around a quarter of their companies’ code is now AI-generated. Developers report time savings of 30–60% on routine coding, testing and documentation.Over the past eighteen months, artificial intelligence has moved from productivity aid to primary author of production code. According to Stack Overflow's 2025 Developer Survey, 84% of professional developers now use or plan to use AI coding tools, with 51% using them daily. GitHub reports that its Copilot tool now generates 46% of code written by active users, up from 27% at launch. Both Microsoft CEO Satya Nadella and Google CEO Sundar Pichai have claimed that around a quarter of their companies’ code is now AI-generated. Developers report time savings of 30–60% on routine coding, testing and documentation.

For businesses claiming R&D tax relief under UK's DSIT guidelines, these numbers demand attention. The statutory test for qualifying R&D has not changed. But the practical baseline against which technological uncertainty is assessed is shifting materially. This paper considers what that shift means for claims strategy, policy risk and governance.For businesses claiming R&D tax relief under UK's DSIT guidelines, these numbers demand attention. The statutory test for qualifying R&D has not changed. But the practical baseline against which technological uncertainty is assessed is shifting materially. This paper considers what that shift means for claims strategy, policy risk and governance.

The Competent Professional Test Under PressureThe Competent Professional Test Under Pressure

The DSIT Guidelines define qualifying R&D as work that seeks to achieve an advance in science or technology through the resolution of scientific or technological uncertainty. For software companies, the test has always hinged on whether a competent professional could readily determine the solution. If they could, the work falls outside R&D. If genuine uncertainty existed, the investigation to resolve it could qualify.The DSIT Guidelines define qualifying R&D as work that seeks to achieve an advance in science or technology through the resolution of scientific or technological uncertainty. For software companies, the test has always hinged on whether a competent professional could readily determine the solution. If they could, the work falls outside R&D. If genuine uncertainty existed, the investigation to resolve it could qualify.

That test made sense in a world where a competent professional was a senior developer with a decade of experience, access to documentation, and a reasonable awareness of the state of the art. It made sense when integrating two complex financial systems involved genuine technical unknowns around data consistency, real-time reconciliation and fault tolerance across disparate architectures. It equally made sense when building a scalable microservices platform demanded novel engineering to handle unpredictable load patterns, or when constructing a real-time data processing pipeline required solving latency and throughput problems for which no established approach existed.That test made sense in a world where a competent professional was a senior developer with a decade of experience, access to documentation, and a reasonable awareness of the state of the art. It made sense when integrating two complex financial systems involved genuine technical unknowns around data consistency, real-time reconciliation and fault tolerance across disparate architectures. It equally made sense when building a scalable microservices platform demanded novel engineering to handle unpredictable load patterns, or when constructing a real-time data processing pipeline required solving latency and throughput problems for which no established approach existed.

The question now is whether that same test produces the same outcomes when the competent professional has access to an AI tool trained on the accumulated knowledge of the entire software industry. A developer can describe an integration challenge in plain English and receive working code within seconds, drawing on training data that spans millions of repositories. Problems that once took teams of engineers several sprints to resolve are being addressed in an afternoon by a single developer working alongside an AI agent. The definition of “readily deducible” shifts accordingly.The question now is whether that same test produces the same outcomes when the competent professional has access to an AI tool trained on the accumulated knowledge of the entire software industry. A developer can describe an integration challenge in plain English and receive working code within seconds, drawing on training data that spans millions of repositories. Problems that once took teams of engineers several sprints to resolve are being addressed in an afternoon by a single developer working alongside an AI agent. The definition of “readily deducible” shifts accordingly.

What Happens to the Baseline?What Happens to the Baseline?

Claimants are expected to demonstrate that their work goes beyond what is already known or achievable in the field. Previously, a literature review might involve searching documentation, consulting peers, or reading academic papers. It was reasonable to miss things. Gaps in available knowledge created genuine uncertainty. Today, AI systems can synthesise information from millions of sources instantaneously. The idea that a solution is not readily deducible becomes increasingly difficult to defend when any developer with access to modern coding tools can surface approaches that would have taken weeks of manual research to uncover just two years ago.Claimants are expected to demonstrate that their work goes beyond what is already known or achievable in the field. Previously, a literature review might involve searching documentation, consulting peers, or reading academic papers. It was reasonable to miss things. Gaps in available knowledge created genuine uncertainty. Today, AI systems can synthesise information from millions of sources instantaneously. The idea that a solution is not readily deducible becomes increasingly difficult to defend when any developer with access to modern coding tools can surface approaches that would have taken weeks of manual research to uncover just two years ago.

Consider the practical implications. A payment gateway integration that required a team to wrestle with undocumented API behaviour, concurrent transaction handling and edge cases in currency conversion involved real technological uncertainty. A compliance workflow engine built to satisfy novel regulatory requirements under tight performance constraints demanded genuine experimentation. These were substantive engineering efforts where the outcome could not be predicted in advance. The concern is not that these challenges were improperly claimed. It is that many of the same problems, which were genuinely difficult at the time, may no longer clear the bar. The patterns have been solved, documented and absorbed into AI training data. The uncertainty has been genuinely eroded by the pace of technological change.Consider the practical implications. A payment gateway integration that required a team to wrestle with undocumented API behaviour, concurrent transaction handling and edge cases in currency conversion involved real technological uncertainty. A compliance workflow engine built to satisfy novel regulatory requirements under tight performance constraints demanded genuine experimentation. These were substantive engineering efforts where the outcome could not be predicted in advance. The concern is not that these challenges were improperly claimed. It is that many of the same problems, which were genuinely difficult at the time, may no longer clear the bar. The patterns have been solved, documented and absorbed into AI training data. The uncertainty has been genuinely eroded by the pace of technological change.

The market is beginning to reflect this. In early February 2026, software equities saw a sharp sell-off amid investor concern that AI could compress the value of commodity SaaS workflows, erasing an estimated $285–300 billion in market capitalisation according to Bloomberg and Wall Street Journal. The significance for R&D is not the valuation correction itself, but what it signals: if AI can replicate the functionality of entire SaaS products and automate implementations that previously took years, then a material portion of what the software industry classified as R&D may no longer involve genuine technological uncertainty. When Microsoft’s CEO describes SaaS applications as CRUD databases with business logic that will migrate to the AI layer, the implications for the uncertainty test are difficult to ignore.The market is beginning to reflect this. In early February 2026, software equities saw a sharp sell-off amid investor concern that AI could compress the value of commodity SaaS workflows, erasing an estimated $285–300 billion in market capitalisation according to Bloomberg and Wall Street Journal. The significance for R&D is not the valuation correction itself, but what it signals: if AI can replicate the functionality of entire SaaS products and automate implementations that previously took years, then a material portion of what the software industry classified as R&D may no longer involve genuine technological uncertainty. When Microsoft’s CEO describes SaaS applications as CRUD databases with business logic that will migrate to the AI layer, the implications for the uncertainty test are difficult to ignore.

Where Genuine R&D Still LivesWhere Genuine R&D Still Lives

None of this means software R&D is dead. The centre of gravity is shifting, and businesses that understand where genuine uncertainty still exists will be better positioned.None of this means software R&D is dead. The centre of gravity is shifting, and businesses that understand where genuine uncertainty still exists will be better positioned.

Greenfield projects in regulated environments remain firmly in R&D territory. Government systems, defence applications and critical national infrastructure often mandate that every line of code be human-written, explainable and auditable. AI-generated code is frequently prohibited by procurement standards or security requirements. The technological uncertainty in these environments is real, the constraints genuine, and the work involves pushing boundaries that AI tools cannot yet navigate.Greenfield projects in regulated environments remain firmly in R&D territory. Government systems, defence applications and critical national infrastructure often mandate that every line of code be human-written, explainable and auditable. AI-generated code is frequently prohibited by procurement standards or security requirements. The technological uncertainty in these environments is real, the constraints genuine, and the work involves pushing boundaries that AI tools cannot yet navigate.

Frontier AI research itself qualifies by definition. Companies developing new model architectures, training methodologies or novel applications of machine learning to previously unsolved problems are operating at the edge of scientific knowledge. Similarly, work that requires novel scientific or mathematical approaches, complex optimisation challenges, new cryptographic methods, or entirely new computational paradigms is likely to continue qualifying.Frontier AI research itself qualifies by definition. Companies developing new model architectures, training methodologies or novel applications of machine learning to previously unsolved problems are operating at the edge of scientific knowledge. Similarly, work that requires novel scientific or mathematical approaches, complex optimisation challenges, new cryptographic methods, or entirely new computational paradigms is likely to continue qualifying.

Equally important, AI is opening doors for companies that were never previously in scope. Businesses in logistics, agriculture, professional services and creative industries are now using machine learning to tackle problems that were simply beyond their technical reach before. A mid-sized logistics firm building a novel route optimisation model, or a legal services company developing proprietary natural language processing, may well be undertaking work that meets the DSIT threshold. Some qualifying opportunities are disappearing. But others are emerging in places where R&D was never previously considered.Equally important, AI is opening doors for companies that were never previously in scope. Businesses in logistics, agriculture, professional services and creative industries are now using machine learning to tackle problems that were simply beyond their technical reach before. A mid-sized logistics firm building a novel route optimisation model, or a legal services company developing proprietary natural language processing, may well be undertaking work that meets the DSIT threshold. Some qualifying opportunities are disappearing. But others are emerging in places where R&D was never previously considered.

What the Numbers ReflectWhat the Numbers Reflect

HMRC’s September 2025 R&D statistics showed that total R&D claims fell to 46,950 for the 2023–24 tax year, a 26% drop from the previous year. SME claims fell by 31%. This follows a similar decline the year before. In two years, the number of companies claiming R&D relief has effectively halved.HMRC’s September 2025 R&D statistics showed that total R&D claims fell to 46,950 for the 2023–24 tax year, a 26% drop from the previous year. SME claims fell by 31%. This follows a similar decline the year before. In two years, the number of companies claiming R&D relief has effectively halved.

This decline has multiple causes, and it would be dishonest to attribute it solely to AI. A significant part of the story is the erosion of business confidence in the scheme itself. HMRC’s compliance campaign from 2023 saw enquiry volumes surge to the point where, according to CIOT, one in five claims was being investigated. The introductions to mandatory Additional Information Form, Claim Notification Form and Mandatory Random Enquiry Programme have improved scheme integrity, and the error and fraud rate has fallen from 17.6% in 2021–22 to 5.9% in the most recent estimates. That is a genuine achievement. But the collateral damage has been considerable. First-time claimants, according to the same HMRC's 2025 statistics, have collapsed from a peak of 19,720 to just 3,765. Claims under £15,000 have fallen most sharply. For many SMEs without the expertise or budget to navigate a lengthy HMRC review, the perceived risk now outweighs the benefit.This decline has multiple causes, and it would be dishonest to attribute it solely to AI. A significant part of the story is the erosion of business confidence in the scheme itself. HMRC’s compliance campaign from 2023 saw enquiry volumes surge to the point where, according to CIOT, one in five claims was being investigated. The introductions to mandatory Additional Information Form, Claim Notification Form and Mandatory Random Enquiry Programme have improved scheme integrity, and the error and fraud rate has fallen from 17.6% in 2021–22 to 5.9% in the most recent estimates. That is a genuine achievement. But the collateral damage has been considerable. First-time claimants, according to the same HMRC's 2025 statistics, have collapsed from a peak of 19,720 to just 3,765. Claims under £15,000 have fallen most sharply. For many SMEs without the expertise or budget to navigate a lengthy HMRC review, the perceived risk now outweighs the benefit.

The workforce data tells a related story. A Stanford Digital Economy Lab study found that employment for software developers aged 22–25 declined by nearly 20% between late 2022 and mid-2025, while older developers saw stable or growing employment. AI is particularly effective at replacing the kind of work that junior developers do: implementing known patterns, writing code from specifications, translating requirements into functional software. These are precisely the tasks around which many software R&D claims were built.The workforce data tells a related story. A Stanford Digital Economy Lab study found that employment for software developers aged 22–25 declined by nearly 20% between late 2022 and mid-2025, while older developers saw stable or growing employment. AI is particularly effective at replacing the kind of work that junior developers do: implementing known patterns, writing code from specifications, translating requirements into functional software. These are precisely the tasks around which many software R&D claims were built.

The falling claim volumes therefore reflect a confluence of pressures: tighter compliance, reduced confidence, lower SME relief rates, compressed qualifying activity and changing workforce composition. All five forces are pulling in the same direction.The falling claim volumes therefore reflect a confluence of pressures: tighter compliance, reduced confidence, lower SME relief rates, compressed qualifying activity and changing workforce composition. All five forces are pulling in the same direction.

Structural and Policy ImplicationsStructural and Policy Implications

The DSIT guidelines were last substantively updated in April 2023. The merged RDEC scheme and Enhanced R&D Intensive Support took effect from April 2024. These were meaningful reforms, but they were designed for a pre-AI world. Several structural issues now warrant serious attention.The DSIT guidelines were last substantively updated in April 2023. The merged RDEC scheme and Enhanced R&D Intensive Support took effect from April 2024. These were meaningful reforms, but they were designed for a pre-AI world. Several structural issues now warrant serious attention.

The definition of a competent professional needs revisiting. Should the baseline assumption include access to AI tools? If so, a substantial amount of work that previously involved genuine uncertainty may no longer qualify. If not, how long can the guidelines ignore the most significant shift in practitioner capabilities since the advent of high-level programming languages?The definition of a competent professional needs revisiting. Should the baseline assumption include access to AI tools? If so, a substantial amount of work that previously involved genuine uncertainty may no longer qualify. If not, how long can the guidelines ignore the most significant shift in practitioner capabilities since the advent of high-level programming languages?

The nature of qualifying expenditure is changing. Where a company’s R&D team once consisted of twelve developers, it may now comprise two or three senior engineers directing AI agents. The economics of claims shift from labour-intensive to tool-intensive, and the current framework does not account well for AI tooling costs. This has a further consequence for smaller businesses. Under both the merged scheme and ERIS, the PAYE cap limits payable R&D tax credit to £20,000 plus 300% of PAYE and NIC liabilities. For a lean, AI-augmented team, those liabilities may be insufficient to unlock the full value of a claim. Large corporates with substantial payrolls will barely notice. But for startups and small software houses, the cap becomes an increasingly tight constraint precisely as AI makes their R&D more capital-efficient. Loss-making SMEs relying on cash credits may find their claims capped at levels that fail to reflect the genuine scale of their research activity.The nature of qualifying expenditure is changing. Where a company’s R&D team once consisted of twelve developers, it may now comprise two or three senior engineers directing AI agents. The economics of claims shift from labour-intensive to tool-intensive, and the current framework does not account well for AI tooling costs. This has a further consequence for smaller businesses. Under both the merged scheme and ERIS, the PAYE cap limits payable R&D tax credit to £20,000 plus 300% of PAYE and NIC liabilities. For a lean, AI-augmented team, those liabilities may be insufficient to unlock the full value of a claim. Large corporates with substantial payrolls will barely notice. But for startups and small software houses, the cap becomes an increasingly tight constraint precisely as AI makes their R&D more capital-efficient. Loss-making SMEs relying on cash credits may find their claims capped at levels that fail to reflect the genuine scale of their research activity.

The UK is not alone in confronting these questions. In the United States, which operates one of the world’s largest R&D credit regimes under IRC §41, the advisory community is already working through the same structural tensions. A recent Bloomberg Tax analysis examined how AI-assisted development interacts with the US credit’s four-part qualifying test, particularly the “elimination of uncertainty” requirement, which serves a comparable function to the UK’s “readily deducible” standard. Their conclusion was not that AI eliminates qualifying uncertainty, but that it reframes it: the uncertainty shifts from the micro-level task of writing code to the macro-level challenge of engineering reliable systems from AI-generated outputs. The analysis also argued that the developer’s role is evolving from hands-on coder to systems architect and lead investigator, directing AI rather than replacing the research process. It noted that traditional documentation metrics such as lines of code are becoming less meaningful, and that firms will need to capture the cognitive effort behind architectural decisions, the evaluation of AI-generated alternatives, and the iterative testing that turns AI output into production-grade software.The UK is not alone in confronting these questions. In the United States, which operates one of the world’s largest R&D credit regimes under IRC §41, the advisory community is already working through the same structural tensions. A recent Bloomberg Tax analysis examined how AI-assisted development interacts with the US credit’s four-part qualifying test, particularly the “elimination of uncertainty” requirement, which serves a comparable function to the UK’s “readily deducible” standard. Their conclusion was not that AI eliminates qualifying uncertainty, but that it reframes it: the uncertainty shifts from the micro-level task of writing code to the macro-level challenge of engineering reliable systems from AI-generated outputs. The analysis also argued that the developer’s role is evolving from hands-on coder to systems architect and lead investigator, directing AI rather than replacing the research process. It noted that traditional documentation metrics such as lines of code are becoming less meaningful, and that firms will need to capture the cognitive effort behind architectural decisions, the evaluation of AI-generated alternatives, and the iterative testing that turns AI output into production-grade software.

No formal IRS guidance or case law has yet addressed AI-assisted work leading to R&D claims directly, so the US position remains interpretive rather than settled. But the direction of thinking is instructive. It suggests that jurisdictions with mature R&D incentive frameworks are arriving at similar questions about where human-led research activity sits when AI handles implementation. The UK’s DSIT guidelines operate on different statutory foundations, but the underlying tension is the same: how to distinguish genuine technological advance from efficient execution of known solutions. If the UK’s framework becomes too restrictive, penalising companies for adopting AI in their development process, it risks discouraging both innovation and honest claiming. If it remains too permissive, the Treasury faces a growing integrity problem. Getting the balance right matters, and the international conversation is moving faster than the UK’s domestic one.No formal IRS guidance or case law has yet addressed AI-assisted work leading to R&D claims directly, so the US position remains interpretive rather than settled. But the direction of thinking is instructive. It suggests that jurisdictions with mature R&D incentive frameworks are arriving at similar questions about where human-led research activity sits when AI handles implementation. The UK’s DSIT guidelines operate on different statutory foundations, but the underlying tension is the same: how to distinguish genuine technological advance from efficient execution of known solutions. If the UK’s framework becomes too restrictive, penalising companies for adopting AI in their development process, it risks discouraging both innovation and honest claiming. If it remains too permissive, the Treasury faces a growing integrity problem. Getting the balance right matters, and the international conversation is moving faster than the UK’s domestic one.

What This Means in PracticeWhat This Means in Practice

For boards, CFOs and technical leadership, the question is no longer whether AI is reshaping software development. The more pressing issue is how quickly internal assumptions about qualifying R&D activity need to be recalibrated. The statutory test for technological uncertainty has not changed. But the practical baseline against which that uncertainty is assessed is moving, and organisations that fail to adjust their approach risk either under-claiming or, more dangerously, defending claims that no longer withstand scrutiny.For boards, CFOs and technical leadership, the question is no longer whether AI is reshaping software development. The more pressing issue is how quickly internal assumptions about qualifying R&D activity need to be recalibrated. The statutory test for technological uncertainty has not changed. But the practical baseline against which that uncertainty is assessed is moving, and organisations that fail to adjust their approach risk either under-claiming or, more dangerously, defending claims that no longer withstand scrutiny.

Evidence standards will need to evolve. In an AI-assisted development environment, it will no longer be sufficient to characterise integration complexity or performance optimisation as inherently uncertain. Companies will need to demonstrate clearly where established approaches proved inadequate, where AI-assisted solutions failed or required material adaptation, and where genuine technical constraints were encountered and resolved through investigative work. AI does not eliminate uncertainty. It raises expectations regarding what a competent professional can reasonably deduce. The burden of proof shifts accordingly. As the US analysis noted, traditional metrics like lines of code are becoming less indicative of genuine research; the focus must shift to capturing the process of experimentation and the cognitive effort of the developer. The same logic applies under the UK framework.Evidence standards will need to evolve. In an AI-assisted development environment, it will no longer be sufficient to characterise integration complexity or performance optimisation as inherently uncertain. Companies will need to demonstrate clearly where established approaches proved inadequate, where AI-assisted solutions failed or required material adaptation, and where genuine technical constraints were encountered and resolved through investigative work. AI does not eliminate uncertainty. It raises expectations regarding what a competent professional can reasonably deduce. The burden of proof shifts accordingly. As the US analysis noted, traditional metrics like lines of code are becoming less indicative of genuine research; the focus must shift to capturing the process of experimentation and the cognitive effort of the developer. The same logic applies under the UK framework.

Governance around AI usage must also mature. Where AI tools form part of development workflows, organisations should be able to articulate the extent and nature of AI assistance, the role of human technical oversight, and the validation and testing processes that sit around AI-generated outputs. In regulated or security-sensitive sectors, these considerations may be determinative. In commercial environments, they will increasingly influence the defensibility of claims under enquiry.Governance around AI usage must also mature. Where AI tools form part of development workflows, organisations should be able to articulate the extent and nature of AI assistance, the role of human technical oversight, and the validation and testing processes that sit around AI-generated outputs. In regulated or security-sensitive sectors, these considerations may be determinative. In commercial environments, they will increasingly influence the defensibility of claims under enquiry.

There is also a portfolio question. Some categories of software work that previously qualified may no longer meet the threshold where implementation has become readily deducible. Conversely, AI may enable investment in more ambitious initiatives, including advanced optimisation, novel model development, proprietary data science and new computational approaches, that represent genuine advances. The strategic opportunity lies not in defending historical narratives, but in aligning future R&D activity with areas where uncertainty remains substantive.There is also a portfolio question. Some categories of software work that previously qualified may no longer meet the threshold where implementation has become readily deducible. Conversely, AI may enable investment in more ambitious initiatives, including advanced optimisation, novel model development, proprietary data science and new computational approaches, that represent genuine advances. The strategic opportunity lies not in defending historical narratives, but in aligning future R&D activity with areas where uncertainty remains substantive.

And the financial architecture of claims needs reviewing. Leaner engineering teams, the shift from labour to tooling costs and the PAYE cap dynamics described above all alter the financial profile of R&D relief. Smaller and scaling businesses should assess how these changes affect projected claims and adjust their funding assumptions accordingly.And the financial architecture of claims needs reviewing. Leaner engineering teams, the shift from labour to tooling costs and the PAYE cap dynamics described above all alter the financial profile of R&D relief. Smaller and scaling businesses should assess how these changes affect projected claims and adjust their funding assumptions accordingly.

The Threshold Has MovedThe Threshold Has Moved

Technological abstraction has historically expanded, rather than diminished, the scope of innovation. When compilers replaced assembly language programmers, the result was not less research but more. When cloud infrastructure eliminated the need to provision servers, it freed a generation of startups to focus on harder problems. There is reason to believe that AI could have a similar amplifying effect on genuine research, provided the incentive structures keep pace.Technological abstraction has historically expanded, rather than diminished, the scope of innovation. When compilers replaced assembly language programmers, the result was not less research but more. When cloud infrastructure eliminated the need to provision servers, it freed a generation of startups to focus on harder problems. There is reason to believe that AI could have a similar amplifying effect on genuine research, provided the incentive structures keep pace.

But policy frameworks must evolve alongside technological capability. If they lag materially, there is a risk of either over-incentivising routine implementation or under-supporting frontier research. The organisations that understand both the technical shift and the regulatory lens through which their work is assessed will be best positioned.But policy frameworks must evolve alongside technological capability. If they lag materially, there is a risk of either over-incentivising routine implementation or under-supporting frontier research. The organisations that understand both the technical shift and the regulatory lens through which their work is assessed will be best positioned.

AI has altered the tools available to competent professionals undertaking R&D activities. It has not removed the requirement for genuine advances in science or technology. The distinction between the two will define the next phase of approach to software and technology related R&D relief in the UK and globally The question for policymakers, advisors and businesses alike is whether the framework can adapt quickly enough to reward the new kind of innovation that is emerging, or whether we risk a growing disconnect between where genuine research is happening and where the incentives are pointing. That is a conversation worth having now, before the gap widens further.AI has altered the tools available to competent professionals undertaking R&D activities. It has not removed the requirement for genuine advances in science or technology. The distinction between the two will define the next phase of approach to software and technology related R&D relief in the UK and globally The question for policymakers, advisors and businesses alike is whether the framework can adapt quickly enough to reward the new kind of innovation that is emerging, or whether we risk a growing disconnect between where genuine research is happening and where the incentives are pointing. That is a conversation worth having now, before the gap widens further.

Souvik Dey, is a Director of R&D Tax at Unity Advisory, with over 15 years' experience leading C-suite and senior-level R&D Tax engagements across major FTSE-listed, PE-backed, global and high-growth businesses. A former Managing Director at a high-growth advisory firm and Big Four Associate Director for R&D Tax and Disruptive Technology, Souvik brings a rigorous, investigative approach to claim strategy, compliance and HMRC defence. His sector experience spans financial services, technology, media, insurance, telecoms, gaming, engineering, manufacturing and life sciences. He holds an MSc in Information Systems from the University of Sheffield and a Bachelors in Computer Science Engineering, and is a vocal advocate for ethics in R&D Tax practice and responsible use of AI in tax advisory.Souvik Dey, is a Director of R&D Tax at Unity Advisory, with over 15 years' experience leading C-suite and senior-level R&D Tax engagements across major FTSE-listed, PE-backed, global and high-growth businesses. A former Managing Director at a high-growth advisory firm and Big Four Associate Director for R&D Tax and Disruptive Technology, Souvik brings a rigorous, investigative approach to claim strategy, compliance and HMRC defence. His sector experience spans financial services, technology, media, insurance, telecoms, gaming, engineering, manufacturing and life sciences. He holds an MSc in Information Systems from the University of Sheffield and a Bachelors in Computer Science Engineering, and is a vocal advocate for ethics in R&D Tax practice and responsible use of AI in tax advisory.

Unity Advisory is a next-generation CFO advisory firm, grounded in experience and built to evolve. Free from audit conflicts and legacy constraints, the firm supports complex, ambitious mid-market organisations with integrated, AI-enabled solutions across finance operations, tax, transactions, digital and transformation. Unity Advisory helps leadership teams move faster, scale with confidence and deliver sustainable enterprise value.Unity Advisory is a next-generation CFO advisory firm, grounded in experience and built to evolve. Free from audit conflicts and legacy constraints, the firm supports complex, ambitious mid-market organisations with integrated, AI-enabled solutions across finance operations, tax, transactions, digital and transformation. Unity Advisory helps leadership teams move faster, scale with confidence and deliver sustainable enterprise value.