Written submission from Gianina Caballero (AIB0072)

Written evidence submission - Business and Trade Committee Inquiry: Artificial Intelligence, business and the future of the workforce

Gianina Caballero, MSc — Independent advisor and researcher, expert in the analysis and delivery of complex systems, with no institutional affiliations, financial interests, or party-political associations relevant to its contents.

INTRODUCTION

This submission presents an independent structural analysis of AI adoption, workforce impact, and government strategy in the UK economy, produced using the EIMAR Framework, a proprietary analytical framework designed to identify systemic patterns in complex systems. It does not summarise or evaluate the reports it draws on. It extracts quantitative and qualitative data from government-commissioned research and applies an independent interpretive lens to identify structural patterns that the reports themselves do not articulate.

The evidence base draws on data published by the UK government between 2025 and 2026, including research commissioned from or produced by DSIT, the AI Security Institute, Skills England, Ipsos, and the Alan Turing Institute[i]. In each case, what this submission uses is the data, not the authors' analysis, interpretation, or conclusions.

SUMMARY

At the surface level, the available data presents the following picture. Of businesses surveyed, 80% do not use AI and have no plans to do so. Among the 16% that do, the median spend is £2,000 and 31% report spending nothing. 75% of adopting businesses perceive productivity gains; 77% report no change in revenues. The proportion of all businesses surveyed demonstrating measurable economic impact attributable to AI is approximately 2%.

Taken together, these figures describe a system in which declared adoption, perceived productivity, and institutional commitment to AI deployment are all present, and in which demonstrable economic transformation is not. The gap between them is the product of specific mechanisms and structural conditions that the data itself contains, visible once the figures are read not in isolation, but as a system. The analysis that follows does not introduce new data, it reads the government's own data structurally, crossing what adoption figures show against what productivity figures reveal, and what skills data exposes against what policy documents declare. In each case, the gap between what the system produces and what it claims to be building is measurable, traceable, and addressable, but only if policy is designed against what the evidence shows rather than what the strategy declares.

THEME 1: AI ADOPTION

Where is AI currently being deployed across the UK economy (by sector, region, firm size) and how is it being used?

AI deployment across the UK economy is low, uneven, and concentrated in ways that limit its productive potential. Of 3,500 businesses surveyed by DSIT, 80% do not use AI and have no plans to do so. The 16% that declare use are not a representative cross-section of the economy: adoption rises with firm size, from 14% among micro businesses to 36% among large ones, and varies sharply by sector. Construction, transport and hospitality cluster around 88–90% non-use; information and communication reach 43%; financial services, 21%. These are not marginal differences: they reflect structural variation in access to capital, talent and technical capacity that the pattern of adoption reproduces rather than corrects.

Within the adopting minority, the pattern of use narrows further. 85% rely on external natural language processing or text generation tools, while agentic AI, the form most associated with process transformation, appears in only 7% of cases. Functionally, use is concentrated in marketing and administration, each at 72%, while R&D and technical engineering represent 1%. What is being described as AI adoption is, in the majority of cases, the integration of accessible external tools into existing support functions. This is operationally rational given the constraints firms face, but it is categorically different from the structural transformation that productivity gains at economy level would require.

The investment data reinforces this reading. The median spends on AI among businesses surveyed is £2,000, and 31% report spending nothing. Among businesses declaring intention to adopt in the coming year, only 28% have a defined budget and 47% have none. Intention and investment are not aligned, and where adoption is not resourced, it defaults to the lowest-cost accessible option, which is precisely what the usage pattern shows.

How does adoption vary across different parts of the economy and why; what are the main barriers to AI adoption?

The barriers that businesses identify when adoption does not occur are not primarily technical. Among active barriers, ethical concerns are cited by 80% of firms, high cost by 76%, and unclear regulation by 72%, lack of tools or lack of interest do not feature with comparable weight. These are not barriers that capability-building programmes address, they are structural conditions that operate independently of whether businesses have the skills to adopt.

The talent market that should enable deeper adoption operates under its own constraint. 77% of organisations recruit from existing AI roles, the dominant strategy is not to develop capability but to capture what already exists. This pool circulates within a narrow geographic and organisational perimeter: outside London and the South, only 5% of roles are filled via graduate routes; in London and the Southeast, that figure reaches 53%, and 46% of all AI vacancies are concentrated in London alone. The organisations best positioned to adopt AI more deeply are also the organisations with access to this talent. The majority of businesses, particularly SMEs, which represent 82% of the labour market survey sample, operate outside this perimeter without access to the talent pipelines that deeper adoption would require. The consequence is predictable: adoption defaults to external tools because building internal capability requires talent the system does not generate or distribute beyond its existing boundaries.

THEME 2: IMPACTS ON WORK AND WORKERS

What evidence is there that AI is having an impact on productivity?

The productivity case for AI in UK businesses rests on evidence that the available data both supports and precisely qualifies. 75% of adopting businesses report improved workforce productivity, but that figure describes perceived efficiency gains, not measured economic outcomes. At the same time, 77% of the same group report no change in revenues, only 12% of adopting businesses report increased income, and only 16% of that group attribute it substantially to AI. Across the full survey universe, the proportion of businesses demonstrating measurable economic impact attributable to AI is approximately 2%. Productivity is perceived, but it does not convert into economic outcomes at scale.

The mechanism that produces this gap is specific and traceable. Reducing time spent on administrative tasks generates efficiency but not value, unless that time converts into additional productive activity, and that conversion is not measured in any of the data systems currently available. Businesses declare productivity improvement because something has changed in how they work; revenues do not move because the change has not altered what they produce or sell. This is not a failure of AI as a technology: it is a predictable consequence of deploying AI in support functions, where 72% of use is concentrated, without redesigning the workflows that determine whether efficiency translates into output. The gains are real within the adopting minority, but structurally bounded by the type of adoption occurring, a pattern consistent with findings from experimental studies that productivity gains at task level have not translated to macroeconomic level.

What impact can AI have on productivity at the economy, sector, organisation and worker levels? What conditions are required to diffuse AI uptake, create productivity gains and what factors might slow AI use?

The distribution of effects across sectors, regions and groups follows the same concentration pattern visible in adoption itself. Sectors with highest adoption, information and communication, financial services, are also those where productivity gains are most likely to be observed; sectors with lowest adoption, construction, transport, hospitality, are not yet in a position where AI impact, positive or negative, is measurable at scale. The geography of impact mirrors the geography of adoption: concentrated in London and the Southeast, with the rest of the country at an earlier stage of exposure. This is not a temporary distributional lag that will resolve as adoption spreads, the talent market data shows the conditions for broader adoption are not being built: the pipeline remains narrow, geographically concentrated, and oriented toward organisations that are already inside the system.

What are the effects of AI on employment and job security?

For workers specifically, the available evidence has a structural limitation the Committee should register. The dominant employer narrative projects continuity: AI as augmentation, not substitution, with no significant anticipated reduction in headcount. That narrative coexists with a finding that warrants attention: employers consistently report that AI will not displace workers in their own sector, while simultaneously acknowledging that AI will replace many jobs across the economy. Both positions are held at the same time, by the same respondents, without resolution. What the data does not capture is workforce contraction that occurs without formal redundancy, not replacing staff who leave, reducing new hiring, or reorganising roles. That contraction is real but invisible to the evidence base, and invisible evidence does not enter policy.

Which sectors, regions and groups are most likely to be affected by AI, and what are the implications for inequality?

The groups most structurally exposed to the effects of AI, whether through adoption pressure, skills displacement, or labour market reorganisation are those already operating at the margins of the current talent system: workers in sectors with low adoption and limited upskilling infrastructure, workers in regions outside the London-Southeast corridor, workers in SMEs without access to formal training pipelines, and workers in roles concentrated in the tasks that current AI tools perform most readily. The concentration of benefit and the distribution of risk do not map onto the same population, that asymmetry is structural, it predates AI adoption and is being amplified, not corrected, by the current pattern of diffusion.

The data on gender representation within the AI talent sector illustrates this dynamic directly. More organisations now employ at least one woman in an AI role, the proportion with no female representation fell from 53% to 41%, yet the overall share of women across the sector fell from 24% to 20% over the same period, broader distribution has not produced greater representation. Over the same period, ethics training nearly doubled, from 24% to 45% of organisations. Training is expanding and more organisations have some female representation but the overall share of women in the sector is falling, and more activity is not producing more representation.

THEME 3: SKILLS, EDUCATION AND TRANSITIONS

What skills will workers and managers need to work effectively with AI?

The skills that workers and managers need to work effectively with AI fall into two categories that the current policy response treats as equivalent, but are not. Skills for use; operating tools, running workflows, generating outputs are the focus of the majority of current training investment; skills for understanding; evaluating outputs critically, identifying errors and bias, judging reliability, governing deployment responsibly are more universally necessary, more difficult to develop at scale, and more systematically underfunded. The AI Skills for Life and Work collection, published by DSIT in January 2026, documents this asymmetry in the data: only 17% of the working population can explain what AI is in detail, only 21% feel confident using it at work, and 84% have received no AI-related training of any kind. The gap is not primarily in technical capability; it is in the foundational comprehension that makes technical capability usable and governable.

For managers specifically, the evidence identifies a failure point that training alone cannot resolve. Only 34% of senior leaders can identify new AI use cases relevant to their business, and 50% of employers do not know what AI training is relevant for their organisation. These are not knowledge gaps that a skills programme fills: they are governance gaps that reflect the absence of the contextual understanding needed to make adoption decisions, evaluate vendor claims, or assess whether deployment is producing the intended results. An organisation whose leadership cannot identify relevant use cases will not generate demand for the deeper skills its workforce would need to use AI productively. The skills deficit and the adoption deficit are the same problem at different levels of the organisation.

Are current education, further/higher education and adult learning systems adequate and flexible enough to meet evolving skills needs?

The current education and training system is not adequate to meet these needs, and the inadequacy is structural rather than a matter of scale or funding alone. A Rapid Evidence Review commissioned by DSIT identifies that nearly 60% of the workforce cannot complete the 20 tasks of the Essential Digital Skills framework, the baseline on which AI literacy must be built. The policy response has concentrated almost exclusively on tertiary education; master's degrees, doctoral training centres, bootcamps, while AI content in primary and secondary curricula remains minimal. The market concentrates investment in skills of use because they produce direct returns, while skills of understanding generate positive externalities that benefit other organisations when trained workers move; and so, the market systematically underinvests in exactly the skills most universally needed.

The deepest structural constraint is the privatisation of advanced knowledge: in the most technically advanced organisations, 65% of AI training is internal compared to 32% external because cutting-edge model development is increasingly proprietary, and that knowledge does not circulate to the education system or the general labour market, leaving universities in a permanent reactive position, teaching to a baseline that industry has already moved beyond. The consequence is a closed circuit: the most valuable formation requires prior entry into the system, and the system becomes progressively less permeable. SMEs, the majority of UK businesses, cannot access this formation at any price, and so operate with tools they do not fully understand, without the governance frameworks that advanced organisations are developing internally, and without realistic pathways to build those capabilities organically.

How can employers support upskilling when introducing AI?

The pattern that emerges from employer behaviour when introducing AI is consistent across the evidence: 87% do not integrate AI into their graduate training programmes and 60% provided no training of any kind in the previous twelve months. When employers introduce AI without developing internal governance or comprehension-oriented training, the result is what the data shows throughout this theme, adoption declared without the capacity to evaluate it, human oversight without procedures to structure it, and advanced knowledge concentrated in the organisations that already have it. The gap between training delivered and capability developed is not a funding problem, it is a design problem, and it operates at every level of the system simultaneously.

THEME 4: GOVERNMENT STRATEGY, REGULATION AND RIGHTS

How prepared is the Government and are regulators for the potential impacts of AI?

The government's preparedness for the impacts of AI on employment and business should be assessed against what its own evidence shows, not against what its strategy documents declare. The AI Opportunities Action Plan was accepted in full on the day of its publication in January 2025, before the research describing the actual state of AI adoption in UK businesses had been commissioned or completed. The One Year On update reports 38 of 50 commitments met. The dashboard at delivery.ai.gov.uk shows measurable progress across all seven areas. What neither document contains is any metric connecting those commitments to economic outcomes. Commitments met is a measure of process, but it is not a measure of whether the process is producing the transformation the strategy requires. The Assessment of AI Capabilities, published by the AI Security Institute in January 2026, identifies as priority evidence gaps establishing causality between AI adoption and employment change, predicting which groups absorb transition costs, and measuring net job creation. These are not peripheral concerns; any strategy genuinely aimed at managing AI's impact on work and workers would need to answer exactly those questions first and the government's own scientific apparatus acknowledges it cannot yet answer them.

The businesses operating within this strategy have identified what they need from government: regulatory clarity and standards, public funding for training, sector-specific capability support, evidence of proven use cases, and clearer information about the real economic value of AI adoption. These demands are not unreasonable, but the available data allows a more precise reading of each one, and that reading has direct implications for what government support should actually look like.

Is the existing UK regulatory framework sufficient to address current and emerging employment-related AI issues? If not, what changes are required?

The demand for regulatory clarity is the most directly supported by evidence and the most precisely actionable. Regulation is the only area of the AIOAP below 67% completion, sitting at 50%, while Compute and Sovereign AI are at 100%. These are not equivalent types of commitment: infrastructure can be built and counted; normative clarity requires definition, consultation and enforcement. The government has delivered where delivery is easiest to measure, and accumulated its largest deficit precisely where businesses report the most significant active barrier.

That deficit is not incidental, the government presents regulatory flexibility as a competitive advantage over more prescriptive jurisdictions, but the firm-level data does not support that framing: 72% of businesses identify unclear regulation as an active barrier to adoption, and 80% cite ethical concerns. Businesses do not experience flexibility as freedom to innovate; they experience it as the absence of the normative clarity they need to make deployment decisions with confidence. 84% of adopting businesses declare human oversight of AI outputs, but that oversight is, in the majority of cases, informal and individual, without established procedures. Without a defined standard, there is nothing against which that informality can be measured or required to improve.

That pattern is not a failure of individual businesses; it is the predictable result of a system that promoted adoption without defining what responsible deployment requires. That sequence creates a specific obligation: not to regulate more, but to define the minimum standard from which businesses can exercise the governance responsibility that is theirs. Without that standard, governance exists in the abstract but cannot be measured, required, or improved.

What support should the Government be offering to businesses, employers and employees?

The demand for public funding for training is plausible in direction but imprecise in content. The evidence does not show a shortage of training provision; it shows a systematic mismatch between what is being trained and what is most universally needed. Only 17% of the working population can explain what AI is in detail, and only 21% feel confident using it at work. Among those responsible for deployment decisions, the gap is equally significant: only 34% of senior leaders can identify relevant AI use cases for their business, and 50% of employers do not know what AI training is relevant for their organisation. These are not gaps that more training of the same type will close, they are gaps in the capacity to understand, evaluate and govern AI, not in the capacity to use it.

The market systematically underinvests in understanding-oriented skills because their returns are diffuse, a trained worker may move to another employer, and employer investment in training has fallen 60% since 1997. The case for public funding is therefore structural, not supplementary, but conditioned on two things: that it develops comprehension rather than use, and that its distribution mechanism reaches SMEs and regions outside the London and Southeast corridor that current provision does not serve. The government has delivered 1 million free AI courses nationally, whether any of those courses are developing the capacity to evaluate AI outputs critically is not measured anywhere in the current evidence base. That is the distinction this paragraph turns on: between training delivered and capability developed.

The demand for sector-specific capability support and proven use cases appears straightforward, but the data reveals why responding to it directly would miss the underlying problem. Businesses are not adopting AI primarily through independent evaluation of value, they are adopting it under perceived competitive pressure. The belief that "most businesses are already using AI", when only 16% are, is operating as an active decision driver. Providing more proven use cases responds to that pressure without changing the underlying logic.

What businesses need is not a larger catalogue of cases but the internal capacity to evaluate whether a given use case produces value in their specific context. The policy response that addresses this is not more case studies; it is investment in the leadership comprehension that makes case evaluation possible. That distinction, between providing cases and developing the capacity to evaluate them, is not reflected in any current policy intervention, and it is precisely what the data supports.

For workers, the gap is of a different order, not a gap in what is being measured incorrectly, but a gap in who is being measured at all. The Mais Lecture of March 2026 declares equipping workers as the fourth and most important pillar of the government's AI strategy, but no measurement system currently captures the worker experience of adoption, whether it is changing the nature of their work, reducing their autonomy, affecting job security through mechanisms that do not register as formal redundancies, or producing the wage gains that augmentation narratives imply. The AI Economics Institute, announced in March 2026, acknowledges this absence implicitly, but its mandate, budget and publication timeline have not been specified. When that mandate is defined, the minimum it must include is longitudinal worker-level data collection as a non-negotiable condition, not an optional supplement to employer surveys. A government that declares workers its highest priority while measuring exclusively from the employer perspective will systematically design interventions for the actor it can see, and remain blind to the impacts experienced by the actor it cannot.

The demand for clearer information about the real economic value of AI is, of all the demands businesses make, the one the government is most structurally positioned to meet and the one it has so far failed to. The government publishes adoption rates; it does not publish economic outcomes of adoption. The result is that 75% of adopting businesses perceive productivity improvements while 77% report no change in revenues and nothing in the current evidence base captures what happens between those two figures.

Concretely, this requires three things: defining what counts as meaningful adoption, distinguishing transformative deployment from occasional use of a free external tool; establishing revenue and productivity baselines that allow economic impact to be tracked at firm level over time; and connecting the talent measurement system to the adoption measurement system so that the chain between capability, deployment and economic outcome becomes visible. The government's declared ambition to achieve the fastest AI adoption rate in the G7 cannot be evaluated against any standard that would distinguish activity from transformation, because that standard does not yet exist, and the government has not committed to creating it.

The case for reorienting that standard is not theoretical. A 2026 reference model for AI policy in Latin America and the Caribbean, published by the Inter-American Development Bank, documents that countries with significantly less digital infrastructure than the UK face the same structural gap between declared adoption and demonstrable economic impact. If infrastructure does not predict conversion, then measuring adoption rates without measuring economic outcomes cannot demonstrate the transformation the government declares as its ambition. The question is not how fast the UK adopts AI; it is whether what is being adopted is producing value and that requires measuring what matters.

ABOUT THE AUTHOR

Gianina Caballero is an independent advisor and researcher, expert in the analysis and delivery of complex systems. She holds a Master's degree from the University of Oxford in Major Programme Management and submits this evidence in a personal capacity.

She has contributed written evidence to UK parliamentary processes on multiple occasions, including to the All-Party Parliamentary Group on Project Delivery and to the Committee of Public Accounts, and has been cited in their official reports.

She submits this evidence having worked within two government contexts, one of them the UK government, giving her first-hand understanding of how public systems operate from the inside. Further information is available at www.gianinacaballero.com