How Do You Automate Work Fairly? New Research from the Institute for the Future of Work (IFOW)
The short answer: Automating work fairly means treating every AI deployment as a set of choices, not an inevitability. New research from the Institute for the Future of Work (IFOW) on 5,000+ UK workers shows good automation flows to higher-grade roles while monitoring concentrates lower down. Workers who perceive stronger rights tend to fare better, and union voice helps preserve that protection. Fairness has to be designed in, negotiated and monitored.
Most AI-at-work conversations start with the technology. What can it do? How fast can we roll it out?
The more useful question is who carries the risk when it lands.
IFOW has just published research that answers that question with data. The answer is uncomfortable. Right now, patterns of automation in the UK are reinforcing existing inequalities rather than levelling them (Soffia et al., 2026).
When I launched FrAImeworks, ThinkWorkPlay's series on AI at work, I said I'd keep coming back to the question most leaders skip: how do you drive adoption without amplifying the trust problems you already have? This report is one of the clearest pieces of evidence I've read on it.
My position is simple. Any attempt to automate work, AI included, has to be designed, developed and deployed in a responsible, human-centric way. Fairness, equity and ethics belong at the start of that process. Not bolted on after the complaints arrive.
What did IFOW's Governing the Innovation Ecosystem report study?
Governing the Innovation Ecosystem: How Perceived Rights and Worker Voice Shape the Distribution of Risk and Benefit was published by IFOW in September 2026. Its authors are Magdalena Soffia, Abigail Gilbert, Joe Atkinson, Semeli Hadjiloizou and Joana Geisler.
The study draws on a weighted survey of over 5,000 UK working adults. It asks three questions. How are different types of automation distributed across demographic groups, occupations and industries? Do workers who perceive stronger rights experience better automation outcomes? And do different voice mechanisms, such as unions and HR practices, strengthen or weaken that relationship?
It measures six "automation archetypes" developed from the Pissarides Review into the Future of Work and Wellbeing. Each archetype is scored as an index from 0 to 100.
Three caveats matter. The data was collected in 2023, so it mostly predates the generative AI wave. The indicators are proxies, because the survey was not originally designed to measure these archetypes. And the findings are associations, not proof of cause and effect. The authors are explicit about all three.
If anything, the timing makes the findings more pressing. These patterns were already visible before generative AI arrived at scale.
What are the six automation archetypes?
The archetypes describe what technology does to the terms, conditions and quality of work, not which tool is involved. Each one reflects a different strategy firms use to capture value.
Archetype | What it looks like | Who the report found most exposed | Link with perceived rights |
High-discretion augmentation ("good" automation) | Technology enhances autonomy, learning and judgement | Managers, directors and professionals; also Black, Black British, Caribbean or African workers | +1.87 points per 1-point rise in perceived rights |
Low-discretion augmentation | Technology codifies and routinises tasks, reducing skill use | Low end of the same scale; elementary occupations report the lowest augmentation scores | Higher perceived rights associated with movement away from it |
Intensification | More tasks, faster pace, more unsociable hours | Fairly even across occupations; Asian workers report higher levels | No significant association |
Telepresence | Remote or flexible work bundled with monitoring and surveillance | Younger workers, sales and customer service roles, Black, Black British, Caribbean or African workers | -0.86 points per 1-point rise |
Matching | Algorithms allocate, monitor and evaluate work | Younger workers (steep age gradient), sales and customer service roles, Black, Black British, Caribbean or African workers | -0.64 points per 1-point rise |
Displacement | Perceived risk of job loss and growing automated workload | Fairly even across occupations; Asian workers report higher levels; lower in public administration | No significant association |
Source: (Soffia et al., 2026), Tables 1 to 4 and Figures 1 to 5. Perceived-rights coefficients are adjusted for gender, age, ethnicity, occupation, sector and technology exposure.
Who gets "good" automation, and who gets monitored?
Occupation is the clearest dividing line. Higher-level roles are more likely to experience high-discretion augmentation, where technology expands autonomy and decision-making. Lower-level roles are more exposed to monitoring through algorithmic management (Soffia et al., 2026).
Ethnicity is the most persistent gap. Workers identifying as Black, Black British, Caribbean or African report higher telepresence and matching than White workers, meaning greater exposure to monitoring and algorithmic control. Workers from Asian backgrounds report higher intensification and displacement.
There is a twist. The same Black, Black British, Caribbean or African group also reports higher high-discretion augmentation. Autonomy and surveillance can rise together in the same jobs. The authors flag smaller sample sizes for some groups, so these results need careful reading.
Age matters too. Younger workers are more exposed to most archetypes, and the gradient is steepest for telepresence and matching. I looked at the telepresence finding in more detail in why hybrid working comes with more AI monitoring.
Can differences in technology use explain these gaps?
Only partly. IFOW re-ran the analysis accounting for each worker's exposure to digital information and communication technologies (ICTs), wearables, AI and machine learning software, and robotics.
Gender differences in telepresence and matching largely disappeared. The age gradient weakened. But ethnic and occupational differences persisted, albeit smaller (Soffia et al., 2026).
The authors conclude that how technologies are designed, deployed and governed is at least as important as which technologies are used. The same tool can produce very different jobs depending on the choices made around it.
Why do perceived rights shape automation outcomes?
IFOW measured perceived rights with a single question: on a scale of 1 to 10, how many rights do you have in your current job?
After adjusting for demographics, occupation, sector and technology exposure, a one-point increase in perceived rights was associated with (Soffia et al., 2026):
a 1.87-point increase in high-discretion augmentation
a 0.86-point reduction in telepresence
a 0.64-point reduction in matching
There was no significant association with intensification or displacement in that model.
Perceived rights are not evenly spread. Older, higher-paid and managerial workers perceive the most. Workers in elementary occupations, process and machine operation, and caring and leisure roles perceive the least.
Here is the part leaders should notice. Among employees, pay, training, access to representation, HR philosophy and other job characteristics explain around 40% of the variation in perceived rights (R² = 0.428). That is far more than demographics or occupation alone.
Perceived rights are not a soft, personal attitude. They are shaped by things employers control.
Is all worker voice equally protective?
No. This is the report's most important finding for HR leaders.
Access to trade unions is associated with higher perceived rights. And union-only voice does not weaken the protective link between perceived rights and better automation outcomes (Soffia et al., 2026).
Non-union-only voice behaves differently. For workers whose only channel is something like an employee forum or consultative committee, very high perceived rights were associated with more intensification, telepresence, matching and displacement, not less. At lower levels of perceived rights (below 7.75 out of 10), non-union voice did not protect against low-discretion augmentation relative to having no representation at all.
The authors describe this as a risk of "false security" without real bargaining power. They are careful to say the data cannot show why. One possibility they raise is that these channels are more conducive to adopting technologies designed to capture value, whether through greater discretion or greater monitoring.
My read: a forum without bargaining power can become a listening exercise rather than a safeguard. If you have one, ask honestly which it is.
Can an employee-centred HR philosophy backfire?
It can cut both ways.
An employee-centred HR philosophy is strongly associated with high-discretion augmentation, and it strengthens the positive link between perceived rights and that "good" automation (Soffia et al., 2026).
But among employees who perceive high rights, employee-centred HR practices were also associated with higher intensification, telepresence, matching and displacement. The authors suggest that, without union support, employees may fail to negotiate better outcomes or may concede higher-risk ones.
Engagement is not the same as influence. Warm HR culture does not automatically translate into a say over how technology changes someone's workload, autonomy or monitoring.
Is the same pattern showing up beyond the UK?
Yes. Several recent studies point in the same direction.
Algorithmic management is already mainstream. A survey of over 6,000 firms by the Organisation for Economic Co-operation and Development (OECD) found that 90% of US firms had adopted at least one tool to instruct, monitor or evaluate workers, against an average of 79% across the European countries surveyed (Milanez, Lemmens and Ruggiu, 2025).
AI can intensify work rather than lighten it. An eight-month study of a US technology company with around 200 employees found that generative AI led people to work faster, take on a broader scope of tasks and extend work into more hours of the day (Ranganathan and Ye, 2026). It is early, single-company research, but it echoes what I wrote in my time-based pay series: when efficiency is rewarded with more work, AI gains get absorbed rather than shared.
Unchecked matching shifts value away from workers. Oxford researchers analysed 1.5 million trips from 258 UK Uber drivers. After dynamic pricing was introduced in 2023, inflation-adjusted hourly earnings fell from over £22 to just over £19 before costs, while Uber's commission rose from around 25% to 29% (Binns et al., 2025).
Exposure is already gendered. In high-income countries, 9.6% of women's employment sits in the highest generative AI exposure category, compared with 3.5% of men's (Gmyrek et al., 2025). Equity has to be considered at the design stage, not after the fact.
What does the 2026 regulatory picture mean for employers?
Regulation is moving, but slowly, and it will not do the work for you.
In the European Union (EU), AI systems used in employment are classed as high-risk under the AI Act. In June 2026 the Council gave final approval to amendments that delay the high-risk rules for stand-alone systems until 2 December 2027 (Council of the EU, 2026). A delay is not a repeal.
The EU Platform Work Directive must be transposed into national law by 2 December 2026. It requires transparency about automated monitoring and decision-making systems and human oversight of significant decisions, and it restricts processing of certain personal data, such as workers' emotional state or data used to predict union activity (European Parliament and Council, 2024).
The UK has no AI-specific employment law. In its July 2026 evidence to Parliament, the Trades Union Congress (TUC) argues that AI adoption is broad but shallow, largely because workers are routinely excluded from decisions about how it is introduced (TUC, 2026).
IFOW's own recommendations go further. They include a statutory right to bargain collectively over the introduction and use of AI, union access to workplaces where AI and monitoring are deployed, a defined role for HR in procurement, implementation and review, and consideration of Good Work Algorithmic Impact Assessments (Soffia et al., 2026).
For UK employers, waiting for the law is itself a decision. It carries its own risk. I set out why governance keeps lagging adoption in why AI adoption is outpacing AI governance at work.
What should CHROs and COOs do to automate work responsibly?
Responsible automation is a lifecycle, not a launch. Here is how I would translate the evidence into practice.
1. Design: name the archetype before you buy. For every AI or automation proposal, ask what it will do to the job. Does it add discretion or remove it? Does it support people or monitor them? Who will feel it most?
2. Development and procurement: put HR in the room early. IFOW recommends HR has a defined role alongside technology and operations teams in procurement, implementation and ongoing review, assessed against Good Work frameworks. Include an equality impact check by occupation, age, ethnicity and gender.
3. Deployment: give people real influence, not just a hearing. Involve workers in decisions about autonomy, learning, workload and monitoring. Where you recognise unions, bargain over AI introduction rather than simply informing. Explain what rights and routes to challenge people have, because perceived rights shape outcomes.
4. Monitoring: track who carries the cost. After go-live, measure workload, hours, surveillance and autonomy by group. Do not mistake intensity for productivity.
5. Ethics: keep humans accountable. Consequential decisions about pay, performance, scheduling or dismissal need meaningful human review and a clear way to contest them.
Frequently asked questions
What are automation archetypes?
Automation archetypes are a framework developed from the Pissarides Review into the Future of Work and Wellbeing. They describe six ways technology changes work: high- or low-discretion augmentation, intensification, telepresence, matching and displacement. Each reflects a different way firms capture value from technology.
Does AI monitoring affect some workers more than others?
Yes. IFOW's 2026 research found that lower-grade roles, younger workers and workers identifying as Black, Black British, Caribbean or African report greater exposure to monitoring and algorithmic management. Ethnic and occupational gaps persisted even after accounting for technology use.
Do trade unions make AI adoption fairer?
IFOW found that access to unions is associated with stronger perceived rights, and union voice did not weaken the link between perceived rights and better automation outcomes. The research shows association, not cause, but it supports a role for collective bargaining in shaping AI at work.
Is an employee forum enough to govern AI at work?
Not on its own, according to IFOW's findings. Where non-union voice was the only channel, very high perceived rights were associated with more monitoring, intensification and displacement. The authors warn of "false security" without real bargaining power.
When do EU AI Act rules for AI used in employment apply?
Following amendments given final Council approval in June 2026, the AI Act's high-risk obligations for stand-alone systems, including those used in employment, apply from 2 December 2027.
Where to go next
The first FrAImeworks piece: How is AI use at work evolving from individual workflow to shared, agentic systems?
Why AI efficiency gains get hidden rather than shared: AI is Breaking Time-Based Pay. Here's What the Evidence Says Comes Next.
Where measuring work by time came from: Time Was Never the Right Way to Measure Work. Here's Where That Idea Came From.
Planning an AI rollout? Get in touch at madi@thinkworkplay.me to pressure-test it for fairness before go-live.
Sources
Soffia, M., Gilbert, A., Atkinson, J., Hadjiloizou, S. and Geisler, J. (2026). Governing the Innovation Ecosystem: How Perceived Rights and Worker Voice Shape the Distribution of Risk and Benefit. London: Institute for the Future of Work. Weighted survey of 5,000+ UK working adults, data collected 2023. https://doi.org/10.5281/zenodo.19349919
Milanez, A., Lemmens, A. and Ruggiu, C. (2025). Algorithmic management in the workplace: New evidence from an OECD employer survey. OECD Artificial Intelligence Papers, No. 31. OECD Publishing, Paris. Survey of 6,000+ firms in France, Germany, Italy, Japan, Spain and the United States. https://doi.org/10.1787/287c13c4-en
Ranganathan, A. and Ye, X. M. (2026). AI Doesn't Reduce Work—It Intensifies It. Harvard Business Review, 9 February 2026. Eight-month, in-progress study at a US technology company of around 200 employees. https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it
Binns, R., Stein, J., Datta, S., Van Kleek, M. and Shadbolt, N. (2025). Not Even Nice Work If You Can Get It; A Longitudinal Study of Uber's Algorithmic Pay and Pricing. ACM Conference on Fairness, Accountability, and Transparency (FAccT 2025). 1.5 million trips from 258 UK drivers, 2016 to 2024. https://arxiv.org/abs/2506.15278
Gmyrek, P. et al. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140. International Labour Organization and NASK. https://webapps.ilo.org/static/english/intserv/working-papers/wp140/index.html
Council of the European Union (2026). Artificial intelligence timeline: Council gives final green light to simplify and streamline AI rules, 29 June 2026. https://www.consilium.europa.eu/en/policies/artificial-intelligence/timeline-artificial-intelligence/
European Parliament and Council of the European Union (2024). Directive (EU) 2024/2831 on improving working conditions in platform work. Official Journal of the European Union. https://eur-lex.europa.eu/eli/dir/2024/2831/oj
Trades Union Congress (2026). Artificial intelligence, business and the future of the workforce: TUC response to the Business and Trade Committee call for evidence, 22 July 2026. https://www.tuc.org.uk/research-analysis/reports/artificial-intelligence-business-and-future-workforce
Comments