ICYMI
Learning from London: Innovation Needs Rules – AI lets employees build and improve their own workflows. Companies need shared rules and services that allow those inventions to work together.
Executive Summary
AI is making it increasingly easy to solve technical problems such as writing code and analysing data sets. The harder problems are the ‘soft’ (organisational) ones. Slow adoption is often blamed on a skills gap, but some employees have reasons not to adopt AI even if they are skilled.
AI redistributes power. It can devalue the currencies managers have built careers on, such as the size of their teams and budgets, their unique knowledge, the approvals they control, and the expertise they are respected for.
The most valuable AI deployments redesign whole workflows. This can concentrate losses on a few people, who organise to resist change, whilst spreading gains thinly across many, who have less incentive to organise. Even where the company has powerful winners, implementation may depend on the support of the losers. This is the Olson Trap.
Coalitions harden around whatever arrangement exists when the dust settles, so it’s best to resolve these political issues early rather than unpick them later.
The politics of deployment needs the CEO’s backing and can’t easily be delegated.
In 1815, the UK Parliament passed a Corn Law banning imports of foreign wheat until the domestic price reached 80 shillings a quarter. This protected the rents of landowners (who were well represented in Parliament) but raised the price of bread for the public.
In 1838, a group of Manchester manufacturers founded an Anti-Corn Law Association, leading to the national Anti-Corn Law League in 1839. The League was well organised and funded. In 1843 alone, its staff of 800 sent out nine million pamphlets.
In response to the rising public disquiet, exacerbated by the potato blight which caused mass starvation in Ireland, the Conservative Prime Minister, Sir Robert Peel, decided to repeal the Corn Laws. On 15 May 1846, the Commons passed repeal by 327 votes to 229, even though many in Peel’s own party voted against him. Years of campaigning by a well-organised group of manufacturers helped build support and then the famine created an urgent reason to act. Both winners and losers had organised to represent their own interests.
Now, companies face a similar challenge with AI. Much of the debate about deployment focuses on skills and training. But AI deployments can also stall because AI redistributes power, and the people who stand to lose it may be few, influential and organised, while many of the people who gain are dispersed. Using an AI assistant day to day is one thing, but redesigning a workflow, which impacts roles, budgets and decision rights, is something different altogether.
The Soft Stuff Is the Hard Stuff
AI is making many technical problems easier. At Google, the share of new code generated by AI rose from more than a quarter in October 2024 to about three quarters by April 2026, according to chief executive Sundar Pichai. Morgan Stanley reported that its DevGen.AI tool reviewed nine million lines of legacy code in five months, saving an estimated 280,000 developer hours. Legacy systems, historically a big barrier to change, are becoming easier to understand and refactor.
But whilst many technical problems are becoming easier to solve, organisational problems remain hard. Gartner found that the average employee experienced ten planned enterprise changes in 2022 (e.g., an organisational restructure or a new IT system), up from two in 2016. Over the same period, the share of employees willing to support change fell from 74 to 43 per cent. AI adoption is now adding to this change load. In a December 2025 Gartner survey, just over half of 110 CHROs said their organisation had redesigned or redefined roles because of AI in the previous year. A 2026 Henley Business School survey of 2,900 UK workers found that 61 per cent felt keeping up with the pace of AI development could be overwhelming.
One response to weak AI adoption is to increase training. But there are reasons to question whether skills are the only constraint. In IBM’s 2026 survey of 2,000 CEOs, respondents said that only 25 per cent of their workforce uses AI regularly, yet 86 per cent believed their employees had the skills to work with it. If people have the skills and aren’t using them, something else is impeding progress.
Executives have started to describe part of the problem. In a 2026 survey of 2,400 executives and employees by Writer, an enterprise AI vendor, 54 per cent of executives said AI was ‘tearing their company apart’ and 56 per cent reported power struggles over it. Twenty-nine per cent of employees reported behaviour Writer classified as undermining their company’s AI strategy, rising to 44 per cent among Gen Z. Executives are responding by upping the stakes, with 60 per cent saying they plan to lay off employees who can’t or won’t adopt AI.
Middle management is one place where these tensions show up clearly. In an Infosys survey of 2,603 employees, 22 per cent of middle managers said their company was trying to transform most or all of its work and processes with AI, compared with 49 per cent of senior executives and 25 per cent of junior staff. Middle managers were also more than twice as likely as senior leaders to downplay their own AI use. This isn’t all due to change resistance. Infosys also found gaps in access, training, guidance and autonomy. There are nevertheless genuine reasons for middle managers to worry. In 2024, Gartner predicted that, through 2026, 20 per cent of organisations would use AI to flatten their structures, eliminating more than half of their middle-management positions.
AI Creates Losers
We can begin to understand the political struggles at the heart of deployment by looking at who loses and why. Managers and executives in large organisations accumulate power in several ways and AI can devalue many of them at once:
Headcount. Many job-evaluation frameworks give significant weight to team size and budget. A manager who uses AI to do the same work with half the team may find her role downgraded at the next review. All the AI training courses in the world won’t counteract that incentive.
Knowledge. Careers can be built on being the person who knows ‘how things really work around here’. This tacit knowledge includes undocumented processes, client history and workarounds for flawed IT systems. That knowledge is a source of influence with colleagues. As AI is used to build enterprise-wide memories that capture documents, conversations and decisions, this can reduce the broker’s advantage. The people who hold the most tacit knowledge may be the least enthusiastic about codifying it.
The right to say no. Enterprise AI requires data to be collated and integrated from across the company. But each function tends to treat its data as its own. Risk, legal, security and procurement teams hold legitimate vetoes over data access. These controls matter, but they can also be used to resist AI deployment. Instead of ‘no’, opponents can say ‘not yet’ or raise a data-protection concern. In Germany, works councils have co-determination rights over technical systems used to monitor employees’ behaviour or performance, including AI tools.
Expertise. People whose careers rest on a skill threatened by AI have little incentive to invest in it. To paraphrase Upton Sinclair, it is hard for an executive to understand a technology when their career depends on not understanding it. This type of resistance may not even be conscious, and may manifest as scepticism about quality or a belief that ‘our work is different’.
The problem is made worse because companies often ask the experts with most to lose to judge the quality of AI output. Senior lawyers are asked whether AI contract review is good enough, for example. The best qualified to judge are also the most exposed.
Of course, sometimes the losers are right. In a controlled trial by METR using early-2025 tools, 16 experienced developers were 19 per cent slower using AI, even though they believed they were 20 per cent faster. Some of AI’s theoretical gains don’t translate to reality.
AI also gives junior staff the tools to outperform senior colleagues. Where junior staff are more fluent with AI than their managers, they can have a relative advantage as expertise gathered through experience is replaced by expertise sourced or augmented by AI.
Credit and blame. When an AI programme saves money, several functions may claim credit, including the business unit, IT, finance and the transformation office. Disputes over who deserves credit can stall programmes even when they are delivering well. Blame is shared less generously. The researcher Madeleine Clare Elish coined the term ‘moral crumple zone’ for the way responsibility falls on the nearest human when an automated system fails. The manager who approved the AI process may carry the blame for its failures, while any gains are spread across the company.
Business model. Some politics are built into how companies earn money. Consultancies, law firms, agencies and outsourcers often charge by the hour or by the head, so work that takes less time or fewer people earns less. So a shift in how work is done or charged for creates winners and losers.
Losers Get Organised, Winners Hide
So AI creates losers, most of whom are acutely aware of the threat to their status or even livelihood. But the population of losers can be relatively small, so why are they so influential?
In The Logic of Collective Action (1965), the economist Mancur Olson explained why small groups often beat large ones. When a change imposes large costs on a few people and small benefits on many, the few have a strong motivation to get organised and react against the change. In the meantime, each member of the larger group gains too little to justify the effort of getting organised to support the change.
The Corn Laws fit this pattern. A few thousand landowners each stood to lose a great deal if the laws were repealed, while millions of consumers each stood to save a little on bread. The political scientist Cheryl Schonhardt-Bailey has shown that the Anti-Corn Law League succeeded because its supporters were themselves concentrated, in the cotton-exporting towns of the North which needed free trade.
AI deployment follows the same logic. Suppose an insurer deploys AI in its claims process end to end and saves £10 million a year across staffing and other operating costs, resulting in the displacement of 20 middle managers. Each of these managers loses their income, status and career path. Their boss, the claims director, loses half her organisation. Even if all the savings were passed on to customers, each would gain just £10 a year, assuming a million customers. For the impacted managers, the losses are concentrated and severe. For customers, the gains are dispersed and small. The claims director may spend months raising concerns about the new process, but customers are highly unlikely to weigh in to support it. If the company retains some of the savings, shareholders also have a reason to support the change. But the change may be impossible to implement without the support of the managers who are losing out.
We might call this the Olson Trap. AI deployments that create substantial value can stall when losses are concentrated and the winners are dispersed, or when the losers are essential to delivering the change. AI assistants such as ChatGPT get high uptake because using them doesn’t immediately threaten anyone’s job, budget or reporting line. This may explain why AI assistants are everywhere, while redesigned workflows remain rare.
Those who gain from AI deployment have a problem of their own. Earlier automation technologies (e.g., robots, ERP systems) were bought and deployed centrally, giving the company more visibility and control over how gains were used. With generative AI, the individual employee may be the first to see and control the time saved, perhaps finishing a day’s work before lunch. That employee then may choose to tell their manager, likely being rewarded with more work. Or they can keep quiet and spend the time as they like.
Many do keep quiet. Ivanti’s survey of more than 6,000 office workers found that nearly a third of those using generative AI at work keep their AI use secret. Reasons given for doing so include wanting to keep a secret advantage over colleagues and fear that their job may be cut. It also means that some of the apparent non-use reported in surveys is really hidden use.
So some losers organise and some winners hide. That does not mean the time saved is always wasted. In the Infosys survey, nearly 80 per cent said they reinvested it in more work. But employees may know more than their bosses about their individual productivity, and may be reluctant to declare gains until their concerns are addressed. This is a trust problem, not a training problem.
The AI Elite and Other Coalitions
In The Rise and Decline of Nations (1982), Olson extended the argument. He found that stable societies accumulate ‘distributional coalitions’, groups organised to protect their share, and that these coalitions slow down the adoption of new technologies. Established companies work the same way. Functions, committees and colleague peer groups can become coalitions with something to protect. This is one reason why AI-native competitors tend to move faster: they use the same models as everyone else, but have fewer coalitions impeding progress.
AI is already creating new coalitions. Writer found that 92 per cent of executives are cultivating an ‘AI elite’ class of employees. Today’s AI champions may become tomorrow’s landowners as they acquire budgets, status and access of their own to protect.
This argues for early action on AI politics. Coalitions harden around whatever arrangement exists when the dust settles, and any delay gives the losers more time to organise. It is possible to break up long-standing coalitions (Olson’s critics point to Britain post 1979). But it’s far easier to shape coalitions while they are still forming.
Isn’t This Just Change Management?
Companies often practise change management as though the change will be good for everyone once they get used to it. Resistance is then treated as a transition cost, caused by uncertainty, inconvenience or habit. The toolkit is well known, including communication, training, celebrating early wins, etc.
But when a workflow is redesigned around AI, some people end up permanently worse off. The best change management approaches already factor in power dynamics and coalitions. But a toolkit reduced to communication and training won’t resolve a distributional conflict. For those with something real to lose, the right response is negotiation.
Why the CEO Must Drive This
‘The CEO needs to own this’ is an overused phrase, particularly by management consultants. In practice, CEOs are rarely a good choice to lead operational initiatives, given the competing demands on their time. So it’s no surprise that with AI, as with previous technologies, new roles have been created to lead the charge. IBM reported in May 2026 that 76 per cent of the organisations surveyed now have a Chief AI Officer, up from 26 per cent a year earlier. Other firms have given the job to the CHRO, because the barriers are about people, or to the CTO, because the tools are technical.
But if we follow Olson’s logic, none of these roles can resolve the politics without the CEO’s backing. Each may lead a function or coalition of their own, so they are to some extent ‘in the fight’ rather than above it. Every function has an interest in protecting its own headcount, budget and remit. What distinguishes the CEO is the authority to settle conflicts between functions and commit the entire company to the agreed terms. McKinsey’s 2025 survey found that CEO oversight of AI governance was the factor most correlated with EBIT uplift at large companies.
Implementation is a different matter. Moderna merged HR and IT under a single Chief People and Digital Technology Officer, Tracey Franklin, so that the people and technology sides of AI could be merged into one function. That is one sensible home for delivery (other models can work too). But unresolved conflicts over the terms need to be escalated to the CEO.
The CEO may have politics of their own, of course, depending on their context. The costs of redesign arrive during a CEO’s tenure, but the payoff may arrive after it. Announcing AI without redesigning anything is one politically rational response, which may explain why 75 per cent of executives in Writer’s survey said their company’s AI strategy was more for show than real guidance.
The CEO Playbook for AI Politics
Own the difficult decisions. Whilst delivery can be delegated, the terms of the change and the backing for them need to come directly from the corner office. In April 2025, Shopify’s Tobi Lütke set out his expectations for AI use, asking that teams demonstrate why AI can’t do a job before requesting headcount increases. Messages must be backed with clear decisions on resources, incentives and disputes.
Decide what is on the table. The objectives of an AI programme shouldn’t be up for debate. But the implementation design should be tested and informed by consultation. Peel drove repeal through Parliament despite the landowners’ opposition by staggering implementation over three years. Resistance to AI deployment can be diluted through several means. These include giving executives new accountabilities such as AI oversight, changing timing or notice periods or offering to protect a manager’s grade and pay. Each of these bears a cost but should be affordable compared with the value at stake. Other giveaways can include a bigger remit in the redesigned process, first refusal on the new roles created, a share of the savings, or a commitment that headcount falls through attrition rather than redundancy.
Organise the winners. Dispersed winners may have little reason to get organised, but if you proactively recruit them to the cause they can build momentum for change. In a study of call-centre agents, Erik Brynjolfsson and colleagues found that AI raised productivity by 15 per cent on average and by 30 per cent for the less experienced and lower-skilled workers in the published 2025 study. Staff who can demonstrate the benefits of AI are natural supporters, so give them visibility, promotion opportunities and a role in the redesign.
Give coalition leaders a stake in the future. Groups are more likely to stay organised while the people doing the organising find it worth their while. If you can offer those coalition leaders a role in the new arrangement, they may bring the coalition along with them. The claims director offered ownership of the redesigned end-to-end process may become an advocate for it, for example. It’s important that any offers made conform to the same company-wide terms that apply to everyone else. A private bargain that leaves her team exposed would undermine trust.
Publish the terms once, for everyone. If every redesign is negotiated separately, the best-organised group wins. Set out a company-wide policy on what happens to time saved through AI, so employees can share their gains without fear of losing their jobs. Agree in advance how responsibility for AI errors will be handled, so the manager who approves a new process is not the only one exposed. Ultimatums, such as Brian Armstrong’s account of Coinbase dismissing engineers who had not tried AI coding tools, may increase usage, but don’t support openness.
Rewire incentives and decision rights. HR frameworks for grading and career development may need to be updated. Managers need to be graded on output rather than team size or budget. Individuals who contribute knowledge to the corporate memory should be recognised and rewarded accordingly. Engineers or operators managing large teams of agents need to be recognised as a new class of manager. Coalitions also find ways to keep their power through control over budgets, staffing and promotions, so these approval rights need to be redesigned too. At the same time, build AI fluency into hiring and promotion criteria.
Use a crisis, or build somewhere new. A missed target, a competitor’s launch or a market shock can loosen established coalitions. Where there is no crisis, building new units from scratch may avoid replicating existing coalitions. But those units still need a route into the main company’s data, customers and operations if they are to prosper.
All of this costs political capital, as Peel discovered. On 25 June 1846, within hours of the House of Lords passing repeal, protectionist Conservatives voted with the opposition to defeat his Irish Coercion Bill. He was forced to resign.
The landowners lost the vote but escaped much of the immediate damage they had feared. The eventual pressure came from outside. AI’s losers may also preserve their position for years, particularly if internal politics keeps delaying redesign. But in the meantime, competitors aren’t waiting for those disputes to be settled. They can use AI to lower costs or improve service, so that your internal political stalemate becomes a full-blown commercial crisis. The CEO then has to resolve the same conflicts, but with less time and less money to soften the losses.
Monday Morning Actions for Executives
Ask HR what happens to a manager’s grade and pay if she halves her team using AI.
For your AI programmes, map who gains, who loses and who controls implementation, and check who has actually spoken to them.
Find the hidden winners. Ask who is already getting the most out of AI, and what is stopping others from doing so.
Review your AI strategy. If it’s not creating winners and losers, it isn’t going to deliver value.
Price the terms. Estimate what retraining, redeployment, pay protection or other incentives will be needed to secure agreement. Compare these costs with the value at stake and the cost of delay.
Questions for the Board
Which of our AI programmes create losers, and who are they?
Which conflicts remain unresolved, and which need the CEO’s intervention?
Do our HR frameworks reward the behaviour AI requires, or the behaviour it replaces?
What have we promised employees about how gains and losses from AI are distributed?
Do the CEO’s incentives reward redesign, or just announcements?
Sources & Notes
Writer and Ivanti are technology vendors, Infosys is an IT services provider, and IBM’s CEO study is self-reported and published by a supplier. Google and Morgan Stanley figures are company-reported. The insurer example is an illustration, not a benchmark.
Corn Laws and repeal: UK Parliament, Corn Importation Bill, third reading (Hansard, 15 May 1846); Robert Peel, government historical biography. The National Portrait Gallery dates the national League to 1839. The pamphlet figures are attributed to G. M. Trevelyan, British History in the 19th Century (1922), p. 270, in the Anti-Corn Law League historical summary, which also cites W. H. Chaloner on later wheat prices.
Sundar Pichai, Google Cloud Next remarks (April 2026).
The Wall Street Journal, reporting on Morgan Stanley’s DevGen.AI (June 2025).
Gartner research, reported in “Employees Are Losing Patience with Change Initiatives” (Harvard Business Review, May 2023); 2022 data.
Gartner, “Top Change Management Trends for CHROs in the Age of AI” (16 March 2026); December 2025 survey of 110 CHROs, including changes to roles because of AI.
Henley Business School, World of Work Institute, “Hype versus reality: A 12 month AI pulse check” (2026); OnePoll survey of 2,900 full-time UK workers, conducted between 31 March and 13 April 2026.
IBM Institute for Business Value, 2026 CEO Study (4 May 2026); 2,000 CEOs across 33 geographies.
Writer, enterprise AI adoption survey (April 2026); 1,200 C-suite executives and 1,200 non-technical employees who use AI at work; broad definition of sabotage.
Infosys, AI and the Future of Work (2026); 2,603 white-collar employees. The transformation percentages concern what respondents say their company is trying to do, not their own participation or resistance.
Gartner, predictions for 2025 and beyond (22 October 2024); forecast about middle-management flattening through 2026.
Germany, Works Constitution Act (Betriebsverfassungsgesetz), section 87: paragraph 1(6) on technical monitoring systems and paragraph 2 on conciliation.
The line about careers and understanding adapts Upton Sinclair, I, Candidate for Governor: And How I Got Licked (1935).
METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” (July 2025); randomised trial with 16 developers; follow-up and its limitations (24 February 2026).
Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction” (Engaging Science, Technology, and Society, 2019).
Mancur Olson, The Logic of Collective Action (Harvard University Press, 1965).
Cheryl Schonhardt-Bailey, From the Corn Laws to Free Trade: Interests, Ideas, and Institutions in Historical Perspective (MIT Press, 2006).
Ivanti, Technology at Work (May 2025); more than 6,000 office workers. The hidden-use statistic refers to respondents using generative AI at work.
Mancur Olson, The Rise and Decline of Nations (Yale University Press, 1982).
McKinsey & Company, “The State of AI” (March 2025; 1,491 respondents).
Moderna, 2025 proxy statement, recording Tracey Franklin’s merged role from November 2024.
Shopify, Tobi Lütke memo (April 2025).
Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work” (Quarterly Journal of Economics, 2025); the published estimates are 15 per cent overall and 30 per cent for less experienced and lower-skilled workers, superseding the 14/34 figures in the earlier working paper.
Coinbase, Brian Armstrong on dismissals in the Cheeky Pint podcast (August 2025), reported by TechCrunch.