A CEO's Guide to the Four Stages of AI Maturity

Four scenarios for enterprise AI adoption, defined by how much of today's manual knowledge work AI can actually replace, and how the Board, the CEO, middle management, and individual contributors should each behave in each one.

Everybody talks about how AI will reshape work. Almost nobody is willing to put numbers on it. The conversation swings between "this changes everything" and "it's just another technology cycle," and most leadership teams I talk to are quietly stuck somewhere in between. They know they should be doing something. They are not yet sure what. The result is a posture of caution that looks reasonable in the boardroom and costs the company two years of optionality.

Scenario planning is one of the most useful tools you can reach for in a moment like this. The point isn't to predict the future. The point is to commit to a specific shape of the future, walk through what your organisation looks like inside it, and figure out what you should be doing differently right now. You'll be wrong about the details. You won't be wrong about the strategic posture, if you do the work properly.

This post is my attempt at that work. Four scenarios, anchored on a single variable: how much of today's manual knowledge work AI can actually do at acceptable quality. 10%, 30%, 60%, 80%. In each one, I walk through the right move for the Board, the CEO, middle management, and individual contributors. There's an opinion at the end about which scenario worries me most, and why. You don't have to agree with the percentages. You just have to pick one and read your row honestly.

The starting hypotheses

Before any of the scenarios mean anything, the assumptions have to be on the table. I'm working with four. If any one of these is wrong, the whole exercise breaks in interesting ways, and that's fine, that's the point. Argue with them.

1. The Displacement Factor. The percentages in this article describe the share of today's manual knowledge work that AI can perform at acceptable quality. They are not predictions of how many jobs disappear. AI is replacing tasks first, jobs second, and the two are connected but they are not the same number. McKinsey's 2025 State of AI report puts 88% of organisations using AI in at least one function, but only around 1% consider themselves mature. That gap, between "we have AI" and "we have an AI strategy that works," is the gap this post is really about.

2. The Profitability Trap. Most organisations will reach for cost-cutting before they reach for value creation, because cost-out shows up on the next quarterly slide and value creation does not. A headline number like "30% AI adoption" lands very differently on a CEO's calendar than on an entry-level analyst's job description, and the cost lever is always more visible than the strategic one. Anthropic's 2025 labour-market analysis of Claude conversations shows AI usage is heavily concentrated in software engineering, writing, and analytical work, not evenly distributed across the org chart. Average percentages mislead. You have to look by level, and you have to be honest about which lever your leadership team is actually capable of pulling.

3. The Token-to-Labour Ratio. The cost of compute will remain significantly lower than the cost of human salaries for the foreseeable future, and that asymmetry is what makes the rest of the maths work. Federal Reserve research published in April 2026 shows generative AI gives the average user about 5.4% of work-hours back, roughly 2.2 hours per 40-hour week. And yet only 29% of organisations report meaningful ROI from generative AI. The time saving is happening. The financial benefit isn't, automatically. Converting one to the other is a separate skill, and most companies don't have it yet.

4. The Structural Lag. Institutional change runs roughly 12 to 18 months behind technical capability. By the time a company has reorganised around what AI could do last year, the model has moved on, and the org is once again behind. Roughly 55,000 of the 1.17 million layoffs tracked by Challenger, Gray & Christmas in 2025 cited AI as a reason. A January 2026 piece in HBR made the point explicitly: many of those cuts were made ahead of any measured productivity gain, on the bet that AI would catch up. Sometimes it does. Often it doesn't. Headcount reduction and productivity gain are separable decisions, and any honest scenario plan has to treat them that way.

If you accept those four, keep reading. If you don't, you'll quickly see why I drew the lines where I did.

The map, in one table

Before the long version, here is the high-level view. One row per stakeholder, one column per scenario. Read across to see how the same person's job changes as adoption rises. Read down to see what each scenario looks like across the org.

The table is the short version. The next ten minutes are the long version, scenario by scenario. If you only have time for one row, read your own.

Scenario 1: 10% adoption — the Low Hanging Fruit

AI can replace about one in ten units of today's manual knowledge work. This is roughly where measured productivity actually sits in 2026: the Federal Reserve's 5.4% time-savings figure, doubled, to be generous about the next twelve to eighteen months. If you find yourself thinking "we're already past 10%," you are probably confusing tool usage with task replacement. They are not the same thing.

Board. 10% is not enough to justify a reorg. The work that gets freed up is too thin and too unevenly distributed to merge departments or rip up the operating model. If you cut headcount based on 10% adoption, you're trading the most fragile piece of an emerging capability for a one-time cost saving, and you'll regret it inside two years. The right board move is to redirect the savings into things with longer-dated returns: AI literacy across the workforce so the next tier is absorbable, sustainability and customer experience programmes that compound over time, and serious investment in the data foundations that everything else will rely on. This is easier to say than to do. The narrative reward for "we cut 200 jobs through AI" is, in the short term, larger than the narrative reward for "we hired 200 people into customer experience," and most boards underestimate how much that asymmetry shapes their own decisions. Resist it.

CEO. Realistically, 10% adoption translates into something like a 5 to 8% improvement in financial performance, not a 10% one. Not all time saved converts to revenue, and a meaningful chunk of it gets absorbed by people doing more of the same work, slightly faster. The CEO's job at this stage is foundation-building, not P&L mining. McKinsey is unambiguous on this: the gap between organisations that are starting to scale AI and those that aren't is widening fast, and most of the gap is built on data, talent, and governance, not on tools. If you skip the foundation in scenario 1, you don't get to play in scenario 2.

Middle management. Standardisation is the first lever. AI replaces repeatable tasks much faster than bespoke ones, so the manager who codifies how their team actually works, through templates, playbooks, and explicit decision rules, gets a multiplier on every hour saved. The shape of the role starts to change here. You move from approving outputs to designing the work itself. Less time spent reviewing whether a draft passed muster, more time spent designing what the work should look like in the first place, and which parts of it should ever pass through a human at all. Managers who do not make that shift will find their role narrowing without anyone formally announcing it.

IC — the Supervisor. ICs at 10% are spending more time checking AI outputs than producing them, and they should be paid attention to, because they are the early warning system for everything that goes wrong later. Reliability matters here. Microsoft Research's 2026 DELEGATE 52 paper showed that even top frontier models corrupt about 25% of document content over long workflows, with no plateau even after 100 interactions. The IC's real job in this scenario is to be the human safety net against silent errors, and to surface them up the chain. If you treat the IC as a routine reviewer, you lose the only honest signal you have about how good the AI actually is.

Scenario 2: 30% adoption — the Reorganisation Zone

AI can replace roughly a third of today's manual knowledge work. This is the threshold where real reorganisation becomes economically rational, and it's the one I'd treat as the base case for the next two to three years. It's also, by some distance, the most dangerous scenario, and we'll come back to that.

Board. 30% is reorg numbers. A 30% reduction in headcount, applied unevenly across the company, is genuinely feasible. Token costs go up, but they remain marginal next to a labour line item, so the maths works. The board ends up with a binary choice: pocket the savings as margin, or reinvest them into growth, new products, new markets, new audiences. Both are defensible. Neither is automatic. Anthropic's CEO has publicly suggested that AI could eliminate roughly half of entry-level white-collar positions within five years. At 30% adoption, that pressure is already showing up in graduate intake and junior consulting and legal roles. The boards that handle this well are the ones that ask, in writing, "what is the next generation of senior talent going to be made out of, if we cut the entry-level pipeline?" The boards that handle this badly don't ask that question until 2029.

CEO. The cost-of-labour reduction is the easy headline. The harder question is which 30%. Cutting evenly across the org tends to remove the wrong people, because institutional knowledge and connective-tissue work doesn't show up cleanly on a productivity dashboard. The 2025 Workday and Amazon layoffs cited AI explicitly; the ones that actually worked were the ones paired with deliberate redeployment of senior talent into AI-leveraged roles, not symmetrical cuts. The CEO who treats this as a finance exercise loses the company quietly. The CEO who treats it as a redesign keeps it.

Middle management. This is where the pain concentrates. 30% adoption means new team shapes, new charters, and a different incentive structure, all at once. Managers have to broaden their own skill range to cover the wider span of work AI now enables, and they have to become coaches and reviewers rather than directors of work. The role they were originally promoted into, the one they spent years getting good at, often doesn't exist anymore. Some managers will quietly thrive in this transition. Many will not, and a meaningful share of them will leave, voluntarily or otherwise. If you don't plan for that, you are planning for it accidentally.

IC — the Survivor. The IC sees colleagues impacted directly, and workload often goes up initially, not down. AI introduces new categories of work, prompt design, output review, edge-case handling, escalation triage, on top of whatever is left of the original role. The paradox of this scenario is that the people who keep their jobs frequently end up doing more, not less, while watching peers leave around them. Career-wise, the most resilient ICs in this scenario are the ones who shift toward designing and supervising AI systems rather than producing outputs that can be replaced. For an IC reading this, the practical guidance is straightforward: get fluent with the tools, document the parts of the job that are tacit, and position yourself as the person who knows where the AI breaks. That is the role with a future.

Scenario 3: 60% adoption — the Structural Pivot

AI can replace roughly 60% of today's manual knowledge work. This requires meaningful improvements over current model capability, but at the rate of progress observed in 2025 and 2026, it's closer than the public conversation makes it sound. Gartner already predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from fewer than 5% in 2025. The infrastructure for this scenario is being laid down right now, even if the capability isn't there yet.

Board. At 60%, the entire operating model is in question. Departments can be merged or eliminated. Whole layers of management may not be needed. Token costs become a real line item, but they remain favourable next to labour, especially as inference prices keep dropping. The board's question shifts from "how much can we save?" to "what is this business actually for, in a world where most knowledge work is done by machines?" That isn't a cost question. It's a strategy question, and most boards are not staffed for it. The boards that thrive at 60% are the ones that brought in operators who have lived through a previous platform shift, retail to e-commerce, on-prem to cloud, manual to digital, and who know what it feels like when the moat moves.

CEO. This is the inflection point where AI stops being a productivity story and becomes a strategy story. The companies that win at 60% adoption are the ones that redefine their value proposition, not the ones that cut hardest. You can technically reduce your labour cost base by half. So can every competitor. The harder, more durable play is using the freed capacity to do things that weren't economic before: bespoke customer experiences, deeply local products, services that previously required too much manual effort to be profitable. The CEOs who pull this off behave less like cost-cutters and more like product founders. If your instinct at 60% adoption is to lean harder on the CFO, you're managing the wrong scenario.

Middle management. The role of middle management as it currently exists shrinks dramatically. What remains looks much closer to product management and orchestration: defining what the AI does, where human attention belongs, and how the team's outputs combine into something coherent that a customer would actually pay for. Hierarchies flatten. Titles stop meaning what they used to. The manager who survives is the one who can hold a portfolio in their head, work across functions without ceremony, and explain the work in plain language to people on either side of them. The manager who survives by being the most senior person in the room does not survive.

IC — the Specialist. An IC at 60% adoption is, in practice, a senior decision-maker, even early in their career. They edit AI-produced work, judge it for fitness, and feed feedback into the systems that produced it. The volume of work they oversee is far higher than what an IC was responsible for in 2025. What separates a great IC from an average one in this world is much less about output volume and much more about empathy, taste, and physical presence with customers, the things AI is worst at. The career ladder distorts: there is more responsibility much earlier, and there are also fewer footholds, particularly in consulting and law where the entry-level rungs were always the apprenticeship. The World Economic Forum's 2025 Future of Jobs report estimates 92 million displaced jobs against 170 million new ones globally by 2030. Whether that arithmetic holds inside any one organisation depends on choices being made right now, in scenario 1 and scenario 2, before the 60% world arrives.

Scenario 4: 80% adoption — the Autonomous Enterprise

AI can replace 80% of today's manual knowledge work. This is the scenario most current models fall short of, and the one most often invoked in public commentary, usually with more confidence than the evidence supports. It's still worth planning for, because the asymmetry is huge: if you're wrong about 80% being far away, the cost is enormous, and the runway to react is short.

Board. 80% isn't "more cost reduction." It's a different company. The capital structure changes. Payroll halves, or worse. Riding the cost curve down is not a strategy, because every competitor is doing the same thing, and the floor gets reached fast. The advantage in this scenario goes to organisations that find new outputs entirely: services priced to a market that didn't exist when they were too labour-intensive to deliver, products that depend on near-zero marginal cost of expert work, or geographies that were uneconomic before. Cost-out plays at 80% adoption are race-to-the-bottom plays. The board's job is to refuse them, even when they look like the safe option.

CEO. The CEO's job becomes capital allocation and culture. Where do you point the freed capacity? What kind of organisation do you want to be when most of the work is being done by systems? What does ambition look like, when "scaling the team" stops being the lever it was? Very few CEOs have run companies where the dominant input is compute rather than headcount, and the leadership models for it are being written in real time. For most CEOs, the role they will occupy at 80% adoption looks materially different from the one they were appointed into, and the leaders who acknowledge that early tend to manage the transition far better than those who do not.

Middle management. Most middle management roles do not survive in their current form. What replaces them is a smaller number of orchestration roles, where people design how AI systems and human reviewers fit together, and a larger number of customer-facing roles where irreducibly human judgement still matters: relationships, taste, ethics, accountability, the things you can't outsource without losing the brand. The career direction for a middle manager looking at this scenario is straightforward: move toward the customer, or move toward the system. The middle does not survive.

IC — the Architect. An IC at 80% adoption is, in practice, a small business of one, capable of producing work that previously required a team. They prompt, audit, and orchestrate whole workstreams that used to belong to entire departments. The role becomes more entrepreneurial, less hierarchical, much more dependent on judgement and taste. The risk concentrates at the entry level: there may be no clear path from "graduate" to "senior" if the work that historically taught the craft is no longer being done by humans. That's not a future the labour market has handled well in any previous platform shift, and there's no reason to expect this one to be different unless we deliberately design it to be. Which brings us to the part most posts skip.

The scenario that worries me most

The scenario that worries me most is not the 80% one. It's the 30%.

That's the one where the temptation to reorganise is highest, the institutional knowledge most easily lost, and the actual capability of the technology most easily overestimated. It's the scenario where layoffs feel rational on a slide and turn out to be expensive in reality, and where the people who get cut are the ones holding the connective tissue that didn't show up on the productivity dashboard. The HBR piece I cited earlier, "Companies Are Laying Off Workers Because of AI's Potential, Not Its Performance," makes this point sharply: many of the cuts being made today are running ahead of measured productivity improvement, on the assumption that the technology will catch up. Sometimes that assumption pays off. Often it doesn't, and by the time you find out, the people you needed are gone.

If I were hedging risks today, this is what I'd do. I'd prepare for the 30% scenario as the base case, not the 60% or 80% one, because 30% is the most likely real-world outcome inside the next two to three years, and it's the one with the most asymmetric downside if you mishandle it. I'd invest in AI literacy at every level, not just at the top, because the bottleneck on AI value capture is almost always lower-down in the org than people expect. I'd over-invest in the people who hold institutional knowledge, the ones who know why a process exists rather than just how it runs, because they become more valuable, not less, when the rest of the organisation is being reshaped. And I'd treat layoffs driven by AI hopes, rather than by measured AI productivity, as a credit signal about the company's leadership, not a productivity signal about the technology. Those are not the same thing, and the market is starting to notice.

How to use this

Pick the scenario you think is most likely for your organisation in the next 24 months. Be honest about it: not the one that flatters the leadership team, the one that actually fits your data, your customers, and your workforce. Read your row, by level. Ask whether your current plan would actually work in that world, not in the optimistic version of it. If it would, you're in good shape. If it wouldn't, you've just identified your next conversation. Have it before someone else does it for you.

Sources: McKinsey, The State of AI in early 2025 · Federal Reserve, Monitoring AI Adoption in the U.S. Economy (April 2026) · Deloitte, State of AI in the Enterprise 2026 · Gartner, 40% of enterprise apps will feature AI agents by 2026 · Anthropic, Labor Market Impacts of AI · CNBC, AI is already taking white-collar jobs · HBR, Companies Are Laying Off Workers Because of AI's Potential, Not Its Performance · Microsoft Research, Laban, Schnabel & Neville (2026), LLMs Corrupt Your Documents When You Delegate · World Economic Forum, Future of Jobs Report 2025 · BCG, AI Maturity Matrix.

All articles on this site are written by me. I use AI to assist with final formatting and editing before publication.