Before AI replaces the job, ask what knowledge you still need

The fastest way to make an AI investment look financially attractive is to put a number beside the jobs it might remove. It is also one of the fastest ways to design the wrong organisation.

Before leaders treat a role as redundant, they need to separate the work AI can absorb from the knowledge the organisation still needs. The tasks may disappear first. The expertise often does not.

The easiest savings case can be the wrong design

OpenAI's account of building an AI-native finance function is instructive partly because of what it is not. It is not a synthetic organisation in which artificial agents constitute the finance department. The human finance team remains the function. OpenAI says it is working towards a zero-day close and continuously updated forecasting. AI is being built around the team to reconcile information, prepare initial explanations, support forecasting and surface exceptions, while finance professionals validate the numbers, apply judgement and retain final sign-off.¹

OpenAI is selling AI as well as describing its own operating model, so the account should be treated as a primary signal rather than independent proof that every finance function can reproduce the same results. Even with that caveat, the design choice is revealing.

That arrangement can be dismissed as an ordinary human-in-the-loop model. The more interesting question is why OpenAI has retained the humans so centrally.

The post describes a small finance team supporting a company growing at exceptional speed. This was not presented as a large back office waiting to be cut. The economic opportunity lay in carrying scarce expertise further: replacing searches, spreadsheet assembly and recurring preparation with live tools that allow the same specialists to oversee more work and support decisions sooner.¹

In that setting, reducing headcount is a narrow ambition. The greater prize is leverage. A finance professional who no longer spends hours reconstructing a variance can spend more time deciding whether it matters, testing the explanation, challenging assumptions and advising leaders while the outcome can still change.

The distinction matters beyond finance. AI can remove work without removing the need for the person who understood the work.

A job is not a list of tasks

Most automation cases begin by listing activities: gather the data, reconcile the records, draft the report, answer the routine query, prepare the forecast. This is useful, but incomplete.

A role usually contains several different kinds of value.

There is execution knowledge: how the task is normally completed. There is exception knowledge: how to recognise when the normal process no longer applies. There is decision knowledge: which differences matter and which can safely be ignored. There is also the why layer: the accumulated rationale behind a control, assumption, threshold or apparently awkward handover.

The first layer is often the easiest to automate. The later layers are frequently less visible because experienced people apply them without narrating every step. They notice that a number is technically valid but commercially implausible. They know that one client promise should alter the forecast while another should not. They recognise when a clean answer is suspiciously clean.

This is tacit knowledge. Some of it can be documented. Some can be demonstrated through examples. Some may be captured indirectly in the data used to train or configure an AI system. But capturing past behaviour is not the same as preserving the ability to respond when the environment changes.

That is why the task-versus-job distinction, long established in automation research, remains so important. Jobs are bundles of tasks, and technology may substitute for some tasks while complementing the people who perform others.⁴⁵ Exposure to AI does not, by itself, tell a leader whether the whole role should disappear. The ILO's 2025 task-level index estimates that one in four jobs worldwide has some exposure to generative AI, but describes transformation rather than automatic replacement as the more likely general effect.⁴

The expert may be feeding the system that appears to replace them

A large workplace study of a generative AI assistant in customer support makes the knowledge problem unusually visible.

Across 5,172 agents, access to AI assistance increased productivity by 15 per cent on average. The largest gains went to less experienced and lower-skilled workers. The researchers found evidence that the system was spreading the communication patterns and effective practices of stronger agents, helping newer workers move down the experience curve more quickly.²

At first glance, that can look like a straightforward substitution story. Once the organisation has captured the best workers' methods, perhaps it needs fewer experts.

The paper points in the opposite direction as well. Top performers supplied many of the examples on which the system learned. Their contribution was not limited to resolving today's cases; they were also producing the training material from which tomorrow's assistance was built. The researchers note that some high-skill workers were already being asked to review AI suggestions and provide better alternatives. They also raise a longer-term risk: if the most capable workers contribute fewer original solutions because they begin following the system, future versions may become less effective at handling new problems.²

This is the knowledge-retention paradox.

The better AI becomes at reproducing expert behaviour, the easier it is to mistake the source of that behaviour for a redundant cost. Yet the system may still depend on experts to identify new exceptions, challenge stale patterns, improve recommendations and decide when yesterday's best practice no longer fits today's circumstances.

An organisation that removes those people too quickly may preserve a copy of its existing knowledge while losing its capacity to create the next version.

Buy-in is part of system performance

Keeping knowledge holders involved also changes the politics of adoption.

OpenAI describes giving finance staff room to identify recurring problems, build tools and test them with colleagues, supported by technical specialists. The use cases came from people close to the work rather than being imposed entirely from above.¹ That approach gives employees a reason to reveal where the real friction sits, which shortcuts are dangerous and which apparently simple tasks conceal judgement.

It also gives them a credible place in the future function.

That is not merely a humane way to introduce technology, although it may be that too. It is an operating advantage. People are more likely to help redesign a process when they are being asked to extend their expertise than when they suspect that documenting the process is the final task before their own removal.

OECD workplace research has found that training and worker consultation are associated with better reported outcomes from AI adoption.³ The evidence does not prove that consultation alone causes a successful implementation. It does reinforce a practical management point: trust, participation and capability-building affect whether people engage with the system, surface its weaknesses and learn how to use it well.

Buy-in is therefore not decorative change-management language. It determines whether the organisation gains access to the knowledge it is trying to encode.

Leaders should be candid about this. Retaining experts solely to extract what they know before dismissing them is unlikely to create durable trust. The future role must be real, with authority, responsibility and a visible contribution to how the AI-enabled function operates.

Move the knowledge holder into the control layer

When AI absorbs more execution work, the experienced employee's role should not shrink into rubber-stamping machine output.

A useful control-layer role is more demanding than that. The knowledge holder may need to:

• define what a good output looks like and which evidence it must carry;
• set thresholds for approval, escalation and exception handling;
• challenge outputs that are plausible but contextually wrong;
• preserve the why layer behind controls and decisions;
• update the system when policy, markets, customers or operating conditions change;
• contribute new examples, edge cases and rejected alternatives;
• remain accountable for consequential judgement.

This is a move from producing every unit of work to managing a larger field of AI-executed work. It resembles the wider shift from prompting to managing: the scarce human contribution is increasingly the ability to specify, test, steer, challenge and improve the system rather than personally perform every step.

That does not make the human role smaller. In a lean, expert-heavy function it may make the role more consequential. One finance professional can oversee more reconciliations, scenarios or explanations, but the organisation becomes more dependent on that person's ability to recognise when the automation should not be trusted.

The design question is therefore not simply how much work the AI can complete. It is where knowledge, authority and learning must sit after the work has been redistributed.

This is not a promise that everyone stays

None of this supports a comforting claim that AI will never remove jobs.

Some functions are bloated. Some roles consist largely of stable, codified and testable tasks. Some organisations will decide that fewer people can safely run the redesigned process. Productivity gains may reduce the number of worker-hours required, and the customer-support study itself notes that the firm could theoretically handle the same volume with fewer hours if demand did not increase.²

Nor is every experienced employee automatically an irreplaceable knowledge holder. Tenure can preserve poor practice as easily as good judgement. Expertise should be evidenced through the person's ability to diagnose, explain, improve and handle exceptions, not assumed from title alone.

The contrarian point is narrower and more useful: headcount reduction should not be the first design decision.

The right sequence is to map the work, identify the knowledge on which good performance depends, redesign the roles, transfer what can be transferred, test the control system and only then decide what capacity is genuinely surplus.

Starting with a target number of jobs to remove reverses that sequence. It invites the organisation to delete the people before it understands what else it is deleting with them.

Flattening can remove the ladder as well as the layer

That danger is not new. From the late 1980s into the 1990s, British organisations joined a broad delayering movement: fewer management levels, wider spans of control and larger jobs for those who remained. A 1995 Institute for Employment Studies review found that 74 per cent of 150 companies surveyed had widened middle managers' spans of control; another survey of personnel directors at 123 top-500 UK organisations found that 64 per cent had undertaken a delayering exercise.⁶

Some firms removed genuine bureaucracy. But the review also warned that an organisation could become so flat that managers were overloaded and had to compromise on the work they performed. It noted another risk: co-ordination roles could be cut away without the need for co-ordination disappearing.⁶

A friend of mine saw a version of this inside a large British furniture retailer. He ran the kitchen-design department in one store, a responsible and reasonably well-paid role. Department managers reported to store management. Area managers sat above the stores.

The company removed the store-management layer and retained the area managers. Almost immediately, the area managers became long-distance store managers across several locations. The reporting lines looked flatter. The practical job did not. Local staffing, trading, customer and operational decisions still had to be made, but the people now expected to carry them were too distant and too broadly stretched.

The less visible damage was to the management ladder. A department manager could plausibly progress to assistant store manager, then store manager, and eventually area manager. Each step expanded the person's judgement and experience. The direct leap from running one department to overseeing several complete stores was too large.

As area managers left, the organisation had fewer plausible internal replacements. It had not merely removed a layer of management. It had removed part of the process by which future area managers were made.

One person's recollection cannot establish the wider effect of the restructuring. The mechanism, however, is well recognised. Research on four large British service-sector organisations found that earlier delayering had opened gaps in internal job ladders, weakened incremental on-the-job development and turned progression into a much larger, sometimes all-or-nothing leap.⁷

This gives leaders a third question to place beside tasks and knowledge: what future capability is created by doing this role?

A role may be valuable for what it produces today, what its holder knows, and what it prepares that person to become next. AI may reduce the first while leaving the other two intact.

AI can make the old mistake easier to repeat because it reproduces visible outputs so quickly. Reports still appear. Questions still receive answers. Dashboards still update. The missing local judgement, escalation route, coaching and succession path may remain invisible until the system encounters something unfamiliar or the next senior vacancy appears.

Removing a role may also remove the route by which the organisation learns to fill the role above it.

The Knowledge-Retention Test

Before automating a role or reducing the team around it, leaders should be able to answer eight questions.

Decision rule

Automate the work. Preserve the intelligence

OpenAI's finance example is not evidence that every organisation should keep every specialist. It is a useful demonstration of a different objective.

When the team is already small and the expertise is scarce, AI is not primarily a headcount tool. It is a way to increase the reach of people who understand the function. The employees retain a place partly because that improves participation and trust, but more fundamentally because their knowledge remains necessary to validate, govern and improve the work.

The same logic applies wherever apparently routine output sits on top of years of accumulated judgement.

Before AI replaces the job, leaders should ask what knowledge will leave with the person, who will know when the system is wrong, how the organisation will continue learning, and where the next generation of experts will gain the experience to take over.

The job may eventually be redesigned out of existence. The knowledge must be deliberately designed back into the organisation.