The Engine Light Is On. Does Anyone Still Know What Is Under the Bonnet?

There is a particular kind of afternoon that begins with the phrase, “This should save me some time.”

You open an AI tool before lunch to help finish a proposal, organise a plan or build a small automation. By 4.42 p.m., there are seventeen tabs open, three nearly convincing versions, two figures that disagree with one another and a cup of tea that has developed the small grey sadness of something long forgotten. Every part of the task looks almost finished. The task itself is nowhere near done.

AI did not steal the afternoon. More disconcertingly, it persuaded everybody that the afternoon had been saved.

I use AI, and I help businesses use it. I have seen it remove genuinely dreary work, organise an untidy mass of information, reveal a useful pattern and give a small business access to capabilities it could not previously afford. Some of what is happening is remarkable. The proper response to a remarkable tool, however, is not reverence. It is judgement.

The question that interests me is no longer simply, “What can AI do?” It is what happens to the business around it: to the people who must check its work, to the knowledge they cease to practise, to the customer who needs an exception, and to the revenue that was supposed to justify the whole cheerful experiment.

The magic is real. So is the muddle.

One of the clearest studies of AI-assisted knowledge work involved 758 consultants and a set of realistic business tasks. On work that sat within AI’s capabilities, people using GPT-4 completed 12.2 per cent more tasks, worked 25.1 per cent faster and produced higher-quality results. That is not a marginal gain. It is the sort of improvement that quite reasonably makes boards, founders and slightly breathless software demonstrations sit up straighter.¹

The same experiment contained a more awkward finding. When the consultants were given a complex task deliberately chosen to sit beyond the model’s capability, those using AI were 19 per cent less likely to reach the correct answer.¹

The researchers called this uneven boundary the jagged technological frontier. It is an excellent phrase because the edge is not where we expect it to be. AI may be brilliant at one task and unreliable at another that looks remarkably similar. It does not cough politely before crossing from competence into improvisation. The prose remains smooth. The confidence remains intact. Only the answer has wandered off.

This is why arguments about whether AI “works” are becoming less useful. A kettle works. A hammer works. AI behaves more like an extraordinarily well-read new colleague who can produce a board paper before breakfast, occasionally invents the supporting evidence and has not yet learnt to look worried when doing so.

Used well, it can increase our reach. Used carelessly, it can increase the volume of things that look finished.

Those are not the same achievement.

We made the beginning cheaper

Andy Lambert recently described the problem neatly: AI did not remove the work; it pushed it downstream.² That is precisely what many founders are now experiencing, although the work does not always arrive downstream wearing a label.

A proposal can be drafted in twelve minutes. Someone must still decide whether it understands the client. A campaign can produce fifty variations before the kettle boils. Someone must still know which promise the business can keep. Code can appear at astonishing speed. Someone must still test what it does when a customer behaves like a customer rather than a tidy example in a demonstration.

We have made production cheaper. We have not automatically made outcomes cheaper.

This distinction matters because businesses do not earn revenue from the number of things they begin. They earn it when something reaches the other side: a customer chooses, a problem is solved, a product works, an invoice is paid, a client returns.

The latest large UK government study makes that gap visible. Published in February 2026, the Department for Science, Innovation and Technology’s research covered 3,500 private-sector businesses with at least five employees. One in six was using at least one AI technology. Among adopters, 75 per cent reported improved workforce productivity, yet 77 per cent reported no change in revenue. Only 12 per cent reported an increase.³

These are self-reported figures, and the absence of immediate revenue growth does not prove failure. Benefits can take time to travel through a business. Cost reductions matter. Better service matters. Learning matters. Nevertheless, the gap should make us curious.

If time has been saved, where did it go?

If productivity has risen, where should the improvement appear next?

If nobody can connect it to capacity, margin, conversion, retention or revenue, perhaps we have measured the sensation of moving quickly rather than the value of arriving.

A firm congratulating itself on the number of AI-generated outputs may be rather like congratulating the kettle for boiling while forgetting to make the tea.

The saved minute has a habit of coming back

There is now an ugly but useful word for AI-generated material that appears polished while leaving the real thinking to its recipient: workslop.

The research that popularised the term came from BetterUp Labs and Stanford’s Social Media Lab. In an online survey of 1,150 full-time American desk workers, 40 per cent said they had received workslop in the previous month. Each incident took nearly two hours to resolve on average.⁴ This was US, self-reported and vendor-linked research, so we should not turn it into a universal law. We should, however, recognise the scene.

A junior colleague is asked for a market assessment. They use AI and return a beautifully formatted document in an hour. The argument sounds plausible; the numbers are wearing sensible shoes. It reaches a manager who asks where the figures came from. Nobody quite knows. The manager spends the evening locating sources, recovering the original question and deciding which sentences mean anything.

Three hours were saved at one end. Five were quietly charged to somebody more senior at the other.

Speed was privatised. The checking cost was socialised.

For a solopreneur, there is no downstream department. The work simply changes hats. At 10 a.m., you are the enthusiastic adopter. At 6 p.m., you are quality assurance, technical support and the tired-looking person trying to remember why the automation has emailed the same lead four times.

Research with 319 knowledge workers, covering 936 examples of real AI-assisted tasks, found that higher confidence in generative AI was associated with less critical-thinking effort. Greater confidence in one’s own ability was associated with more. The researchers also found that critical work had not vanished; it had shifted towards verification, integration and stewardship.⁵

That shift is economically important. The first draft may now be cheap, but the person capable of spotting what is missing is often experienced, busy and expensive. The saved minute at the front of the workflow returns as a dearer minute later, on the desk of the person whose judgement is scarcest.

Judgement has not been removed from work. It has migrated.

The dashboard is not the engine

We have watched a version of this happen with cars.

It is tempting to blame electric vehicles, but the change began earlier, as cars became increasingly electronic and software-dependent. Lift the bonnet of a modern vehicle and there is less invitation to poke around thoughtfully with a spanner. A warning light may tell you something is wrong, but diagnosis often requires one machine to have a private conversation with another while the owner drinks indifferent coffee in reception.

The car may be safer, cleaner, more efficient and altogether cleverer. The driver may also understand less about how it works.

Our businesses are acquiring the same curious quality.

A founder can see a lovely dashboard of leads, sentiment, conversion, churn, campaign performance and predicted cash flow. AI may have helped to generate the content, classify the enquiries, route the work, summarise the calls and interpret the numbers. Everything glows. Everything has a percentage beside it.

Then revenue falls.

Which assumption changed? Which customer segment stopped responding? Was the campaign wrong, the data untidy, the routing flawed, the offer weaker or the automated follow-up simply rather strange? Can anybody trace the decision back through the system, or does everyone stand around the dashboard waiting for a more qualified garage?

A warning light is not a diagnosis. A dashboard is not understanding.

This is the deeper risk I see for founders. A business can automate its work and accidentally outsource its understanding of the work.

What a business knows

Much of a company’s intelligence has never lived in a manual.

It lives in the account manager who knows that a particular client writes “fine” when things are decidedly not fine. It lives in the operations person who remembers that the monthly report only reconciles if one awkward field is refreshed first. It lives in the founder who can hear, in the extra second before a customer answers, that the price is not the real objection.

Some of this knowledge should be documented. Some of it can be improved by AI. Yet knowledge is not merely information stored somewhere. It is context, practice, pattern recognition and responsibility gathered over time.

I think of a business’s ability to retain that understanding as operational sovereignty: knowing how the company creates value, where judgement enters, what happens when the normal case fails and who can restore the system when it does.

This is not an argument for founders to approve every email personally until the end of time. That way lies a different sort of breakdown. It is an argument for keeping human understanding close to the places where money, trust and exceptions move through the firm.

The English High Court offered a bracing example in 2025. Considering cases involving actual or suspected use of unchecked AI-generated legal material, the court noted that publicly available large language models can produce coherent, plausible answers that are entirely wrong, including invented cases and quotations. It made equally clear that the professional duty to check remains with the lawyer, whether that person used AI directly or relied on somebody else’s AI-assisted work.⁶

The drafting could move. The accountability could not.

That principle travels rather well beyond the law. If an AI-assisted forecast affects hiring, the founder owns the decision. If generated copy makes an untrue promise, the company owns the promise. If an automated customer journey traps a frightened person in a conversational roundabout, the brand owns the frustration.

Delegation does not dissolve responsibility. It merely changes where a responsible person must stand.

The skills we only miss when we need them

There is another concern, and it deserves careful language because the evidence is still developing.

People become skilled partly by doing the ordinary work: noticing patterns, making small judgements, being wrong and correcting course. When automation removes routine practice, it can also remove the repetitions through which expertise is built and maintained. Researchers have warned that AI assistance may conceal this decay because performance can remain high while the tool is present.⁷

A 2025 natural experiment in Poland provides one sober example. After AI was introduced into colonoscopy practice, clinicians’ unassisted adenoma-detection rate fell from 28.4 per cent to 22.4 per cent over the following three months.⁸ One medical setting cannot carry the whole argument for business, and it should not be asked to. It does reveal the shape of the risk: assisted performance can improve while independent capability quietly weakens.

For founders, this matters when early-career people no longer get enough practice at the work they will later be expected to judge. If a junior marketer never learns to research a market without a generated summary, how will they recognise a shallow one? If a new developer mostly approves suggestions, where is the slow accumulation of instinct that notices an elegant error? If nobody writes the first draft, who develops an ear for the difference between a sentence that is correct and one that is alive?

We may be removing the apprenticeship and keeping the final examination.

Eventually, an exception arrives. The system is unavailable, the case is novel, the customer is furious or the model is confidently wrong. At that moment, the business needs the very judgement it has stopped rehearsing.

Humans are not the expensive bit left over

Some companies have already discovered that replacing human capacity is a more complicated calculation than reducing a headcount line.

Klarna became famous for saying its AI assistant was doing work equivalent to 700 customer-service agents. The tool was not useless; it continued to handle about two-thirds of enquiries and delivered substantial speed improvements. Yet the company began recruiting people again. Its chief executive acknowledged that cost had been too dominant a factor and that the result was lower quality.⁹

That is not an AI failure story. It is a judgement correction.

Commonwealth Bank of Australia made a smaller but equally instructive reversal. It announced 45 customer-service redundancies after introducing an AI voice bot, then apologised and reversed the decision after the union reported rising call volumes; the bank said it had not considered all the relevant business factors.¹⁰

The automation had changed part of the work. It had not removed the surrounding demand, the exceptions or the humans wanting help.

Companies that replace people too quickly may save a visible salary and lose something spread across less obedient columns: customer trust, product feedback, commercial instinct, resilience, training capacity and the person who knows what to do when Tuesday refuses to resemble the process map.

Humans are not the expensive residue left behind when the clever work has been automated. In many businesses, they are the part that knows what the work is for.

Revenue is where the story must eventually arrive

For founders, time is not an abstract efficiency measure. It is inventory.

Five hours a week spent correcting, prompting, reconnecting and supervising an AI workflow is 260 hours a year. If that time is conservatively valued at £100 an hour, the opportunity cost is £26,000 before counting a single missed sale, delayed launch or weary decision.

The more important cost is often harder to see. It is the proposal not followed up because the afternoon went into repairing the proposal-generating system. It is the customer conversation replaced by an hour polishing content nobody was waiting for. It is the service issue that reaches a human only after the customer has ceased to feel warmly towards the company.

A useful AI investment should eventually leave footprints in the commercial world.

If AI improves…	The gain should become visible in…
Customer work	Faster resolution, lower abandonment, stronger satisfaction, retention or repeat purchase.
Sales and marketing	Better-qualified demand, improved conversion, shorter sales cycles or a clearer cost of acquisition.
Delivery and operations	Shorter end-to-end cycle time, fewer errors, less rework, more capacity or healthier margin.
Founder capacity	More time spent on customers, decisions, relationships, product and revenue-producing work.

The unit of progress is not the prompt. It is not the generated page, the automated step or the number of licences distributed. The unit of progress is what the business can now do better, more reliably or more profitably than it could before.

This is where management matters. Office for National Statistics research found that firms with stronger management practices were more likely to adopt advanced technologies and to follow through on planned AI adoption.¹¹ Technology does not float above the business that buys it. It lands inside existing habits, incentives, definitions, data and decisions. A confused process with AI inside it is still a confused process, only now it can reproduce itself at speed.

Keep a hand on the knowledge

Human-led AI is not a timid compromise in which a person ceremonially approves everything the machine has done. It is an operating choice.

It means deciding what success looks like before generating the work. It means measuring the whole journey from request to accepted outcome, including checking and repair. It means keeping sources and decision trails where the stakes justify them. It means designing a clear route to a human before reducing the humans available. It means giving less-experienced people enough unaided practice to build the judgement the company will later depend upon.

Most of all, it means staying close to the living parts of the business. Founders still need to hear customers speak, understand how delivery really happens and know why revenue moves. The aim is not to clutch every task possessively. It is to avoid becoming a passenger in your own machinery.

One question reveals a great deal:

If the AI disappeared for a week, could your business still explain what is happening, decide what matters and recover?

The goal is not to prove that you can run without it. Few of us would enjoy returning to a world without search engines, spreadsheets or satnav. The answer shows whether AI is carrying labour or carrying understanding.

Modern drivers do not need to rebuild an engine at the roadside, and founders should not spend their lives tightening every operational bolt. Someone inside the business must still know what the warning light means, which journey matters and when it is wise to pull over.

AI can help us travel further and faster. I am grateful for that. We should simply take care not to become so enchanted by the speed that nobody remembers how to steer.

The destination was never more output. It was a business that works, earns, serves and can still find its way home.