Everyone Has an AI Copilot. Does Everyone Need One?

The productivity pitch behind generative AI has always been unusually simple.

Give a worker an AI copilot and the worker gets more done.

Emails take minutes instead of half an hour. Reports appear from a prompt. Code is generated before a developer has finished describing the problem. Presentations, summaries, research notes and first drafts can all be produced at a speed that would have looked absurd only a few years ago.

And unlike many technology promises, there is now serious evidence that some of these gains are real.

In a controlled experiment involving 453 college-educated professionals, researchers from MIT gave participants occupation-specific writing assignments such as reports, emails and analysis. Workers using ChatGPT completed the tasks around 40% faster, while independent evaluators rated their output 18% higher in quality.

A much larger study followed 5,172 customer-support agents after a generative AI assistant was introduced into their work. Productivity, measured as customer issues resolved per hour, increased by around 15% on average. The gains were particularly strong among less experienced and lower-skilled workers, while the strongest employees benefited much less.

Consulting produced a similar result. In an experiment involving 758 Boston Consulting Group consultants, people using GPT-4 on tasks that fell within the model's capabilities completed 12.2% more tasks, worked 25.1% faster and produced work rated more than 40% higher in quality than participants working without AI.

So it would now be difficult to argue that AI copilots are merely productivity theatre.

They can make people faster.

But that is only the first condition the larger claim needs to satisfy.

Because productivity is not the same thing as producing something quickly.

A consultant does not create value by generating a slide. The slide has to contain the right analysis.

A developer does not create value by generating code. The code has to work inside a real software system.

An employee does not become more productive because an email took thirty seconds to draft if someone else later spends twenty minutes working out what it was supposed to say.

The relevant unit of productivity is not therefore how quickly AI produces an output.

It is how quickly useful work reaches completion.

That distinction changes the picture considerably.

The same BCG study that produced some of the strongest evidence for AI-assisted productivity also found something less convenient. When consultants were given a task that sat outside GPT-4's capabilities, those using AI were more likely to arrive at the wrong answer. The researchers described this uneven landscape as a jagged technological frontier: AI can be remarkably capable at one task and unexpectedly unreliable at another that looks superficially similar.

That makes the question of where to use an AI copilot almost as important as whether to use one.

And the boundary is not always obvious to the person doing the work.

Feeling faster is not the same as being faster.

Software development provided one of the clearest examples.

In 2025, research organisation METR conducted a randomised trial with experienced open-source developers working on their own mature codebases. The developers expected AI tools to reduce their completion time by 24%. Even after completing the experiment, they believed the tools had made them about 20% faster.

The actual result went in the opposite direction.

Tasks completed with access to AI took 19% longer.

The finding was striking partly because the developers themselves did not notice the slowdown.

That does not mean AI coding tools generally make developers slower. METR explicitly cautioned against making that generalisation, and the technology has moved quickly since the experiment. In a February 2026 follow-up, the organisation said its newer data suggested late-2025 AI tools were probably speeding up developers more than the earlier systems had, although problems with the study sample meant it could not reliably estimate how large that improvement was.

That update is important.

The lesson is not that AI failed at coding.

It is that the productivity effect is conditional and moving.

It changes with the model, the task, the worker, the existing codebase, the amount of context the AI has and the amount of checking required afterwards.

Which makes broad claims that "AI increases productivity" technically true in some settings and surprisingly unhelpful in practice.

A better question is: which part of the work did it make more productive?

Because AI is extremely good at making the first part of many tasks cheaper.

Producing the first draft.

Generating the first version of the code.

Summarising the documents.

Suggesting the options.

Creating the analysis structure.

Turning notes into something polished.

That shifts where the expensive part of work sits.

Before generative AI, producing something often required much of the effort.

Now the effort increasingly moves towards deciding whether what was produced is actually any good.

And organisations are beginning to see the cost of that shift.

The time saved can come back as rework.

A 2026 Workday study surveyed 3,200 full-time employees who actively use AI across North America, Europe, the Middle East, Africa and Asia-Pacific.

The headline number initially looks extremely positive.

85% said AI saves them between one and seven hours per week.

But Workday also found that around 37% of the time employees reported saving was being lost again through correcting errors, rewriting material and verifying AI-generated output. Only 14% of respondents consistently reported what the study classified as clear positive net outcomes.

This is survey evidence rather than a controlled productivity experiment, and Workday is itself an enterprise AI vendor, so it should not be treated as a universal measure of AI's economic impact.

But the underlying problem is difficult to ignore.

AI can reduce production time while creating verification time.

A separate survey from BetterUp Labs and Stanford's Social Media Lab gave that problem a memorable name: workslop.

The researchers used the term for AI-generated work that appears polished but lacks the context, substance or accuracy needed to be useful. In their survey, 40% of US desk workers said they had received this kind of work during the previous month. Employees reported spending close to two hours dealing with each incident.

The paradox is obvious.

One employee saves ten minutes generating something.

Three colleagues each lose twenty minutes trying to understand or correct it.

From the perspective of the person who pressed "generate", AI was productive.

From the perspective of the organisation, it may not have been.

This is why measuring AI productivity through output alone can become misleading.

The employee drafted more emails.

The marketer produced more concepts.

The developer generated more code.

The analyst created more slides.

But more is only useful when the extra output survives the rest of the system.

AI has reduced the effort it takes to produce something.

It has not reduced the effort required to decide whether that something is good.

And in some environments, reducing the cost of production may actually increase the amount of judgment required.

If producing ten ideas once took a day, people naturally limited how many ideas they produced.

If producing 100 ideas now takes minutes, somebody still has to decide which three are worth pursuing.

The bottleneck moves.

This does not mean the productivity story is weak.

There is a risk of swinging too far in the opposite direction.

The evidence is not telling us that AI copilots are useless.

Quite the opposite.

One of the most consistent findings across research is that they can be extremely effective when the task is well defined and the model is operating within an area where it performs reliably.

Customer service is a good example because the objective is relatively clear. A worker needs to understand a customer's problem, find relevant information and resolve the issue efficiently. The AI assistant can draw on patterns from previous interactions and provide useful suggestions in real time.

The 5,172-worker customer-support study found not only greater productivity but also an unusually interesting distribution of the gains. Less experienced workers benefited substantially more than the strongest workers. The researchers found evidence that the AI was effectively spreading some of the practices used by more experienced employees to people who had not yet learned them.

That could be one of the technology's most important effects.

AI may not simply make existing expertise faster.

It may make some forms of expertise easier to access.

A field experiment at Procter & Gamble pushes that idea further. Researchers studied professionals working on real product-development challenges and randomly assigned them to work individually or in teams, with or without generative AI.

Individuals using AI were able to produce work comparable in quality to traditional two-person teams without AI. The system also appeared to help employees cross functional boundaries: R&D specialists produced more commercially balanced solutions, while commercial employees generated more technically informed ideas. The final published study involved 791 professionals.

That is more consequential than shaving five minutes from an email.

It suggests AI can sometimes change who is capable of doing a task, not simply how fast the existing expert can complete it.

And economically, that is where the case becomes particularly interesting.

A relatively inexpensive software tool that allows a junior worker to perform closer to the level of an experienced worker, helps one professional perform some work that previously required two disciplines, or lets a customer-support agent resolve more cases without reducing quality can have a compelling return.

The problem is that those economics do not automatically transfer to every task performed by the same employee.

The same person who receives a major benefit from AI when summarising a large document may receive almost no benefit when making a sensitive strategic decision based on it.

AI therefore scales technically much more easily than it scales productively.

Giving 10,000 workers access to a copilot is easy.

Knowing which of their millions of individual tasks should be delegated to it is much harder.

Adoption is running ahead of redesign.

This may explain one of the stranger features of the current AI boom.

AI tools have been introduced into workplaces at extraordinary speed, but the structure of work around them often looks remarkably familiar.

Employees still have the same meetings.

The same approval processes remain.

The same responsibilities exist.

The same documents move between the same people.

AI is frequently inserted into an existing process as another tool rather than used as a reason to redesign the process itself.

Workday's 2026 survey found that in 89% of organisations represented, fewer than half of roles had been redesigned to reflect what AI could now do.

That distinction matters.

Imagine an analyst normally spends four hours preparing a report and another employee spends one hour reviewing it.

AI reduces the drafting time from four hours to one.

That looks like a huge productivity improvement.

But if the organisation responds by asking the analyst to produce four times as many reports, the reviewer may suddenly become the bottleneck.

Nothing about the AI solved that part of the system.

This is a familiar pattern in technological change.

Making one stage dramatically faster can expose inefficiencies elsewhere that were previously hidden.

Generative AI is doing that to knowledge work.

It makes producing information cheaper.

So the scarce resources increasingly become attention, judgment, context and accountability.

Who decides whether an answer is correct?

Who notices what the model omitted?

Who understands the organisation well enough to recognise that a perfectly written recommendation is strategically wrong?

Who takes responsibility when AI-generated analysis becomes the basis for a decision?

Those are not edge cases.

They determine whether faster output becomes genuine productivity.

The real productivity opportunity may therefore look different.

The biggest benefit of AI copilots may not ultimately come from asking every employee to do the same job slightly faster.

It may come from changing how work is divided in the first place.

Let the AI handle the parts where generation is cheap and verification is straightforward.

Let humans spend more of their time where context, judgment and responsibility matter.

That sounds obvious, but the distinction is important because different tasks have very different verification costs.

If AI formats a spreadsheet, the result can often be checked quickly.

If it summarises a meeting transcript, a worker may already know enough context to recognise an obvious mistake.

If it drafts a legal argument, assesses a strategic acquisition or writes unfamiliar production code, verification may require almost as much expertise as producing the work manually.

The economics therefore depend partly on a simple relationship:

How much work does AI remove compared with how much checking it creates?

Where that ratio is favourable, copilots can deliver substantial productivity gains.

Where it is not, they may simply move labour from creation to correction.

This is why the productivity debate increasingly looks less like a verdict on AI itself and more like a question of workflow design.

The evidence that AI can make workers faster is already strong.

The evidence that it makes every worker, on every task, more productive is not.

And it probably never will be, because work is too varied for a universal effect.

The more useful question is becoming much more specific.

Where should humans stop doing the first draft, and where should they never stop making the final call?