Human-AI Collaboration: Why AI Often Adds Work Instead of Saving It

AI can make a task quicker and still leave the business carrying more work. Both things are true at once, and that is the part the boardroom keeps skipping past. Most investment cases assume the opposite, that every hour saved becomes an hour freed for something better. It is a comfortable assumption. It is rarely tested.

Someone has to make the saving real, on purpose. Leave it to chance and the freed-up space fills with something else: another report, another experiment, another stream of output that somebody now has to check. The machine looks productive. The person at the end of the line is still working late.

Deloitte's 2026 Global Human Capital Trends offers a better idea than the familiar race to automate: humans and machines working in concert. I like the phrase. Read it next to research in the Harvard Business Review, though, and a harder question surfaces. Who decides how that combined capability gets used, and who actually benefits?

I have argued for a while at 360 Strategy that culture comes first. These findings drag that argument out of the strategy deck and into the working day. A culture that values judgement has to give people the time, the authority and the confidence to use it. Without that, the human advantage is a flattering phrase bolted onto an increasingly exhausting way of working.

Productivity can bring its own workload

Start with the uncomfortable evidence. Aruna Ranganathan and Xingqi Maggie Ye (2026) spent eight months inside one US technology company of about 200 people, watching how the work actually changed and backing it with more than 40 interviews. Nobody was ordered to use AI. They used it anyway, and as they did they took on more responsibilities, ran more tasks at once, and let work seep into the gaps that used to be breaks.

Then came the knock-on. Experienced engineers found themselves checking and correcting AI-assisted work from colleagues who had wandered into unfamiliar technical territory. The saving in one person's workflow quietly became a demand on somebody else's. Nothing had been automated away. It had moved.

This is one company, and the research is still in progress, so I would not hang a universal law on it. What it does earn is a question worth asking in your own business. Has the time saved in one role simply reappeared as work in another?

None of this means the productivity gains are not real. They are. A 2023 study of customer support staff given an AI assistant recorded a productivity lift of around 14% on average, with the least experienced workers gaining the most (Brynjolfsson et al., 2023).

Both findings can be true at the same time, because they measure different things. One counts how fast a defined task gets done. The other counts what the whole working day feels like by six o'clock. The leadership mistake is to reach for whichever number flatters the decision you have already made.

Jevons gives us a useful warning

There is an old idea that explains why. In The Coal Question, William Stanley Jevons (1865) noticed something awkward: making coal more efficient to burn did not shrink demand for it, instead, it grew it. Cheaper useful power opened up new uses, and the new uses swallowed the saving per unit and then some.

The parallel with AI is useful, as long as we do not inflate it into a law of office life. Lower the effort per task and you invite more tasks. Whether the total human workload actually climbs depends on three things: real demand, the review burden left behind, and the choices managers make about the capacity that opens up.

Picture it with round numbers. A team spends an hour preparing each of ten reports, so ten hours in all. AI halves the preparation to thirty minutes, and five hours fall out of the week. Good news, until the business, seeing how cheap a report has become, commissions thirty of them. Now preparation alone runs to fifteen hours, and the checking sits on top of that.

The numbers are invented to make the shape visible, not to prove AI caused the extra demand. That is rather the point. A business case that quotes the saving per task and stays silent on the expected volume has told you only half the story.

Figure 1. Author's illustration informed by Jevons. Preparation time only; assumed values demonstrate a possible rebound in workload. They are not observed AI results.

Thirty reports might be a fine decision. If the extra volume puts a valuable service within reach of customers who were priced out before, that is growth worth having, and it can justify hiring, training and investment. Equally, those thirty reports might include twenty that nobody asked for and nobody reads. The dashboard counts both with the same cheerful enthusiasm.

So the decision about that extra demand belongs in the business case, out loud, before the tools go in. Leave it unspoken and the promised benefit becomes remarkably hard to find when someone eventually goes looking for it.

What working in concert requires

This is where Deloitte's framing earns its keep. It describes reimagining work so the best of humans and the best of machines pull towards the same outcome, with real attention paid to how the two interact (Deloitte, 2026, pp. 3, 9).

The concert in that phrase is doing more work than it looks. An orchestra is not loud people playing quickly; it is listening and timing. Hand everyone an instrument and tell them to play faster, and you have not made music; you have organised a noise complaint. Someone still has to hold the whole thing together: what the piece is meant to be and where each part comes in.

Take something ordinary, like putting a proposal together. AI can gather the material and draft the options in minutes. The account lead brings what the model cannot, a feel for the customer. A delivery specialist pressure-tests whether what is being promised can actually be built. When they disagree, the work goes back round before anyone commits to it, and what they learn on this proposal sharpens the next one.

That view also changes what we mean by adoption. Installing a tool and counting logins tells you people opened it, nothing more. It says nothing about whether the proposal got better, the customer understood it, or the team made a smarter commitment. Those outcomes come from the quality of the whole arrangement, and that includes the moment somebody looks at the machine's suggestion and decides it is not good enough. It is the same instinct I set out in the five stages of AI adoption.

In a smaller business, that is what orchestration means to me: a clear purpose, clean handovers, and room to push back. Let the roles follow the task and the evidence, not the org chart. Resist the two lazy assumptions that every human touch adds value and every machine output needs fixing, because both quietly undermine the thing you are trying to build.

The distance between intent and progress

Deloitte's survey puts a useful number on the gap between knowing and doing. On designing human-machine interactions, 66% of leaders recognise it matters, 57% have something underway, and just 6% say they are making great progress. On organising people, skills and resources to get work done, it is 88%, 77% and 7% (Deloitte, 2026, pp. 9, 40).

These are self-assessments, progress included, so read them with that in mind. The middle numbers are the ones that matter. They show a lot of activity already underway, and writing off everyone outside that final few per cent as idle would misread what the survey actually says.

Figure 2. Redrawn from Deloitte's 2026 survey figures on printed pp. 9 and 40 (PDF pp. 11 and 42). Self-reported responses to importance and progress questions; categories are not additive or a conversion funnel.

For a UK SME the survey is a prompt, not a template. Its global numbers still need reading against your own resources and your own customers. I would use it to go and look at one specific handover that keeps failing, rather than to copy a multinational's org chart, which tells you almost nothing about why your own process keeps jamming.

A faster curve still needs a destination.

Deloitte also sketches a compressed S-curve, with the moments of reinvention falling closer together for the strongest performers. It is a concept, not a calendar. Nothing in it tells your particular business the week to walk away from what is working now (Deloitte, 2026, p. 3).

Figure 3. Adapted from Deloitte's 2026 report, Figure 1, printed p. 3 (PDF p. 5). Conceptual illustration, not measured growth or a forecast.

After more than thirty years building ventures, I have learned to distrust any story about innovation that forgets the cost of getting the timing wrong. Move early and you can open a market before anyone else. Move too early and you tie up cash and attention long before you understand what customers will actually pay for.

Working in concert should help a team see those moments coming and respond together, rather than lurching. My reading of the curve is a simple one: shorter learning cycles, with the room to change your mind. Reinvention as a permanent state of emergency would be a very expensive way to read the same chart.

Human judgement needs conditions in which to work

I want to be careful with the phrase 'human advantage' because it can turn smug very quickly. People bring bias, sloppy reasoning, and a stubborn loyalty to decisions they have already sunk money into. A human in the loop reassures nobody if that human lacks the expertise, feels unable to say no, or has thirty seconds to approve something they never properly read.

That caution includes me. Experience helps me spot patterns, and it also makes a familiar explanation feel sturdier than the evidence deserves. The discipline I try to hold to is a single question. What would change my mind? It matters most when the machine's answer happens to agree with me.

Moreover, on decisions that carry real consequences, I would want the accountable person to set out the evidence, the strongest argument against, and the conditions under which they would stop or turn back. AI can help build every part of that case. Accepting it, though, still has to have a name attached.

The same discipline applies to the data underneath. An AI-generated summary should always stay distinguishable from the source it is drawn from. Let an unsupported claim slip into an internal knowledge base, and months later it gets pulled back out as established fact, and now the organisation is quietly recycling its own mistake.

Keeping the original evidence, and logging corrections when they are made, gives people something to push back with. The danger here does not need a model quietly retraining itself in the background. Pulling an old mistake back out and trusting it is more than enough.

The capacity to challenge

There is a training question buried in all of this. When junior colleagues lean on AI to get through unfamiliar work, they need feedback that explains why an answer is right, not just that it passed. Skip that, and you build the appearance of competence on top of judgement that has never actually been tested.

I would carve out protected time for people to talk through their reasoning and to work a few problems the hard way, without the tool. They will need that depth on the day the stakes are high enough to challenge what the machine recommends. Put it in the plan, not in the nice-to-have column.

Decide where the saved capacity goes

Give the outcome an owner. I would put the business leader responsible for the workflow in charge of the combined result, human and machine together. Finance validates the benefit, technical colleagues judge how reliable the thing is, and the people actually doing the work help establish what has really changed. Then give that owner the authority to settle the competing demands that follow.

This is the sort of decision I work through with owner-managers and their boards, and it is what 360 Strategy provides AI consulting in Scotland to help them get right.

Count the work at both ends

Back to that proposal process. Measure the whole of it against the old way: preparation, review, corrections, and how the customer responds. Remember that human first drafts always needed checking too, so separate the quality assurance you would do anyway from the rework the AI created, and count both properly. Then ask the unglamorous question. Is the work quietly bleeding into people's evenings? A quicker first draft answers almost none of the investment case on its own.

Then decide, deliberately, where the freed capacity goes. Some of it can chase new customer demand. Some can go into learning, into raising quality, or simply into letting an overstretched team breathe again. There is no allocation that is right for every company. There just has to be a conscious one for yours.

Finance has to hold one line firmly: available time is not the same as a realised saving. Freeing five hours does not cut the wage bill by itself, and it does not magically become five hours of billable work. The benefit is whatever the business actually does with that capacity, once you have paid for the tools and the effort of running them reliably. This is the point where a promising demonstration has to meet the accounts.

Agree the measures before the trial starts, not after, and put a review date in the diary. Dig into the disappointing results, scale the ones the evidence backs, and be willing to switch the whole thing off. Good judgement should be as quick to recognise a win as it is to expose a weak claim.

The people doing the work should have a say in how the benefit is shared. That might be protected learning time, a workload that is actually sustainable, or a real stake in the upside they helped create. What is fair will vary from one business to the next. What cannot wait is the answer, because if every efficiency gain simply resets as next quarter's target, ask yourself what earthly reason a capable person has to tell you how much time they just saved.

Underneath all of it sits trust. Talk about augmentation while quietly turning every saving into a headcount cut, and people will read the gap between the words and the outcome long before you do. They will keep using the tools. They just will not hand over the learning that would make those tools far more useful. Culture is not the soft, optional part of this strategy. It decides how much of the truth ever reaches the people making the decisions.

Give oversight the authority to matter

When the review capacity is stretched, the workflow owner needs to be able to ease off the pace, not just watch the queue build. Keep the controls proportionate to what an error would actually cost, and give staff a clear route to pause or escalate. Responsibility without the authority to act on it protects almost nobody.

A board conversation worth having would look at three things: where AI improved the finished outcome, where it just shifted the effort onto someone else, and which of last year's assumptions now need tearing up. I would want to hear about the workflow that was stopped for good reasons just as clearly as the one that was scaled after it earned it.

The leadership choice behind the technology

The promise of people and machines working in concert is worth chasing. Done well, it lets a business serve customers better, take on work that used to be out of reach, and build real capability across the team. None of that arrives on its own. It depends on deliberate choices about how the work fits together and how the rewards get shared.

That brings me back, as it always seems to, to culture. The real test is a simple one. What happens in your business when a capable person says the output needs another look, the workload has tipped into unreasonable, or the decision on the table is not backed by the evidence? Whatever happens next tells you more about your readiness than any adoption dashboard ever will.

Before we celebrate the hours AI has saved, the better question is who now controls those hours, what they are being used for, and whether the people doing the work would even recognise the improvement we are claiming.

Mark Evans is the founder of 360 Strategy, an AI readiness, governance and strategy consultancy based in Scotland.

Does AI adoption reduce a team's workload?

Not on its own. AI adoption can make individual tasks faster and still raise the total amount of work, because cheaper output invites more of it and someone has to check what the machine produces. The productivity gain is real. Whether it reaches the bottom line depends on what you do with the time it frees, so measure the work at both ends before you count the saving.

How should an SME approach AI adoption and scaling without burning out its team?

Start narrow. Prove it on one workflow, then scale what the evidence supports rather than everything at once. AI adoption and scaling tend to go wrong when the pilot's saving quietly becomes next quarter's target and nobody decides where the freed capacity should go. Give one person ownership of the workflow, count the review and rework as honestly as the time saved, and hold some of the gain back for learning and recovery.

What kind of AI training do teams actually need?

Less tool tuition than most people expect, and far more judgement. A team can learn the buttons in an afternoon; what takes longer is the confidence to challenge an answer, spot a weak one, and know when the stakes are high enough to slow down. Good AI training protects time for people to reason through real problems, rather than watch another demo.

How do you choose the right AI company or consultant?

Look for one that measures the workload you'll carry as closely as the time you'll save, and will tell you when not to use AI. Be wary of anyone selling adoption as a dashboard of logins. That's the standard we work to at 360 Strategy (linked to your service page), across AI readiness, governance, adoption and team training for SME founders, owner-managers and boards.

What's the honest way to measure the time AI saves?

Compare the whole workflow against the old way, not just the first draft. Count preparation, review, corrections and rework, and ask whether the work is creeping into people's evenings. A faster first draft that triggers three rounds of checking hasn't saved anything. Available time isn't the same as money saved either, so decide deliberately where any real capacity goes.

References

Brynjolfsson, E., Li, D. and Raymond, L.R. (2023) Generative AI at work. NBER Working Paper No. 31161. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w31161 (Accessed: 17 September 2026).

Deloitte (2026) 2026 Global Human Capital Trends. Deloitte Insights. Available at: https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends.html (Accessed: 17 September 2026).

Jevons, W.S. (1865) The Coal Question. London: Macmillan. Available at: https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ.html?chapter_num=9 (Accessed: 17 September 2026).

Ranganathan, A. and Ye, X.M. (2026) 'AI doesn't reduce work—it intensifies it', Harvard Business Review, 9 February. Available at: https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it (Accessed: 17 September 2026).