What are the humans for?

Designing work when AI can do more of it

By Ross McCulloch, Third Sector Lab

I read a piece by Anna Aston recently, When Machines Do the Thinking, What is Left for Us?, and it has been rattling around my head ever since. It asks a question I think most organisations are avoiding, mostly because the other questions are easier.

Which AI tool should we buy? How do we get everyone using Copilot? Which tasks can we automate? Those are procurement questions dressed up as strategy. The harder one underneath is: what do we actually want the humans to be responsible for?

That used to answer itself. For most of the digital era, technology helped people do the jobs they already had. A database replaced a filing cabinet, email replaced letters, Teams replaced a few meetings you never wanted to attend anyway. The person was still obviously doing the work.

That is no longer obvious. A system can now research a subject, summarise a meeting, interrogate thousands of documents, analyse survey responses, draft a strategy, write a funding application and carry out multi-stage work with very little intervention. So the interesting question is not where AI might be useful. It is how we deliberately divide responsibility between people and machines, and I do not think that division should be decided by what the technology happens to be capable of this month.

Capability is not permission

Just because AI can do a task does not mean your organisation should give it that task. I think this is the principle everything else hangs off.

Researchers at Stanford went and asked workers what they actually wanted automated, rather than measuring what could theoretically be automated. Across 1,500 workers, 104 occupations and 844 tasks, they found that workers consistently wanted more human involvement than AI experts thought was technically necessary – in 47.5% of the tasks studied, people preferred a higher level of human agency than the experts said the technology required (Stanford, Future of Work with AI Agents).

That gap is the whole story. The engineers are answering “could an AI system do this?” The people doing the work are answering something quite different: “should this be done without me?”

Picture a charity supporting someone facing homelessness. AI could probably analyse that person’s circumstances, compare them against previous cases and recommend a course of action. It might even produce a good recommendation. But does the person in front of you expect another human being to hear their story? Does understanding their situation depend on context that never makes it into a form? Who carries the can if the recommendation is wrong? Do you want people in crisis to feel that decisions are made with them rather than about them?

Technical capability answers none of that, because none of it is a technical question. They are questions about values, relationships and what your organisation is for. Charities in particular should be wary of letting technical possibility quietly harden into organisational policy.

Stop designing jobs around tasks

Aston makes a useful distinction between execution and judgement. AI is increasingly good at producing, classifying, searching, summarising and recombining. Human value shifts towards deciding which questions matter, reading context, handling ambiguity and owning the consequences.

The World Economic Forum has made a similar argument, describing the emerging human roles as framing the work before AI starts and assessing what should happen after it produces something. AI may execute, but someone still has to turn messy reality into a problem worth solving, set the limits, review the result and decide what happens next (WEF, defining roles for humans and AI).

That suggests a completely different starting point. Do not take your existing job descriptions and go through them asking which bullet points AI can handle. Start with the outcome you want. Ask what has to happen for that outcome to be good, safe and consistent with your values. Only then decide what people do, what people do with AI support, and what you are willing to hand over entirely.

That is work design, and it is a great deal more interesting than shaving ten minutes off someone’s inbox every morning.

Five questions for deciding what stays human

There is no universal chart telling you where the line sits. It moves depending on the organisation, the people affected, the consequences of getting it wrong and the purpose of the work. But these five questions are a decent place to start.

1. Does someone need to be accountable for this?

AI can recommend an action. It cannot accept responsibility for one, and that distinction gets more important the more capable the systems look.

Say an AI tool assesses applications for a small grants programme and recommends who gets funded. It might spot patterns a panel would miss. It will certainly be faster. It may even land close to what an experienced grants team would have chosen. But someone still has to decide what fairness means here. Someone has to decide how much weight lived experience carries. Someone has to notice that a small grassroots group has written a weak application while doing extraordinary work. And someone has to be able to look the unsuccessful applicant in the eye and explain the decision.

“Computer says no” was not acceptable twenty years ago. Giving the computer a much better vocabulary does not change that. The WEF calls this human role stewardship: validating outputs against real-world context, then deciding whether to accept, modify, reject or escalate. So if a decision has meaningful consequences, do not just ask whether a human reviews it. Ask which human owns it.

2. Does the work need judgement rather than a correct answer?

A lot of organisational work has no correct answer. Should we fund this programme? Should we close this service? How do we respond to a complaint that is half justified? Is this campaign appropriate? What level of risk are we willing to carry?

AI can give you evidence, options and arguments, and that is genuinely useful. But choosing between competing values is a different activity from calculating the most probable answer.

MIT Sloan researchers Roberto Rigobon and Isabella Loaiza analysed around 19,000 tasks across 950 job types and pulled out five clusters where humans still matter most. They call it EPOCH: empathy, presence, opinion and judgement, creativity and hope (MIT Sloan, choose the human path for AI).

I like the inclusion of opinion. Organisations often pretend that big decisions are simply what falls out of enough evidence. They are not. Eventually someone has to say “I think we should do this”, and sometimes “the evidence points one way, but our values are taking us the other way”. Loaiza puts it well: people sometimes decide to do something not because the data says it is possible, but because they believe it should be done. In our sector, that is not a bug in the process. It is quite often the point of the organisation.

3. Is the relationship part of the service?

This one matters enormously for charities, care organisations, public services and anyone working with people in difficult circumstances, because it is so easy to separate the information from the interaction that carries it.

A bereavement service is not transferring advice. A youth worker is not answering questions. A fundraiser is not transmitting facts about a cause. A manager is not a task allocation system. People notice whether someone listened. They notice whether someone cares, and whether the person on the other end understands the history behind what they are saying.

EPOCH separates empathy from presence, which I think is the useful move. An AI system can absolutely produce empathetic language. That is not the same as someone in grief or fear wanting your organisation to swap human presence for extremely convincing empathetic text.

The OECD’s 2026 work on AI and skills lands in a similar place: empathy, communication and teamwork remain essential in many jobs, but algorithmic management may quietly reduce how much organisations appear to value them. The OECD’s argument is that we should be monitoring job quality, social interaction and wellbeing, not just productivity (OECD, AI and skills: what we know so far). That seems right to me. It would be a poor outcome to automate human contact and then work out five years later that the contact was the bit that made the service work.

4. Does doing the work develop the person?

This is the question I think we will overlook, because the cost of getting it wrong arrives years after the decision.

Sometimes the value of a task is not the output. A junior colleague writes a weak first draft. Their manager pushes back. They rewrite it, discover the statistic they leaned on was misleading, go and talk to someone with more experience, and rewrite it again. Eventually they get good. If AI produces the polished draft immediately, you have saved a couple of hours. What has your colleague learned?

Harvard Business Review has been circling this. David Duncan argues that experienced workers get enormous value from AI precisely because they already have the judgement to tell whether the output is any good, while less experienced colleagues struggle for exactly the opposite reason (HBR, How do workers develop good judgment in the AI era?). A later piece puts it more bluntly: as producing polished work gets easier, the scarce skill becomes knowing what to trust, what to challenge and what to improve (HBR, Help employees get better, not just faster, with AI).

This is a real dilemma, not a rhetorical one. AI is very good at exactly the tasks people used to develop expertise through: research, first drafts, note taking, basic analysis, summarising a discussion, comparing options, testing an argument. Look at any one of those in isolation and it reads as low-value work to be stripped out. Look at a career and it is the apprenticeship.

So it is worth distinguishing between work that is inefficient and work that is developmental. Sometimes friction is doing a job.

5. What will we do with the time we save?

I suspect this is the most neglected question in every AI strategy currently being written. Ten people each save five hours a week. Excellent. What happens to the fifty hours?

If the answer is fifty more hours of output, you have not redesigned work. You have sped up the system you already had. The alternative is to spend it on the things that always get squeezed: talking to the people you support, visiting projects, coaching staff, testing ideas, building partnerships, reading research, working out why something is not working, reaching communities you have never reached. Or rest, which is a legitimate answer.

Microsoft’s 2026 Work Trend Index surveyed 20,000 AI users across ten countries. Sixty-six per cent said AI let them spend more time on higher-value work, but the more interesting finding is about the most advanced users: they were more likely to stop before starting a piece of work and decide deliberately what AI should do and what a person should do. They were also more likely to do some work without AI on purpose, to keep their own skills sharp (Microsoft 2026 Work Trend Index).

That is a healthier measure of maturity than prompts per employee. If technology hands capacity back, decide what the capacity is for.

A simple way to redesign work

I think most organisations could start by sorting their work into four rough categories.

Human work. Activities where the human involvement is the point: building relationships, supporting someone through something hard, making consequential value judgements, leading people through uncertainty, negotiating, coaching, listening, carrying responsibility. AI might still provide background, but you deliberately put a person at the centre.

Human-led work with AI support. Probably where most knowledge work belongs. Strategy, research, communications, service design, fundraising, management, policy. AI searches, challenges, structures, summarises and offers alternatives, while the person stays intellectually involved throughout rather than turning up at the end to approve an answer.

AI-led work with human accountability. Meeting transcription, routine reporting, classification, data cleaning, first-pass document review, formatting, comparing large volumes of information, straightforward admin. The execution shifts, but a named person still owns the quality and the consequences.

Automated work. Repetitive, low consequence, easily checked, reversible, no developmental or relational value. Automate it and do not look back. Nobody gets a prize for preserving boring administration.

The aim is not to maximise human involvement. It is to put human attention where it matters.

The uncomfortable bit

We should be straight about something. If organisations genuinely automate significant amounts of work, some will need fewer people to produce the same output. Pretending otherwise helps nobody.

But even there, choices remain. You could cut headcount and deliver exactly the same services. You could use the same capacity to support more people. You could shorten the working week. You could move staff out of administration and into frontline or relationship-based roles. You could invest in research, learning and improvement. You could reduce workloads in parts of the sector where burnout has become the norm. You could run services that were previously unaffordable.

Technology changes the options on the table. It does not make the choice. MIT Sloan Dean Richard Locke makes this argument explicitly: the direction of AI adoption is not predetermined, and organisations can pick technologies and working practices that either replace human capability or extend it.

That choice matters more in our sector than most. Charities should absolutely use technology to remove unnecessary work – every hour spent copying information between systems is an hour not spent on something useful. But charities are not factories for maximising output. The work runs on trust, power, dignity, relationships and judgement. An efficiency that damages those things is not efficiency. It is just cheaper.

Be careful what you stop practising

There is a second risk here that gets far less attention: humans get worse at things they stop doing.

If we hand over writing, do we become worse writers? If we hand over analysis, do we become worse analysts? If we always ask a model for options before thinking, do we get less imaginative? If managers receive a generated summary of every conversation, do they gradually stop listening as carefully? If leaders always start from an AI-generated recommendation, how much does that recommendation anchor everything that follows?

A recent HBR article warns that organisations may be quietly weakening leaders’ independent judgement by surrounding them with systems designed to supply ever more analysis and recommendations (HBR, AI is undermining leaders’ judgment).

None of that is an argument for avoiding AI. It is an argument for treating human capability as something you maintain on purpose. We understand this everywhere else. Pilots train for the situations where automation fails. Cybersecurity teams run exercises before the incident. First aiders refresh skills they hope never to use.

Knowledge organisations may need the equivalent. Space where people analyse without AI, write without it, generate ideas before consulting it, form a judgement before seeing the recommendation. Not because the human answer will always be better, but because I would be wary of building an organisation where nobody can function independently of the systems around them.

This is a governance question

All of which moves AI out of the technology conversation and into the boardroom, where it should have been sitting all along.

Boards and senior teams should be asking more than “are we using AI safely?” They should be asking what role we want humans to play in this organisation. Where do we insist on human accountability? Where does human contact matter? Which decisions need genuine judgement? Which capabilities do we need people to keep developing? Which work can disappear completely? What do we want staff to do with the capacity we hand back to them?

And the hardest one: are we designing work around what is best for our mission and the people we serve, or around what the technology makes possible?

Microsoft’s research makes a relevant observation here. The biggest determinant of whether people report real value from AI is not individual proficiency. It is organisational: culture, management support, how work is structured. Which makes sense. You can give everyone the most capable tools available, but if nobody knows who is responsible for decisions, what good looks like, when human involvement matters or what the saved time is for, you have not transformed anything. You have added AI.

Not what AI can do

There will always be another model. It will write better, analyse more, run longer and handle more complicated work. Any strategy built around the current limits of the technology will be out of date by the spring.

So start somewhere more durable. Decide what you value about human contribution first. Decide where judgement matters, where presence matters, where relationships matter, where originality matters, where accountability matters, where doing the work is how someone becomes good at their job, and where another human being simply deserves another human being.

Then work out what the machines are for.