Human plus AI should beat AI alone. In one famous trial, it barely registered. And the reason is the most important lesson I know about deploying this technology.
In a 2024 randomised trial of diagnostic reasoning (Goh et al., JAMA Network Open), GPT-4 alone scored 92%. Doctors using conventional tools scored 74%. Doctors given GPT-4 scored 76%, a difference the researchers found was not statistically significant.
Read that again. The model on its own was far better than any human setup. Yet handing it to the doctors changed almost nothing. They scored 76% with it and 74% without it. The benefit was sitting right there, and the pairing did not pick it up.
It would be easy to read this as "humans are the problem, take them out." That is exactly the wrong lesson, and getting it wrong is how good organisations talk themselves into bad decisions.
What that result actually shows
The doctors were not redundant. They simply failed to capture a benefit that was on the table, and the study itself points at why. The AI Index, summarising this work, attributes the flat result to workflow, training, and interface design, not to the doctors' competence. My own read, having watched this pattern up close more than once, is that capable people lean on the habits that have always served them and treat an unfamiliar tool as an optional second opinion rather than a colleague. So they kept their own score, instead of climbing towards the model's.
That is not a verdict on human judgement. It is a skills and process gap, and it is fixable. The technology did its part. What was missing was the training, the protocol, and the design of the collaboration around it.
So the headline I would write is the opposite of the obvious one. Human plus AI is still the best model. But only when it is designed, not assumed.
There is a harder case hiding in this study, and I want to meet it head on. For some narrow, well-bounded tasks the data really does point to full automation, and pretending otherwise is its own form of denial. The doctor example is interesting precisely because diagnosis is not one of those tasks. Liability, edge cases, and accountability mean a clinician stays in the loop. The art is telling the two kinds of work apart, and most of the value is in getting that line right.
The comfort blanket problem
Naive "human in the loop" has quietly become a comfort blanket. We bolt a person onto the workflow, feel reassured, and never check whether they are improving the output or just slowing it down. A human whose only measured contribution is over-riding correct outputs, while adding nothing on the cases the system gets wrong, is not a safeguard. The whole job is designing the role so the human sits on the cases the model gets wrong, not the ones it gets right.
This matters because the instinct to add a human everywhere is well-meaning and wrong in the same breath. It feels responsible. It looks like governance. But a person placed in a process without training or a clear remit does not reduce risk. They add noise, and they give everyone false comfort that the workflow is safe because someone is "watching." Position that same person on the eight percent the model gets wrong, train them for it, and they become the most valuable part of the system.
What AI does brilliantly, and why that helps people
Let's be honest about what the technology is good at, because pretending otherwise wastes time and, ironically, wastes people.
AI is better than any of us at processing large volumes of structured information, matching patterns across datasets too large to hold in one head, holding consistency across thousands of decisions, and working at scale without fatigue or mood. None of that is a threat. It is the boring, high-volume work finally being done well, so the humans can go and do the work only humans can.
McKinsey (Economic Potential of Generative AI, 2023) estimated generative AI could automate work activities that absorb 60 to 70% of employees' time today. That is an economy-wide ceiling on automatable activity, not a target, and not a claim that two thirds of jobs disappear. My own evidence is narrower and concrete. The agentic system I built for a European insurance brokerage reached 67% autonomous case resolution, with a 23% lift in sales KPIs. Not because the technology was extraordinary, but because most of those cases genuinely did not need human judgement. They needed pattern matching, data retrieval, and rule application, which machines do better.
The economy-wide ceiling and our single workflow are different measurements, and I would not pretend the matching numbers corroborate each other. What they share is a direction. That 67% was not 67% of people made redundant. It was 67% of cases that no longer pulled a skilled person away from the third where their judgement was decisive. Same people. Far more of their time on the work that actually needs them.
The frontier is real, and it cuts both ways
A 2023 Harvard Business School field experiment (Dell'Acqua et al.) with 758 BCG consultants mapped what they call the "jagged frontier" of AI capability. Consultants using GPT-4 completed 12% more tasks, 25% faster, with 40% higher quality, but only for tasks within the AI's capability frontier. For tasks outside it, consultants using AI were around 19 percentage points less likely to produce correct solutions. They trusted the model when they should not have. The best performers were the ones who knew exactly where the frontier was, and handled the other side themselves.
Notice what that study is really saying. The people who lost ground were not stupid, and the tool was not broken. They simply had not learned where the model is strong and where it is not. The people who won had. That is the whole game: not human versus AI, but trained human plus AI versus untrained human plus AI. The gap between those two is enormous, and it is a question of skill and process.
This is the same lesson as the doctors, in a different setting. When you understand that an AI is a probabilistic system, not an oracle, you stop deferring to it blindly and you stop over-riding it out of reflex. You learn the frontier. That learning is the job now.
Where human judgement is decisive
So where do you put the people? Not everywhere, out of habit. In the places where their judgement genuinely changes the answer. Four capabilities are, for now, where the deliberate human earns their place.
Strategic intent. Knowing what is worth doing
AI can optimise any objective you give it. What it cannot do is decide which objective matters. It can tell you the most efficient path to a goal. It cannot tell you whether the goal is worth pursuing.
This is not a small distinction. The most expensive mistakes in business are not execution failures. They are strategic ones. Building the wrong product perfectly. Optimising the wrong metric efficiently. I have seen AI projects fail not because the technology did not work, but because nobody asked whether the problem was worth solving. The high failure rate in AI projects is rarely a technical problem. Someone has to own the intent and say "this is worth doing" or, more usefully, "this is not." That someone is human.
Taste, judgement, and opinion
Business taste matters. The intuition that comes from twenty years in an industry. The ability to look at a proposal and know, before the data confirms it, that something is off. The judgement to say "this is technically correct but strategically wrong."
The model hands you the average answer, and average is rarely what you are paid for. The organisations deploying AI well are not the ones deferring to AI outputs. They are the ones whose people have strong enough judgement to know when the model is right and when it is confidently wrong. Data tells you what happened. Judgement tells you what it means.
Human-to-human relationships
No amount of AI sophistication replaces the trust that forms between people who have worked through hard problems together. When I co-founded GeoVS, the deal that took the company from PhD research to a sale to a listed marine systems group was not closed by a model or a deck. It was closed across a table, in conversations where the other side had to believe I understood their world and would not let them down. A client does not hire me because an algorithm recommended me. They hire me because a conversation reveals that I understand their situation in a way that feels different from the others.
This is economics, not sentiment. B2B sales cycles, board decisions, partnerships, key hires. The moment where someone commits resources, takes a risk, or changes direction is human. AI can prepare the analysis and draft the deck. It does not own the relationship. The shift from individual contributor to orchestrator makes this more important, not less: as the machine handles more of the execution, the human work of communication and trust becomes a larger share of what decides success.
Accountability and contextual wisdom
AI systems optimise. They find the path of least resistance to whatever objective you defined. They do not ask whether the objective is ethical, do not weigh second-order consequences on people absent from the training data, and do not feel uncomfortable when something is technically legal but plainly wrong.
I have sat in the expert-witness chair, where a judgement on someone else's technical work carried real consequences for real people, and no amount of model output relieves you of owning that call. Every organisation I have worked with has faced the same in milder form: decisions where the right answer was not the most efficient one, where serving a customer well meant absorbing a cost. The governance questions around shadow AI are a good example. A system will happily process sensitive data if you let it. Knowing when it should not, and being willing to own that call, requires human judgement about risk and responsibility. Someone has to be accountable when the decision is hard. A model cannot be.
Deliberate humans, not fewer humans
Put all of this together and the strategic move is the opposite of "take the humans out."
Automate the work that is genuinely routine, so your people are not spending their hours where a machine does better. For the narrow tasks where automation is total, accept that and move on. Then put humans exactly where their judgement is decisive: the high-trust calls, the strategic intent, the moments where someone has to own the outcome. And teach them to work with the machine rather than around it, because that is the part most organisations skip.
The dual-stream approach I advocate is precisely this: on one track, automate every process that can be automated; on the other, deliberately develop the human capabilities and collaboration skills that machines cannot replicate.
Restructuring roles, not cutting them
This means restructuring roles, not eliminating them. Same headcount, very different output. The advantage is in the deployment ratio and the training, not the technology, which is why it is so hard for competitors to copy. It means telling your team something like: "We are going to take away the parts of your job a machine does better, and invest in making you exceptional at the parts only you can do, including how to work alongside the AI." That is a harder conversation than "AI is coming for your jobs." It is also more honest, and far more productive.
The "for now" honesty
I want to be honest about the boundary. The human edge moves, and it moves in one direction, towards the machine. Five years ago I would have put competent prose on the safe list. Today agentic systems are making architectural calls that are, often, good enough. Anyone who tells you that strategic thinking, relationship, and judgement are permanently safe is making a prediction about a technology whose trajectory they cannot honestly know.
There is a quieter risk I see on the ground too. Critical thinking atrophies in teams that lean too hard on AI outputs. Junior staff who do not first form their own view before consulting the model lose, over time, the muscle that makes them a useful check on it. That is not an argument against the tool. It is an argument for designing the collaboration so people keep thinking, which is the whole point of this piece.
So how long does the edge last? No one can honestly forecast it. Plan as if the window is short, one to three years, and treat anything beyond that as a bonus. The right response to a shrinking edge is not to pretend it is not shrinking, and certainly not to surrender it by pulling people out. It is to design the pairing well, train for it, and make the advantage count while it is decisive.
What this means for you
The practical shape of this is simple to say and hard to do. Stop adding humans out of habit, and stop removing them out of panic. Both are lazy. Find the work where a person genuinely changes the outcome, and put them there on purpose. Then train for the collaboration: the doctors did not fail because they were doctors, they failed because no one taught them to work with a model, and teaching capable people where the frontier is and how to work across it is the single highest-return investment in most organisations right now. And be honest about the timeline, building your strategy around an edge that is real but not permanent.
The most valuable people in any organisation have always been the ones who know what is worth doing, not just how to do it. AI has not changed that. Paired well, it makes those people count for more, not less.
If you are working out where a human genuinely belongs in your AI workflows, and how to train your people to make the pairing better rather than slower, that is one of the most useful things an outside pair of eyes can help with. Let's talk.
Related: You're not a 10x engineer. You're an orchestrator · You're not talking to an LLM the way you think you are · Most companies are adopting AI. Few are adopting it well · Traditional moats are dissolving
