The human side of human–AI teams

The performance gap is human

A 2024 MIT meta-analysis of 106 experiments found large, systematic variation in human–AI team performance.⁴ The determining factor was almost always the same: not the capability of the AI, but the quality of the human–AI interaction. On tasks where human and AI capabilities were genuinely complementary—where humans brought contextual understanding, judgement and accountability that the model lacked—well-designed human–AI combinations consistently outperformed either alone. Where the interaction was poorly structured, performance consistently fell short of what either could achieve independently. The technology was the same. What changed was how humans worked with it.

Research consistently identifies three failure modes in human–AI teaming. The first is over-trust: humans accept AI outputs without sufficient scrutiny, a tendency known as automation bias. The second is under-trust: humans disregard AI outputs even when they are more accurate than their own judgement. The third is misallocation: AI is applied to the wrong tasks, or the boundary between human and machine contribution is left undefined.

Each failure mode is a human behaviour problem, not a technology problem. And human behaviour problems respond to design, leadership and deliberate practice. That is both challenge and opportunity.

The psychological disciplines

Workers who consistently extract superior performance from AI systems—measurably faster, higher quality or more accurate than peers working with identical tools—share a set of common disciplines. Three of those disciplines are psychological: they are about how people think and behave in relation to AI, not about what they technically know how to do.

Trust calibration

The tendency to over-rely on automated systems, automation bias, is well-documented across high-stakes environments.⁵ In clinical settings, studies have found physicians overriding their own correct diagnoses in favour of erroneous AI recommendations. In aviation, automation bias has been a contributing factor in several accidents, including the 2009 crash of Continental Connection Flight 3407, where the crew failed to override an incorrect automated response despite having both the capability and the time to do so.⁶ The pattern recurs whenever humans interact with systems that are highly reliable in most situations but silently wrong in specific ones.

Under-trust is equally costly and harder to measure. Workers who dismiss AI outputs reflexively leave most of the available performance gain untouched. High performers neither defer nor dismiss. They actively calibrate trust: they know what the system is good at and where it is unreliable, and they adjust their reliance accordingly. In practice, this means knowing the model's track record on the type of task at hand, not its general reputation. This is not intuition. It is a learnable discipline.

Cognitive vigilance

AI outputs require interrogation, not acceptance. Language models can produce reasoning that is plausible, fluent and wrong. High performers treat AI outputs as a strong first draft and an able sparring partner, not as a verified answer. They ask: What is this based on? What has it missed? What would change this conclusion? Critically, they also ask the model itself: what would argue against this? What have you not considered? The habit of structured challenge, applied both to the output and back to the model, is the difference between AI that amplifies thinking and AI that replaces it.

Intellectual humility

Effective human–AI collaboration requires honest acceptance that AI will outperform humans on specific tasks, and the readiness to intervene with confidence when it will not. The BCG study introduced the concept of a 'jagged technological frontier': AI capability is uneven, high in some tasks and low in adjacent ones, in ways that are not always intuitive.³ A model that writes compelling executive communications may systematically misread financial risk. A model excellent at synthesising research may produce confident nonsense on a niche regulatory question. Knowing where the frontier runs for your specific work is itself a capability that must be built and regularly updated as models change.

The capability skills

Beyond mindset, three practical skills drive performance in AI-augmented work. Taken together with the psychological disciplines, they are what separates high-performing human–AI collaboration from average use of the same tools.

Prompt discipline

The ability to interact productively with AI is an analytical skill, not a technical one. High performers do not look for prompt shortcuts. They frame problems clearly, specify constraints and context, request reasoning paths and iterate deliberately. The BCG study found that even a brief prompt-structuring overview meaningfully improved outputs.³ The underlying skill—formulating a problem precisely enough to explain it to a machine—is the same discipline that distinguishes strong analytical thinking from weak. It is also learnable at scale, which makes it one of the most tractable capability investments an organisation can make.

Output interrogation

Asking AI to critique its own reasoning, argue the opposite case, or test a conclusion against alternative assumptions produces materially better outputs. This is the capability dimension of cognitive vigilance: where vigilance is the mindset, interrogation is the practice. High performers develop a consistent repertoire of challenge prompts—'What would argue against this?', 'Where is this most likely to be wrong?', 'What have you assumed that I haven't told you?'—and apply them as a matter of discipline, not as an occasional check. Models that undergo this kind of interrogation produce substantially stronger outputs than those that do not.

Task orchestration

The highest-value question in AI-augmented work is allocation: what should the human do, and what should the machine do? The BCG research identified two distinct high-performing patterns. Centaurs cleanly partitioned work: AI for synthesis and exploration, human for judgement and decision. Cyborgs integrated continuously, moving fluidly between both throughout a task.³ Both patterns outperformed those who used AI arbitrarily. The common factor was intentional task design: an active, ongoing decision about where human judgement adds most and where machine processing does. This is not a one-time setup. High performers rebalance continuously as tasks evolve and model capabilities change.

The six disciplines of human–AI performance

These six capabilities converge into a framework that organisations can use to design, assess and develop human–AI collaboration at scale. Three are psychological disciplines: how people think in relation to AI. Three are capability disciplines: how people act. These patterns appear repeatedly across the research literature and in early enterprise deployments.

Human synthesis: the step that matters most

All six disciplines serve a single purpose: human synthesis. This is the step that transforms AI output into organisational value, and it is worth being precise about what it involves.

A language model can analyse options, draft communications, synthesise research and generate recommendations at speed and scale. What it cannot do is understand the full context in which a decision sits: the history of the relationship, the unstated priorities, the organisation's risk appetite, the credibility of a recommendation in the hands of the person delivering it. It cannot weigh institutional relationships, apply accumulated professional judgement or take accountability for the outcome. Only the human can.

Human synthesis is not a passive review step. It is an active, skilled integration: bringing together what the machine has produced with what only the human knows. The discipline matters because without it, AI produces outputs. With it, AI produces decisions that are grounded, credible and can be owned. The final step in any high-value task is always human. Not because AI is incapable, but because accountability, context and consequences are human territory.

The organisation design implications

Most organisations approach AI capability as a technology deployment problem. They invest in tooling, run training programmes and track adoption rates. This is necessary. It is not sufficient.

The six disciplines are not produced by access to tools. They develop through deliberate practice, explicit norms and supported experimentation. Building them at scale is an organisation design question, not a technology rollout question. Three lessons from early AI deployment in complex environments apply:

Start where verification is straightforward

The equivalent of 'lowest technical debt' in human–AI interaction is where humans have clear expertise to evaluate AI outputs: where they can calibrate trust quickly, catch errors confidently and build the interrogation habit at low risk. Financial analysis checked against known data, communications refined against established templates, research validated against existing expertise: these are natural starting points. Success here builds organisational confidence and generates the shared experience base from which operating norms develop. It also allows leaders to observe and codify what good human–AI interaction actually looks like in their specific context, before deploying into higher-stakes work.

Build the governance before you need it

Governance, oversight protocols, and accountability structures cannot be retrofitted after AI-supported decisions have led to consequential errors. The organisations making most progress have clear answers to the questions that matter before they become urgent: Who is accountable when an AI-supported decision turns out to be wrong? How are AI outputs documented when they inform material choices? How does the organisation know when a model's performance has drifted from what it was when trust was initially calibrated? These are governance questions that are easier to answer in design than in crisis.

Define the new human roles explicitly

Every AI deployment changes what people do, not just how much of it they do. The emerging role distinctions in AI-augmented professional work include those who own the human synthesis step, those who manage prompt discipline and task orchestration across a team, and those responsible for maintaining cognitive vigilance and output quality at scale. These are real roles with real skill requirements. Defining them before deployment, rather than after resistance or failure surfaces the need, is the difference between AI integration that compounds over time and AI adoption that stalls after the initial enthusiasm.

Capability development, leadership expectations and operating norms must all change together. Workers need structured practice, not just access. Managers need to model the disciplines and create the conditions for constructive challenge.

Emerging research shows that how leaders behave around AI tools directly shapes whether teams feel safe to question AI outputs, or whether they default to accepting them.^7,8 Operating norms matter more than policies. The difference between teams that use AI well and teams that do not is rarely a function of access. It is whether there are clear shared expectations: that AI outputs are challenged, that reasoning is made visible, and that human judgement is exercised rather than deferred.

Conclusion

The future of work will not be simply human or machine. It will be human judgement operating through machine intelligence. The organisations that succeed will not simply deploy AI. They will design how humans work with it.

The six disciplines are not a training curriculum. They are an organisation design agenda: what norms to establish, what roles to define, what governance to build, and what leadership behaviours to model. The organisations that treat them as such—rather than as a checklist for individual workers—are the ones that will extract sustained value from AI, not just initial productivity gains.

References

1. Noy, S. & Zhang, W. (2023). 'Experimental evidence on the productivity effects of generative artificial intelligence.' Science, 381(6654). doi.org/10.1126/science.adh2586
2. Brynjolfsson, E., Li, D. & Raymond, L.R. (2023). 'Generative AI at Work.' NBER Working Paper 31161. nber.org/papers/w31161
3. Dell'Acqua, F. et al. (2023). 'Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.' Harvard Business School Working Paper 24-013. ssrn.com/abstract=4573321
4. Vaccaro, M., Almaatouq, A. & Malone, T.W. (2024). 'When combinations of humans and AI are useful: A systematic review and meta-analysis.' Nature Human Behaviour, October 2024. arxiv.org/abs/2405.06087
5. Parasuraman, R. & Manzey, D.H. (2010). 'Complacency and bias in human use of automation: An attentional integration.' Human Factors, 52(3), 381 – 410. See also: Center for Security and Emerging Technology (2024). 'AI Safety and Automation Bias.' Georgetown University.
6. National Transportation Safety Board (2010). Aircraft Accident Report: Colgan Air / Continental Connection Flight 3407. NTSB/AAR-10/01.
7. Center for Security and Emerging Technology (2024). 'AI Safety and Automation Bias.' Georgetown University. cset.georgetown.edu
8. Reeves, M. & Whitaker, K. (2026). 'How to Foster Psychological Safety When AI Erodes Trust on Your Team.' Harvard Business Review, February 2026. hbr.org