Grund InstituteMost organizations treat AI adoption as a deployment and skills problem. Everyone gets the same tools, attends similar training and is encouraged to bring AI into daily work. The outcomes still differ widely.
One employee is soon using AI to explore problems, challenge assumptions and build solutions. Another finds the same tool unpredictable and mentally tiring. A third uses it every day and checks every response. A fourth accepts recommendations quickly, sometimes without enough checking. These differences are usually put down to digital skills, motivation or willingness to change. That explanation is incomplete.
People differ not only in their ability to use AI, but in how naturally their thinking aligns with the way AI systems work.
That is the principle behind Human-AI Cognitive Fit, a central concept in Grund Institute's research. It raises a management question every AI programme has to answer: should people adapt to one way of working with AI, or should AI-enabled work be designed for different ways of thinking? This briefing opens the third theme of the series, on human cognition, trust and AI collaboration.
§The overlooked human variable
Generative AI is fundamentally different from traditional business software. Conventional applications run defined functions on structured inputs and return predictable outputs. Generative systems are probabilistic: their responses depend on context, instructions, available information and patterns learned in training. People can ask open questions, explore alternatives, refine through dialogue and delegate increasingly complex work.
That power brings uncertainty. A response can be useful without being fully correct. A complete-looking answer can rest on unsupported assumptions. The same request can produce different outputs. Working well with AI means handling ambiguity, adjusting instructions, deciding when to trust the output and knowing when independent judgment is needed.
People experience those demands differently. Some are comfortable exploring uncertain possibilities and refining ideas as they go. Others prefer clear requirements, structured processes, explicit evidence and predictable results. Neither is better. What matters is whether the interaction and the workflow support the needs of the person and the task.
A system that feels intuitive to one person can create unnecessary cognitive effort for another.
Organizations that miss this end up treating an interaction-design problem as a training or motivation problem.
§AI benefits are not evenly distributed
Research shows that the gains from AI vary across people, tasks and circumstances.
Brynjolfsson, Li & Raymond · Quarterly Journal of Economics, 2025
The same AI capability did not produce the same benefit for everyone.
Studies of algorithmic trust show that people differ in their willingness to rely on AI recommendations and in how they respond to explanations and uncertainty. Research on AI-assisted decisions shows that the benefit depends on whether users can tell reliable assistance from questionable advice.
Human-AI performance depends on the interaction between the person, the system and the task.
Differences in performance should not be put down to cognitive style by default. Experience, task difficulty, domain knowledge, AI reliability, organizational context and access to information all play a part. Cognitive fit is one potentially important factor inside a wider organizational system.
§What cognitive fit means
Grund Institute's Second Mind framework treats AI as a possible extension of human cognitive activity, not only a tool that executes commands. It asks how people think, reason and solve problems when they work with probabilistic systems.
How they handle ambiguity, control, trust and engagement
Openness, transparency, autonomy, reliability
Risk, structure, need for judgment and evidence
Strong or weak, for this person, this tool, this task
When fit is strong, the interaction supports sound reasoning, manageable effort, well-calibrated trust and efficient completion. When it is weak, people experience confusion, excessive checking, frustration, lower confidence or difficulty using what AI produces.
Cognitive fit is not intelligence. It does not say whether someone is a good employee, a capable professional or suited to technical work. Nor is it a permanent trait. People change how they work with AI as they gain experience, develop strategies, meet new tasks and use systems designed differently. The same person can have strong fit with AI-assisted brainstorming and poor fit with an autonomous system making opaque recommendations.
Cognitive fit is relational and depends on context. It is not a ranking of people.
§Four dimensions of interaction
Grund Institute's research on interaction styles looks at four dimensions that help explain why people experience the same technology differently. Each is a range, not a score.
Tolerance for ambiguity
Generative AI often presents uncertain, incomplete or open-ended information. Someone who prefers certainty can find an open-ended assistant demanding even when it is technically capable.
Design response More structured interaction and clear acceptance criteria.
Control preference
Some people want to approve each intermediate result; others let AI propose approaches or run bounded work. This matters most once agents act autonomously. Too little control erodes oversight; too much creates friction.
Design response Clear decision boundaries, sized to the task. See Briefing 11.
Trust calibration
Effective use needs neither blind trust nor reflex skepticism. Some accept plausible output too readily; others re-verify even low-risk information. The right level depends on the system's track record, the evidence and the cost of error.
Design response Visible evidence and risk-based checks. See Briefings 03 and 08.
Cognitive engagement
Some people use AI as an interactive thinking partner, testing hypotheses and refining ideas. Others prefer to receive structured information and evaluate it themselves. This shapes how much effort a workflow demands.
Design response Interaction modes that suit the way the work is reasoned through.
These dimensions belong alongside task requirements, professional expertise, organizational process and system reliability, not in place of them.
§Four interaction patterns
GrundMind uses four interaction archetypes to describe recurring patterns. Each brings strengths and risks.
Engage through iterative exploration, questioning and building ideas together with AI. Strength: creativity, problem exploration, synthesis. Risk: unnecessary iteration or over-reliance if output is not evaluated carefully.
Evaluate suggestions, test reliability and adjust trust to the task. Strength: careful decisions and output quality. Risk: verification becomes inefficient without clear evidence or review criteria.
Prefer structured interaction, explicit requirements and control over how AI contributes. Strength: consistency, accountability, predictable execution. Risk: friction with open-ended tools unless the workflow supplies structure.
Prefer familiar approaches, need stronger proof of value, or have little interest in direct AI interaction. Strength: protect proven expert methods. Risk: misread as resistance, when conventional or human-led work is sometimes the better choice.
These are descriptive interaction patterns, not measures of worth or fixed predictions of performance. The same person can show different patterns depending on the task, the tool and the context. The objective is not to turn everyone into the same type. It is to design work that produces good outcomes across different interaction needs.
§The hidden cost of poor fit
When tools and people are poorly aligned, the cost goes beyond personal frustration.
These effects feed AI Value Leakage and reduce AI Net Efficiency. But the symptoms alone do not prove poor fit. Unreliable output, missing information, weak workflow design, thin guidance or the wrong choice of task can produce the same picture, as Briefings 05 and 06 showed. A credible diagnosis has to tell these apart.
§Why identical training is not enough
Most AI programmes teach people how to use the tools: prompting, document generation, research assistance, analysis, basic automation. Those skills matter. They do not address the different cognitive demands of working with AI. Someone who prefers structure can master prompting and still find open-ended interaction draining. Someone technically fluent can accept AI output too readily. A deep domain expert can struggle with tools that hide their reasoning or sources. More generic training rarely fixes any of these. This is the capability gap described in Briefing 04, The AI Perception Gap. Different needs call for different support.
| Interaction need | Support that helps |
|---|---|
| Prefers structure | Predefined workflows, task templates, explicit acceptance criteria |
| Works through exploration | Flexible tools for iterative reasoning, with quality controls in place |
| Needs stronger evidence | Accessible sources, transparent limitations, defined verification steps |
| Prefers limited AI involvement | AI on selected activities, without replacing established professional judgment |
Successful adoption does not need identical AI behavior across people. It needs reliable outcomes across the organization.
§One tool, three experiences
Take a hypothetical organization that introduces a generative AI assistant for business analysis. Three experienced analysts get the same platform and the same training. The first uses it to explore explanations for business performance, compare hypotheses and build recommendations through dialogue; the tool suits the way they work. The second prefers structured analysis with documented evidence; they find AI useful for organizing information and frustrating when outputs lack visible assumptions or sources. The third has deep domain knowledge and already works efficiently with established methods; for them, AI suggestions take more checking than they save.
| Analyst | Observed experience | Potential improvement |
|---|---|---|
| Exploratory | Uses AI well for hypothesis generation | Reliable source access and clear decision checkpoints |
| Structured | Finds unstructured output hard to evaluate | Templates, explicit evidence, defined review criteria |
| Expert-led | Little benefit, high verification effort | Narrower AI tasks with demonstrable gains |
Hypothetical example, not an empirical GrundMind case study.
A shallow reading calls the first analyst a successful adopter and sends the other two to more training. A better assessment asks whether the system and the tasks fit each analyst's needs, and may conclude that the organization needs different interaction modes, better evidence presentation, or more selective use of AI.
§Designing for cognitive diversity
People need different levels of structure, control, explanation and autonomy. We recommend six design principles.
Where practical, support structured workflows for defined tasks and flexible interaction for exploration.
Separate established facts, assumptions and open questions where the system can do so reliably. People should not have to infer uncertainty from polished prose.
People should know which decisions stay with humans and when AI can act on its own.
Brainstorming, analytical review, regulatory documentation and operational decisions need different levels of control and checking.
Ask whether AI reduces or raises the mental effort of doing the work reliably. Time saved alone does not describe the interaction.
Give people room to develop new strategies, give feedback and adjust how AI supports them. Preferences should inform design, not become labels.
§The GrundMind view
GrundMind assesses AI adoption across five connected dimensions. Cognitive fit is one of them, and it is read alongside the other four.
AI Fit Analysis examines whether AI tools and ways of working fit the organization's business needs, workflow requirements, human interaction patterns and governance expectations. Its cognitive dimension looks at how people experience uncertainty, control, trust and effort when working with AI, and interprets that against evidence from the other four. A report of high verification effort might reflect poor cognitive fit. It might equally reflect unreliable output, missing information or weak workflow controls. The diagnosis identifies the explanation the evidence best supports. The findings connect to AI Net Efficiency, which tests whether AI reduces total human effort, and AI Value Leakage, which locates where expected benefits are lost.
The archetypes give organizations a shared language for different ways of working with AI. Results are reported at team level. They are not used to rank individuals or inform employment decisions. The objective is organizational: AI-enabled work that supports different human needs while holding common standards of quality, accountability and performance.
§Five questions for the leadership team
- Are we assuming everyone should work with AI the same way?Uniform access does not require uniform interaction.
- Do we know why some people benefit more than others?Check experience, task, workflow conditions, tool reliability and fit before blaming motivation.
- Do our AI systems give the right structure, control and explanation?Interaction design should match the work and the people doing it.
- Are we measuring cognitive effort as well as time saved?AI can speed up output while making verification and interpretation harder.
- Can people adapt AI workflows without weakening governance or quality?Allow flexibility where it is safe, and keep accountability clear.
AI adoption should adapt to human cognition, not ignore it.
People do not experience AI identically. Trust, tolerance for ambiguity, need for control, ways of engaging with problems, expertise and the task itself all shape how they work with it. Programmes judged only on access, training and usage miss this, and pay for it in uneven gains, unnecessary cognitive effort and capabilities left unused.
Cognitive fit explains why an implementation that works in one context fails in another. It does not replace workflow analysis, value measurement or governance. It completes them.
The goal is not to make everyone think like an AI system. It is to design AI-supported work that makes good use of the different ways people think.
Book an AI adoption diagnostic pilot.
Understand how your people actually work with AI, where cognitive and operational friction sits, and how to design adoption around it.
About this briefing
The GrundMind Adoption Briefings are short, leadership-facing notes from the Grund Institute on the measurement and design of AI adoption in organizations. The series accompanies the longer-form Grund Institute white papers on cognitive fit, the Second Mind framework, and the GrundMind diagnostic.
Sources
- Brynjolfsson, E., Li, D. & Raymond, L., Generative AI at Work, Quarterly Journal of Economics, 2025.
- Grund Institute, Evaluating Human-AI Interaction Styles: Dimensions, Measurement and Framework, 2026.
- Grund Institute, Bridging the Cognitive Gap in AI Adoption: Second Mind Framework and Cognitive Fit, 2026.