What drew you to behavioural science and cognitive neuroscience, and how has your perspective evolved over the years?
My interest began with a personal question: why do intelligent, well-intentioned people make decisions that harm their own interests? Growing up, I heard the familiar advice to trust your gut. I also saw how intuition sometimes led people close to me into unnecessary conflict. Later, coming of age during the dot-com boom and bust, I watched apparently sophisticated decision-makers make enormous mistakes. I wanted to understand why such errors were predictable and how people could avoid them.
That led me into the history of behavioural science, behavioural economics, and research on the cognitive neuroscience of decision-making. I earned my PhD in the History of Behavioural Science at UNC-Chapel Hill and subsequently taught and researched decision-making, cognitive biases, and risk management at Ohio State.
The underlying principles still matter enormously. Loss aversion, confirmation bias, overconfidence, and the tendency to follow familiar habits show up repeatedly in leadership decisions and organisational change.
What has expanded in my perspective is the emphasis on systems. You can teach an individual to recognise a bias, but that lesson may have little effect if organisational incentives, managers, and workplace norms continue to reward the old behaviour. My work on hybrid work and now AI adoption has strengthened that conviction. If you want lasting change, you must address the individual's psychology and the surrounding conditions that shape behavior.
Why is it so hard for organisational learning to produce behaviour change and measurable impact?
Training often solves the wrong problem. Someone attends a workshop, understands the material, and even performs well in an exercise. Then they return to a workplace where the manager expects the old process, the necessary tools are unavailable, deadlines discourage experimentation, or there is no reward for applying what they learned. Knowledge improved, but the conditions governing behaviour stayed the same.
There is also a motivational problem that trainers frequently underestimate. In AI adoption, for example, employees may know how to use a tool but avoid it because they fear it will eliminate their job, undermine the skills that define them professionally, or make colleagues think they are cheating. Another prompting course cannot resolve those concerns. You have to diagnose why someone would want to use the skill in the first place.
I recommend starting a programme by specifying the observable behaviour you want people to perform at work. Then investigate what currently prevents that behaviour, including skills, motivation, incentives, access, workflow design, social norms, and management support. Build the training around real tasks and follow it with opportunities to practice, feedback, peer support, and reinforcement from the person's manager.
The transition from learning to behaviour change happens after the workshop as much as during it. Trainers can teach and demonstrate a technique. The organisation must make using that technique practical, worthwhile, and safe.
How can trainers become more aware of their own cognitive biases and avoid assumptions about learners?
The most useful intervention is to stop relying on self-awareness alone. Knowing that confirmation bias exists does not make someone immune to it. Trainers need processes that reveal when their assumptions are wrong.
The curse of knowledge is a common trap. Once you know a subject well, you underestimate how difficult it is for a newcomer. False consensus makes you assume learners share your priorities and motivation. Confirmation bias encourages you to remember the enthusiastic participants and discount the resistant ones. The planning fallacy makes you pack more content into a session than learners can realistically absorb and apply.
I would start by interviewing a representative sample of learners before designing the programme, including those who are skeptical or struggling, rather than only the enthusiastic volunteers. Observe the actual work if possible. Ask what makes the target behaviour difficult, what employees already know, what they have tried, and what they fear might happen if they change their approach.
Then test assumptions explicitly. Write down what you expect learners to find useful, where you expect them to struggle, and what evidence would prove your design wrong. Pilot the material with people who differ in experience, seniority, and attitudes. Ask them to demonstrate the skill, rather than merely say whether the session was helpful.
Finally, examine the failures and unexpected reactions with as much interest as the positive feedback. A trainer who learns that a popular exercise did not change anyone's work has obtained valuable evidence.
How should training providers demonstrate impact and return on investment?
Start with a business outcome, work backward to the behaviour that influences it, and then identify the skills that enable that behaviour. If a leadership programme promises better management, define what that means operationally. It might involve more effective one-on-ones, faster resolution of team problems, lower avoidable turnover, or better delegation. A satisfaction score from the workshop will not establish whether any of those things occurred.
Record a baseline before training. Measure competence through realistic tasks where feasible, then check whether participants apply the target behaviour 30, 60, and 90 days later. Compare relevant quality, speed, error, customer, engagement, or retention indicators with the baseline. If practical, use a comparable group that has not yet received the intervention, while recognising that organisations rarely offer the clean conditions of a laboratory experiment.
For an AI programme, I would distinguish attendance, access, usage, quality of use, and business results. Logging into a tool tells us little about whether the work improved. A better measure might be how long a recurring process takes, whether error and rework rates fall, how many successful workflows employees share, and whether the organisation converts saved time into additional output or better service.
ROI requires comparing reasonably attributable financial benefits with the full costs, including training, technology, employee time, support, and quality assurance. Some outcomes, especially in wellbeing and leadership development, are valuable without being easily converted into pounds or dollars. Report those honestly rather than assigning speculative monetary values. Above all, separate measured change from a claim that training alone caused it.
What consulting case demonstrates training that actually changed behaviour?
An anonymised regional accounting firm with about 200 employees offers a particularly useful example from my book. Its leaders initially promoted generative AI through speeches and broad encouragement, but experienced accountants remained hesitant. They had valid concerns about reliability, confidentiality, and the value of their established professional methods.
We worked with department leaders and the learning and development team to build short, role-specific modules around real accounting work. Tax specialists practiced summarising regulatory changes and checking the source material. Auditors worked on analysing transaction records and flagging anomalies within approved, secure tools. Advisory staff experimented with first-draft scenario analysis. We paired the modules with supervised practice, peer mentors, and feedback about inaccurate AI outputs.
One influential turning point came when the head of audit, initially skeptical, attended a workshop and found that the approach saved him an afternoon of spreadsheet work. His public endorsement helped colleagues take the training seriously. The firm also began recognising employees for demonstrated applications at work, rather than course completion alone.
Over nine months, the firm reported wider day-to-day adoption and a 25% reduction in auditors' time spent searching transactions for irregularities. By the end of the second year, more than 80% of employees reported in an internal survey that AI skill-building had meaningfully improved their productivity; the firm also reported a 14% improvement in morale. Those figures combine operational reporting and employee perceptions, rather than results from a randomised trial, so I would not attribute every change exclusively to training.
What matters about the example is the mechanism. The programme connected learning to work people actually performed, gave them permission and support to experiment, established quality checks, and reinforced useful behaviour through managers and peers. That is how a training event became an ongoing change in how people worked.
What role should AI play in learning design and training?
I see substantial value in using AI to make learning more relevant, interactive, and continuous. A trainer can use it to develop draft scenarios tailored to a specific role, adapt explanations for different levels of expertise, simulate a difficult management conversation, give provisional feedback on practice exercises, or build an on-demand assistant that helps learners apply a method in the middle of actual work. The trainer still needs to verify the material and decide what good performance looks like.
The larger opportunity is to connect learning to workflow redesign. In conventional software rollouts, employees often must learn a standardised interface. Generative AI is far more adaptable to individual work. People can describe a problem in ordinary language, refine a prompt, build a reusable assistant, and collaborate with colleagues to improve it. That calls for learning programmes that teach experimentation, verification, and sharing rather than stopping at tool demonstrations.
My book argues that adoption also depends on psychological readiness. I use three profiles to diagnose recurring barriers. AI Alarmists fear threats to their livelihood and need credible answers about economic security and the value of developing AI skills. Pragmatic Resistors worry that AI diminishes their professional or creative identity, so I begin with tedious tasks that get in the way of the work they value. Reluctant Adopters are already using AI but conceal it because they fear being judged as less capable. They need clear norms, leadership modelling, and recognition for responsible use.
A training programme that overlooks those distinctions can create more technically competent people who remain unwilling to apply or share what they know.
Is AI encouraging people with shallow subject expertise to produce training they are unqualified to deliver?
Yes, and I think this is a serious risk. Generative AI can produce a polished-looking syllabus, slide deck, case study, or assessment remarkably quickly. Fluency and presentation quality can conceal weak underlying expertise. Someone who has never managed a team, treated a clinical condition, or handled a sensitive workplace investigation can generate material that sounds as if it came from extensive professional experience.
The solution is to distinguish content production from subject-matter competence. AI can reduce the time required to draft and organise materials, but a qualified professional must still establish the learning objectives, check the evidence, assess where advice might fail, and decide whether the exercises reflect genuine workplace conditions. Trainers also need to be candid about their own experience. An AI-generated fictional case should never be presented as a case the trainer personally handled.
I would require an expert review for any programme dealing with consequential employment, legal, safety, or wellbeing decisions. Test the training with realistic scenarios, including cases where the easy answer is wrong.
Evaluate whether learners can apply it correctly, not whether the material looks credible. If a provider lacks the relevant expertise, the right response is to partner with someone who has it, narrow the scope, or decline the work. AI can make good practitioners more productive; it does not confer professional qualifications.
How should a training company build an AI tool while minimising hallucinations, false claims of experience, and other risks?
Begin with a tightly defined purpose. A leadership development company might build an assistant that helps managers prepare for feedback conversations using its own approved training framework. A wellbeing provider might build a tool that explains its published educational materials and helps users find the right resources. Those are much safer starting points than an open-ended bot that gives unrestricted management or mental health advice.
Next, assemble a small, vetted knowledge base: the provider's validated materials, policies, licensed content, and other authoritative sources. Configure the tool to ground factual answers in those sources, show where relevant information came from, and say when the material does not support an answer. The tool must never claim personal experience, invented qualifications, or a client story that nobody actually experienced. A clear instruction helps, but instructions alone do not make those rules foolproof.
Test the tool before release with ordinary questions, ambiguous requests, missing-information scenarios, misleading prompts, and deliberately difficult examples. Keep records of where it invents facts, overstates certainty, gives inappropriate advice, or ignores limits. Have domain experts evaluate those failures and revise the source material, workflow, and safeguards. Repeat testing after substantial model or content changes.
Set boundaries according to the consequences of an error. For a low-risk brainstorming tool, human review before using the output may be sufficient. For employment decisions, legal compliance, or sensitive wellbeing questions, require qualified human oversight and clear escalation procedures. A wellbeing assistant should not diagnose conditions, present itself as a therapist, or improvise crisis advice. Protect employee and client information, restrict access to approved systems, and avoid uploading confidential information without appropriate authorisation.
No design can guarantee that a generative model will never hallucinate. The goal is to reduce the likelihood of error, make unsupported answers visible, and prevent an unchecked answer from causing harm.
NIST's Generative AI Profile within its AI Risk Management Framework is a useful reference for building such controls. I'd also make the design a collaborative effort between the provider's subject experts, intended users, and whoever is responsible for security and privacy.
Which learning and training design trends do I expect to grow over the next few years?
I expect the centre of gravity to move away from one-off content delivery and toward continuous performance support. Organisations will increasingly expect employees to practice realistic tasks with AI-supported simulations, receive feedback, and access approved guidance while doing the work. The relevant question will become whether the learner performs better a month later, rather than whether they completed a course.
I also expect more employees and frontline managers to help build and refine their own AI learning assistants, prompts, and workflows using no-code tools. That can make learning more relevant, but only if organisations maintain standards for evidence, privacy, quality, and human review. We will need to teach people how to question AI output, verify claims, recognise uncertainty, and exercise judgment instead of treating fluent answers as proof of competence.
Beyond AI, I expect greater recognition that managers, peers, and workplace culture determine whether learning lasts. Psychological safety, time to practice, and incentives aligned with the new behaviour will remain decisive. Technology may make practice easier to deliver. It cannot make people care about learning or give them permission to change how work gets done.
How do I keep learning and improving my own work?
I try to combine research with direct exposure to how people actually behave. Academic findings help identify patterns, but they are only a starting point. Conversations with executives, middle managers, and frontline employees reveal what happens when those patterns encounter deadlines, conflicting incentives, organisational politics, and genuine uncertainty.
That is also how I approached The Psychology of AI Adoption at Work. It draws on behavioural science, more than 100 consulting projects, and over 50 interviews with business leaders and practitioners. I learn from what succeeds in an implementation, but I pay particular attention to where a theoretically sound approach fails to change behaviour. Those cases tell me which assumptions or organisational constraints I have missed.
Writing and teaching are useful disciplines because they force me to explain ideas clearly enough for other people to test them. I want the frameworks I use to become more specific and actionable as the evidence changes. Given the pace of AI development, the willingness to revise a recommendation is as important as familiarity with the newest tool.