A business can make almost every technical decision about AI correctly and still create surprisingly little value.
It can choose a capable model.
Connect it to the right information.
Integrate it with existing systems.
Redesign the surrounding workflow.
Put appropriate controls in place.
Deploy it successfully.
And then watch employees continue working almost exactly as they did before.
This is sometimes described as an adoption problem.
That’s true, but I think the word adoption makes the issue sound smaller than it really is.
It suggests the technology has been implemented and the remaining challenge is persuading people to use it.
But people aren’t sitting at the end of the AI value chain.
They’re part of it.
They decide when to use the technology, what to ask it, whether to trust the answer, whether to challenge it, how to incorporate it into their work, what to do with the time it saves, when human judgement still matters and whether the capability ultimately changes how the organisation operates.
This leads to a principle I think businesses need to take seriously:
Technology creates capability. People turn capability into organisational value.
Deployment is not adoption#
Software has always suffered from a simple problem.
Buying it is easier than changing behaviour.
AI doesn’t escape this.
An organisation might give 500 employees access to an AI assistant and announce that it has deployed AI across the business.
Technically, that’s true.
But what changed?
Perhaps some employees use it every day.
Some tried it once.
Some use public AI tools instead.
Some don’t know what they’re allowed to do with it.
Some think it’s useful for writing emails but little else.
Some don’t trust it.
Some are worried about what using it means for their job.
Some have discovered valuable ways of working that nobody else knows about.
The deployment statistic tells us almost nothing about the economic outcome.
Even usage statistics have limits.
Imagine 80% of employees use an AI tool every month.
That sounds impressive.
But use it for what?
Did the work improve?
Did customers receive a better service?
Did capacity increase?
Did decisions improve?
Did salespeople spend more time selling?
Did errors fall?
Did cycle times change?
Did margins improve?
Did the organisation become capable of doing something it couldn’t do before?
These are different questions.
AI adoption is an input. It isn’t a business outcome.
The behaviour has to change somewhere#
If AI is going to create value, something eventually has to become different.
Imagine an employee spends five hours each week preparing reports.
AI reduces that to two.
Three hours of capacity have theoretically been created.
But if the employee continues producing exactly the same output and nothing else changes, what did the business capture?
Perhaps there is a quality improvement.
Perhaps the employee feels less overloaded.
Both can matter.
But if the original business case assumed greater capacity, somebody has to decide what happens to those three hours.
Does the employee handle more work?
Spend more time with customers?
Analyse problems that previously went unexplored?
Support colleagues?
Sell?
Improve quality?
Reduce overtime?
Allow the business to grow without additional hiring?
The technology doesn’t make that decision.
The organisation does.
This is why productivity improvements can look enormous in demonstrations and much smaller in financial results.
Saving time creates an opportunity to capture value.
It doesn’t automatically capture it.
The employee’s behaviour, the surrounding workflow and management decisions determine what happens next.
Trust isn’t binary#
One of the most important influences on that behaviour is trust.
But businesses need to be careful when they talk about building trust in AI.
The objective shouldn’t be to make employees trust AI as much as possible.
Imagine an AI system that produces a useful recommendation 95% of the time.
One employee doesn’t trust it.
They manually reproduce all of its work before acting.
The AI might be useful, but much of the productivity benefit disappears.
Another employee trusts it completely.
They stop checking anything.
Eventually, one of the 5% of problematic recommendations matters.
Neither behaviour is particularly desirable.
What we need is something closer to calibrated trust.
Employees need enough confidence to use AI where it is useful, while understanding enough about its limitations to recognise when judgement, verification or escalation is required.
That balance will differ according to the work.
An AI-generated first draft of an internal email carries a different consequence from an AI recommendation affecting a major customer.
Summarising meeting notes isn’t the same as interpreting a contract.
Suggesting a marketing headline isn’t the same as approving a payment.
Trust should therefore be related to consequence.
The question isn’t:
Do our employees trust AI?
It’s:
Do they understand when they should trust it, when they should challenge it and what they remain responsible for?
Training has to go beyond teaching people how to prompt#
This has implications for AI training.
There has understandably been enormous interest in teaching employees how to use generative AI.
How to write better prompts.
How to structure requests.
How to iterate.
How to use different tools.
Those skills can be useful.
But they aren’t sufficient.
The more deeply AI becomes integrated into work, the less useful generic prompting training becomes on its own.
An employee needs to understand what AI means inside their role.
Which parts of their work can it help with?
Which information can they provide?
Which systems can it access?
Where is the output reliable enough to use directly?
Where does it need checking?
What remains their responsibility?
What happens when the AI is uncertain?
What should be escalated?
How does the organisation expect the released capacity to be used?
The relevant capability isn’t simply:
I know how to use ChatGPT or Copilot.
It’s:
I know how to work effectively with AI while doing my job.
That’s a much more organisation-specific skill.
Leaders and employees aren’t experiencing the same AI transition#
There is already evidence of a gap here.
Microsoft’s 2025 Work Trend Index found that leaders were ahead of employees across every measure it used to assess what it called an “agent boss mindset”.
Globally, 67% of leaders said they were familiar with AI agents, compared with 40% of employees.
Seventy-nine per cent of leaders believed AI would accelerate their careers, compared with 67% of employees.
At the same time, leaders expected significant changes in how people work. Forty-one per cent expected their teams to be training agents within five years and 36% expected them to be managing agents.
These are expectations rather than certainties about the future.
But the gap matters.
Senior leaders may spend their time discussing AI strategy, attending demonstrations, meeting technology providers and experimenting with frontier tools.
An employee somewhere else in the organisation may experience AI very differently.
They may hear that it will transform the business without understanding what that means for them.
They may be given a tool without enough time to learn it.
They may see colleagues experimenting while being uncertain about company policy.
Or they may reasonably wonder whether the productivity improvements leadership wants will eventually mean fewer people are required.
If those perspectives aren’t acknowledged, leaders can mistake their own enthusiasm for organisational readiness.
False reassurance won’t create trust#
This is particularly important when discussing jobs.
There is an understandable temptation for leaders introducing AI to reassure employees that nothing important is going to change.
But if the organisation simultaneously describes AI as transformational, that message can become difficult to believe.
If AI works, some work will change.
Tasks may disappear.
Other tasks may become more important.
Roles may broaden.
Responsibilities may move.
People may manage AI systems that didn’t previously exist.
Some expertise may become easier to access.
Expectations around productivity may increase.
The organisation may eventually need different skills.
In some situations, AI may affect future hiring or the number of people required to perform particular work.
Nobody can know precisely how all of that will develop.
Pretending otherwise isn’t reassuring.
It’s just uncertainty dressed up as certainty.
A more credible approach is to be clear about what the organisation knows, what it doesn’t know and what it is trying to achieve.
Employees don’t necessarily need a promise that their work will never change.
They need enough clarity to understand the direction of travel and how they can participate in it.
The people doing the work know things the implementation team doesn’t#
There is another reason employees matter.
They understand the work.
In June, I wrote about the importance of organisation-specific context.
Some of the most valuable context in a business doesn’t live in a database.
It lives in people.
The employee performing a process every day knows where it breaks.
They know which steps are pointless.
They know which information can’t be trusted.
They know which exceptions matter.
They know what customers actually ask for.
They know which approval usually takes too long.
They know which workaround everybody uses but nobody has documented.
They know when the official process and the real process are different.
That knowledge becomes extremely valuable when redesigning work around AI.
A technically elegant AI system designed without the people who understand the process may automate an imaginary version of how the organisation works.
Employee involvement therefore isn’t merely a way to make people feel included.
It can improve the implementation itself.
The people most affected by an AI workflow may also possess some of the knowledge required to make it work.
Resistance can contain useful information#
This also means businesses should be careful about labelling employees as resistant to change.
Sometimes people resist technology because they don’t understand it.
Sometimes because they don’t want their work to change.
But sometimes they’re telling you something important.
Perhaps the AI recommendation isn’t reliable enough.
Perhaps the new workflow creates extra work elsewhere.
Perhaps employees are duplicating the old process because the new system doesn’t handle important exceptions.
Perhaps the performance measure rewards behaviour that conflicts with the new workflow.
Perhaps the technology saves management time while making frontline work harder.
Perhaps the employee understands a risk the implementation team has missed.
Not every objection is correct.
But treating all resistance as a cultural problem can cause businesses to ignore useful evidence.
The objective isn’t to eliminate resistance.
It’s to understand it.
Some concerns require education.
Some require better communication.
Some require changes to the workflow.
Some reveal genuine flaws in the technology.
And some require management decisions that have been avoided.
Over-adoption can be a problem too#
There is an opposite risk.
An organisation can become so enthusiastic about AI that employees begin using it everywhere.
That can look like successful adoption.
It isn’t necessarily.
People may use AI for tasks where conventional software would be cheaper or more reliable.
They may introduce unnecessary complexity.
They may rely on generated answers where authoritative information already exists.
They may use tools with sensitive information inappropriately.
They may automate work that benefits from human interaction.
They may stop developing expertise they still need in order to judge the AI’s output.
This returns us to a theme that has run throughout this series.
The objective isn’t to maximise AI use.
It’s to maximise business value.
Sometimes those things align.
Sometimes they don’t.
A mature organisation shouldn’t celebrate an employee merely because they used AI.
It should care whether AI was an appropriate capability for the problem.
Incentives will shape adoption#
Behaviour also follows incentives.
Imagine a salesperson can use AI to prepare for meetings more effectively.
The company says it wants them to use the technology.
But management still measures them primarily on the volume of administrative activity completed in the CRM.
Or an AI system allows a customer-service employee to handle more complex cases.
But performance targets reward the number of tickets closed as quickly as possible.
Or employees are encouraged to experiment with AI while simultaneously being punished when experiments occasionally fail.
The formal AI strategy says one thing.
The operating environment says another.
Employees usually respond to the operating environment.
This is why AI adoption can’t be separated from management.
If leaders want behaviour to change, they need to examine the measures, incentives, responsibilities and expectations surrounding the behaviour.
Otherwise, AI can remain an additional tool sitting on top of an unchanged organisation.
Human judgement becomes more important in some places#
There is a popular assumption that as AI becomes more capable, human judgement becomes less important.
Across some tasks, that may be true.
But there is another possibility.
AI may allow organisations to remove human effort from areas where judgement creates little value and concentrate it where judgement matters most.
Last month I argued that the best AI workflow may not be the one with the least human involvement.
It may be the one that uses human judgement only where human judgement creates distinctive value.
That means employees need to understand those boundaries.
If AI gathers the information and prepares a recommendation, what exactly is the person reviewing?
If an agent performs routine actions automatically, what constitutes an exception?
If AI produces ten possible options, what knowledge should the human contribute when choosing between them?
If an employee remains accountable for a decision, what level of understanding do they need before approving the AI’s recommendation?
The role of the human can’t simply be:
Check the AI.
That’s too vague.
Human involvement has to be designed with the same care as machine involvement.
AI changes the management job too#
It’s easy to frame AI adoption as something employees need to do.
Managers have to change as well.
If employees can produce work more quickly, managers may need to rethink workloads.
If AI makes expertise more accessible, managers may need to reconsider how work is allocated.
If employees begin supervising agents, managers may need to evaluate a combination of human and machine output.
If routine work decreases, expectations around judgement and initiative may rise.
If information becomes easier to access, some management layers built around moving information may need to evolve.
And if AI changes what an individual employee can accomplish, managers will need to decide what good performance looks like in that new environment.
AI adoption therefore isn’t something management does to the workforce.
Management itself is part of what has to adapt.
Same AI. Different organisation.#
Imagine two competing businesses deploy exactly the same AI capability.
The first treats deployment as the finish line.
Employees receive licences.
Training explains the features.
Usage grows.
Management reports adoption.
The second starts with a different question:
What behaviour needs to change for this capability to create value?
It involves employees who understand the work.
It identifies where AI should and shouldn’t be trusted.
It redesigns responsibilities.
It makes clear where human judgement matters.
It teaches people how AI applies to their actual roles.
It listens when the workflow doesn’t work.
It changes measures and incentives where necessary.
And it decides what should happen to the capacity AI releases.
Same underlying technology.
Very different organisational capability.
This is why I don’t think people should be treated as the final stage of AI implementation.
They aren’t something to deal with once the technology has been deployed.
They are part of the system through which the technology produces an outcome.
AI can provide intelligence.
It can retrieve context.
It can accelerate tasks.
It can participate in workflows.
It can increasingly take actions.
But the organisation still has to decide how those capabilities change the way work happens.
And ultimately, organisations are collections of people making decisions and doing things differently.
The AI doesn’t have to adopt your organisation.
Your organisation has to adopt what the AI makes possible.
Technology creates capability. People turn capability into organisational value.