AI & Digital Change · The leadership behind transformation
AI adoption is a leadership challenge
By Isabella Brusati
An AI rollout needs a clear agreement about the work people should change, the judgement they must exercise and the responsibility they retain.
Imagine a team given access to an approved AI tool to prepare internal briefings. Some colleagues generate drafts immediately. Others hesitate because they do not know what information they may enter or how thoroughly they must check the result. The manager asks everyone to experiment more. Usage rises, but reviewers spend longer tracing claims and correcting confident errors.
This hypothetical rollout has created activity without a shared definition of useful work. Another demonstration of the tool may increase confidence while leaving the underlying responsibilities unclear.
My argument is that leaders should define the operating agreement around an AI use case before treating wider access as progress. That agreement needs a specific task, a quality standard, a human owner and a way to learn from the results. Technology specialists help establish the boundaries; managers make them usable in daily work.
Why access is an incomplete adoption measure
If leaders reward usage without defining an acceptable output, people may optimise for producing more drafts. Verification then falls to whoever receives the work. A local time saving can become another team’s checking burden, while uncertainty encourages other employees to avoid the tool altogether.
These are different responses to the same missing agreement. The practical task is to make expectations clear enough that both enthusiastic and cautious users can exercise informed judgement. The steps below are recommendations for managing adoption, not a claim that AI will improve every workflow.
1. Choose a task with an assessable result
At the start of a trial, ask the operational owner to describe the work in one sentence. “Prepare a first draft of an internal briefing from approved material, with material claims checked against their sources” is a testable task. “Use more AI” is not.
Agree what would count as improvement: less repetitive preparation, a more consistent result or more time for analysis. Record the current effort and quality for comparable work. Include preparation, checking, correction and handover so that the comparison covers a completed output.
Keep the first trial bounded enough to understand. If several teams use different tools for unrelated tasks, a single adoption number will reveal little about where the approach helps.
2. Give the human owner usable boundaries
Before people begin, agree who reviews the work, who may approve it and where an uncertain result goes. Involve the relevant technology, security and governance specialists in guidance on approved tools and appropriate information.
Then test whether the guidance answers a real working question. Can the reviewer access the original material? Do they have the expertise and time to check it? Does the user know when to stop and ask for help?
Make those responsibilities part of the workflow. An instruction to check everything is weak if deadlines leave no time for review. A manager may need to narrow the task or adjust the expected volume before the arrangement is credible.
3. Build learning around difficult examples
During the trial, bring a successful output and a plausible but flawed one to a short team review. Ask how each was checked, which assumptions needed attention and what the user would do differently next time.
Listen for the specific concern beneath reluctance. Someone may need practice assessing an output, clarification about accountability or an honest conversation about how their expertise will be valued. Those needs cannot all be met by another interface demonstration.
Be clear about what is known and what remains undecided about roles. Avoid promises of certainty the organisation cannot support. Explain how colleagues’ experience will inform the next decision.
Decide what the trial has earned
At an agreed review date, compare completed work with the starting point. Examine quality, total effort and the receiving team’s experience. Ask where errors escaped review and whether a new bottleneck has appeared.
Choose whether to expand, adjust or stop the use case. An unsuitable application is a useful finding. More licences or prompts cannot compensate for a task that remains unreliable or uneconomic after checking and correction.
Before the next expansion
Ask the use-case owner to bring one finished output and explain its purpose, review responsibility, full effort and remaining uncertainty. Resolve one missing condition before increasing the adoption target, then check its effect in the next operating cycle.
AI adoption becomes a leadership achievement when people can explain how the tool improves the work and when their judgement requires them to do something else.
An original leadership perspective. The briefing example is hypothetical. These recommendations complement the organisation’s specialist technology, security and governance advice.