The Quiet Failure Inside Your AI Programme
Last week, I wrote about the widening disconnect between CEOs and boards on AI. This week, the evidence points to the next failure point: what happens when the story from the top meets the people expected to remake their work.
For two years, leaders have treated slow AI adoption as an employee problem. People are resistant. They lack the right skills. They are too anxious about change. There is some truth in all of that, but it is an incomplete diagnosis.
Harvard Business Review has just published research based on three years of work, 23 interviews, and three leadership workshops across 11 European IT-services firms. Its conclusion should stop a few executive teams in their tracks. When AI adoption stalls, the brake is frequently the leadership team itself.
Leadership drift is what happens when senior people speak enthusiastically about AI, issue broad encouragement to experiment, then leave the difficult questions unresolved.
Which decisions can an agent make?
Which roles will change first?
What is the threshold for human oversight?
How will performance be measured?
What happens to people whose work is genuinely reduced?
Employees can live with uncertainty. What they cannot work with for long is an organisation that asks for trust while refusing to be specific.
McKinsey’s new analysis makes the business consequence clear. Trust in the organisation and its leaders is one of the strongest predictors of employees’ readiness to use AI and of the enterprise value an organisation captures from it. Around one in five respondents across organisational levels say they feel anxious about AI-related workplace change. The figure rises to one in four among middle managers. Employees who report low trust in their organisation’s support during transformation are 1.5 times more likely to report anxiety.
That middle-manager figure matters. Senior leaders announce the transformation. The middle is where it becomes real. Managers determine if employees have time to learn, if new workflows are adopted, and whether early mistakes become learning opportunities or prompt a return to old practices. When managers are unclear, anxious, or privately sceptical, the strategy often fails during team meetings and performance reviews.
There is a temptation to smooth this over with reassurance:
“AI is here to help.”
“There are no current plans for redundancies.”
“We are all learning together.”
Employees hear the gaps between those statements. They know AI will change roles, progression paths, team structures, and the value placed on different kinds of expertise. A vague promise becomes a credibility problem the moment a role disappears or a team is restructured.
The more credible approach is harder, but far more effective. Be explicit about the work being redesigned. State what leaders know, what remains undecided, and how decisions will be made. Put real investment behind career pathways, internal moves, role-specific learning, and fair transitions. Then give managers the language, time, and authority to lead the conversations that follow.
For executive search and talent acquisition leaders, this changes the brief. It is no longer enough to appoint one senior person to “own AI”. Organisations need a leadership bench that can make difficult trade-offs in public, translate strategy into operating decisions, and keep capable people engaged while the ground shifts beneath them.
Three moves for the C-suite:
1) Treat the AI plan as a change contract. Publish the business outcomes, the workforce assumptions, the guardrails, the decision makers, and the next review date. A plan can evolve. Silence is interpreted as avoidance.
2) Make middle managers a funded workstream. Give them practical role redesign tools, protected learning time, forums to surface failure early, and access to senior decision makers. A manager cannot build confidence with a slide deck and no answers.
3)Track trust alongside adoption. Usage figures and productivity dashboards tell only part of the story. Track manager confidence, internal mobility, capability growth, and whether people believe the organisation is investing in a future that includes them.
Many AI programs do not fail because of the technology itself. Instead, the real problem is the gap between the future leaders talk about and the future employees think is being created for them.
Your Agent ‘Estate’ Has a Shadow Problem
Businesses are starting to talk about more than just AI copilots. Employees are now wiring together agents that can search, draft, trigger workflows, reach internal systems, and pass work to other agents. That creates real potential. It also creates an estate that many leadership teams cannot see clearly enough to manage.
BCG’s new work on an Enterprise AI Control Plane describes the pattern bluntly. Agents are spreading across platforms, business units, and use cases faster than governance frameworks. Each platform may have its own permissions, audit trail, spend controls, and local owner. Put enough of those together and the enterprise ends up with duplicated work, uneven policy enforcement, rising cyber exposure, and no central view of what actually exists.
The key word here is ‘estate’. An agent is not simply another piece of software on a licence register. It can take action in connected systems, call tools, access data, and be repurposed by another team. It has an identity, a scope of permission, an owner, a cost profile, and a potential failure mode. If any of those are unknown, governance is already behind the deployment.
The economics are changing as well. Gartner forecasts that worldwide spending on AI-optimised infrastructure will reach $42.3 billion in 2026, up 96% year on year. It expects inference, the cost of running models continuously in live systems, to overtake model training this year: $23.3 billion versus $19 billion.
That shift deserves more attention than it is getting. The cost of AI is shifting from experimentation into day-to-day operations. Every agent that runs continuously, checks a system, calls an API, generates a decision, or co-ordinates another agent creates an ongoing consumption cost. Boards that approved limited innovation budgets will soon face questions about operating models and cost structures.
For research and insights teams, the risk is especially acute. An agent that can interrogate a knowledge base, pull survey data, synthesise customer calls, create a brief, and distribute an output may be genuinely useful. But it also touches proprietary methodologies, client data, source material, and judgement built over years. The responsibility is to know precisely which systems it can reach, who is accountable for its outputs, how its findings are checked, and when the human researcher must intervene.
BCG’s answer is an enterprise AI control plane: a common layer for identity and authentication, an agent and tool registry, runtime policy enforcement, and ‘golden paths’ that bake governance into the quickest route to deployment. The practical point is important. Teams will work around governance that feels like paperwork. They will use governance that gives them a ready-made, compliant way to build. BCG says those pre-governed paths can reduce setup from weeks to a day.
Four questions should now be routine in every executive AI review:
1) What agents do we have, what can each one access, and who is accountable for it?
2) Where do our agents create value, and where are they simply creating activity and token spend?
3) Which decisions require a named human reviewer, particularly where client relationships, quality, fairness, or commercial judgement are involved?
4) Can a team build and launch a compliant agent faster than it can launch an ungoverned one?
If the answer to the fourth question is ‘no’, the shadow estate will keep growing. The control plane is not a technical nice-to-have. It is the management infrastructure that allows an organisation to scale agents without losing sight of risk, cost, or accountability.
This piece first appeared in Talent Pools Nexus, our weekly read for leaders in research, insights and consulting. Subscribe on LinkedIn.
