AI adoption: Supporting middle managers to lead
Part 2 of a 5-part series: The human AI system.
In part one, I argued that AI adoption fails at the human layer rather than the technical one, and that the deepest fault line usually runs straight through the middle of the organisation. This piece goes into that fault line. The people we are counting on to drive adoption are, at the same moment, among the most threatened, the most depleted and the least supported. Until that is dealt with, no amount of tooling or training will make adoption stick.
An impossible brief we hand to the middle
Picture a capable team manager. She has been told to champion the new AI tools across her team. She is also, privately, unsure what half of them do. She has read the same headlines as everyone else about organisations flattening their management layers, and some quiet part of her wonders whether those headlines are about her. Her own organisation is going leaner. Nothing has come off her plate to make room for the new remit. And on Thursday morning she is expected to walk into her team meeting and project a confidence in her relationship with AI that she does not feel, in front of people who read her cues far more accurately than she realises.
That is an impossible brief for a middle manager to deliver upon, without sacrificing trust and integrity. We give the person in the middle the hardest brief in the organisation and the least room to deliver it, then treat it as a mystery when adoption stalls at exactly their level.
Three things that are true at once
Middle is where AI adoption is won or lost. Strategy gets decided at the top and delivered at the front line, which means it has to be translated somewhere in between. Managers are the layer that turns a slide about AI ambition into what people actually do on a Tuesday. When that translation does not happen, the strategy never reaches the ground, and the investment behind it quietly disappears.
The middle is the most exposed. Gartner projects that through 2026, one in five organisations will use AI to flatten their structures, eliminating more than half of their current middle-management positions, and firms including Oracle, Meta and Google have already cut heavily into this layer while increasing their AI spend (Dennison, 2026). So the people we are asking to champion the change are the people that change is most likely to remove.
Many managers are already running on empty. Shin and Sucher interviewed partners, managers and junior consultants at two large consulting firms and found that AI was elevating the work of junior and senior staff while overloading the layer in between. Managers were the ones validating AI outputs, catching what the authors call “workslop”, the polished-looking content that says very little, coaching their teams and holding the quality line – all under the same delivery pressure as before, and with almost no formal support (Shin & Sucher, 2026).
Salesforce’s survey of more than 500 managers points the same way. Nearly half (48%) feel pressure from leadership to demonstrate AI adoption, while only 32% work somewhere with a formal way of tracking it. Two-thirds are optimistic about AI. Half (51%) say they are anxious about keeping up with it themselves (Salesforce, 2026).
Put the three together and the picture is stark. We have built a layer that is pivotal, exposed and exhausted, and then asked it to lead the very change doing the exposing and the exhausting.
Why a threatened regulator cannot regulate
This is where the neuroscience earns its place, because it explains why the paradox is self-defeating and not simply unfair.
Emotion runs downhill in organisations, and it gathers speed on the way. Decades of research show that a leader’s affective state measurably shapes the mood, coordination and effort of the group around them. Barsade’s work on the “ripple effect” showed how one person’s emotion spreads through a team and changes how it behaves (Barsade, 2002). Sy and colleagues found that groups whose leader had been put into a negative mood took on that tone, coordinated less well, and had to expend more effort to get the same work done (Sy, Côté & Saavedra, 2005). We also watch the person in charge more closely than we watch each other, so whatever state they are in tends to travel further than they intend.
A manager who is quietly in a threat state, then, does not only feel the threat. She transmits it. The team picks up the physiology long before it processes the talking points, and no amount of upbeat messaging will override a nervous system that is signalling danger. You cannot co-regulate a team through a change you have not yet metabolised yourself.
Her own threat is not incidental to this, either. The three psychological needs that underpin human motivation – autonomy, competence and relatedness (Deci & Ryan, 2000) – come under pressure for her all at once:
- Autonomy: did I choose this, or was it handed to me?
- Competence: am I still good at the job I built my identity around?
- Relatedness: where do I belong if my whole layer is in question?
These are the same three needs Part One identified at the front line. Here they land on the person we have asked to reassure everybody else.
The loop this creates
Asked to model a confidence they do not feel, most managers do the reasonable thing and perform certainty. But performing certainty means hiding the learning curve, and a manager who cannot be seen to learn will not build a team where it feels safe to learn. The gap between the confident words and the guarded nervous system leaks. The team calibrates to the leak rather than the words. The effort to look adopted ends up suppressing the real adoption underneath.
Why the usual fix makes it worse
Most “manager enablement” for AI turns out to be a frontline programme with a leadership badge stapled to it: the same content everyone else receives, plus an instruction to go and cascade it. Three things commonly go wrong, and all of them are predictable:
- It adds responsibility without touching the threat, and without taking anything off the plate. An overloaded, exposed manager is handed one more thing to be accountable for.
- It asks managers to advocate for a message they have not yet believed themselves. What has not been metabolised does not get cascaded; it gets performed. And teams discount performance almost immediately.
- It rewards the appearance of adoption. Prize the managers who look adopted without ever checking underneath, and the real uncertainty goes underground – which is, more or less exactly, where unsanctioned AI use starts to breed. That is the subject of Part Three.
What actually helps: an activation pathway for the middle
A room of managers admitting the same uncertainty is among the cheapest and most powerful interventions available to you. It is also where a layered rollout and visible role-modelling do most of their groundwork.
The answer is not more exhortation. It is to give managers their own adoption journey before we ask them to lead anyone else’s. Here are five moves can make a disproportionately big difference .
-
Let managers be learners before you make them leaders. Ring-fence a period in which managers use the tools on their own real work, clumsily and out loud, with no expectation that they teach anyone yet. Adoption you can model beats adoption you can announce. This is acknowledgment and experimentation applied to the manager rather than to their team.
-
Treat self-regulation as a core AI-leadership skill. Managers need interoceptive and metacognitive tools to notice their own threat state and settle it before they walk into the room: the metacognitive pause, a breath before the translation, and affect labelling – the simple act of putting a feeling into words, which dampens the brain’s threat response (Torre & Lieberman, 2018). The instruction is short. Self-regulate, then translate. A settled manager is the most valuable adoption asset you have.
-
Restore the three needs deliberately. Autonomy: real agency over how AI lands in their team, not only over whether it does; Competence: protected, low-stakes practice time, because access is not the same as capability; Relatedness: peer cohorts, so that managers co-regulate with one another rather than performing alone.
-
Redistribute the load. If you add validation, coaching and quality-holding to the manager’s role, take something else out of it. Protecting manager capacity is not a wellbeing nicety bolted on at the end; it is one of the mechanisms by which adoption happens at all. An overloaded regulator cannot regulate, however willing she is. And capacity has to be a standing design choice, not a one-off gesture.
-
Change the job description in your own head. Stop asking managers to enforce adoption and start resourcing them to design the conditions for it – the people who make experimentation feel safe, rather than the people who police compliance. That is a trust move at heart. The manager’s job is to build the psychological safety in which learning can happen in public, and learning in public is what adoption actually looks like.
Measuring what matters
If your management team is the lever, measure the lever. Tool logins tell you who opened the door. They say nothing about whether a manager can lead their people through the change on the other side of it. The signals worth watching sit at the manager tier itself: their own confidence and regulation, the psychological safety they create, and their willingness to experiment where their team can see them. Team usage numbers are a lagging indicator of all three.
The image from Part One still holds. Middle managers are the load-bearing walls of the organisation, and you do not strengthen a load-bearing wall by leaning more weight against it and painting it to look confident. You strengthen it by taking the paradox seriously: by giving the middle its own path to adopt, its own tools to regulate, and its own permission to learn in public.
Do that, and the layer that is currently the bottleneck becomes the multiplier. Leave it as it is, and every adoption strategy above it will keep arriving at a wall.
Next, in part three of this series I’ll be exploring shadow AI and what it’s really telling you: why the AI use happening off the books is the most honest signal in your organisation, and what the squeeze on the middle has to do with it.
References
Barsade, S. G. (2002). The ripple effect: Emotional contagion and its influence on group behavior. Administrative Science Quarterly, 47(4), 644–675. https://doi.org/10.2307/3094912
Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268. https://doi.org/10.1207/S15327965PLI1104_01
Dennison, K. (2026, May 21). Why companies cutting middle managers to fund AI is a mistake. Forbes. https://www.forbes.com/sites/karadennison/2026/05/21/why-companies-cutting-middle-managers-to-fund-ai-is-a-mistake/
Salesforce. (2026). Middle managers aren’t obsolete. AI just made them more important. https://www.salesforce.com/news/stories/middle-managers-vital-age-of-ai/
Shin, J., & Sucher, S. J. (2026, June 26). AI adoption is overloading your middle managers. Harvard Business Review. https://hbr.org/2026/06/ai-adoption-is-overloading-your-middle-managers
Sy, T., Côté, S., & Saavedra, R. (2005). The contagious leader: Impact of the leader’s mood on the mood of group members, group affective tone, and group processes. Journal of Applied Psychology, 90(2), 295–305. https://doi.org/10.1037/0021-9010.90.2.295
Torre, J. B., & Lieberman, M. D. (2018). Putting feelings into words: Affect labeling as implicit emotion regulation. Emotion Review, 10(2), 116–124. https://doi.org/10.1177/1754073917742706*
