Two contract teams sit on the same floor of the same legal department. They got the same AI drafting tool on the same day, after the same all-hands where the general counsel said the same encouraging things. Six months later, one team runs nearly every first draft through it. The other barely opens it. The difference is not the lawyers, the contracts or the training. It is what happens at ten o'clock on Monday, when each team lead runs the weekly review.
One lead starts the meeting by asking what the tool's draft got wrong last week and what the team changed in response. The other lead never mentions it, keeps redlining by hand, and once remarked, lightly, that she prefers documents "written by someone who went to law school". Nobody issued a policy. Nobody needed to.
Strategy from above, use from one floor down
Most organisations talk about AI adoption as if it flows in two directions. From the top comes the strategy: the board wants a plan, the executive team announces priorities, somebody gets a new title with "AI" in it. From outside come the tools: vendors, pilots, licences, a demo that made the leadership team go quiet in a good way.
Both of those matter, and neither decides anything. The decision about whether a tool becomes part of how work gets done is made by the person who allocates the work, reviews it and signs it off. In most companies that is a middle manager. They are rarely in the strategy meeting and never in the vendor negotiation, and they hold a quiet veto over both.
What a team lead actually controls
On paper, middle managers have modest power. They cannot set budgets of any size, choose software or change company policy. In practice they control almost everything that determines whether a new tool survives its first month.
They decide what counts as done. If work is accepted when it is generated and checked, people will generate and check. If the manager sends back anything that looks machine-written, people will stop. They decide where time goes, including whether anyone is allowed a slower week while they learn something new. They decide what gets praised in front of the team and what gets a raised eyebrow. And they own the templates, checklists and recurring meetings that make up the real operating system of a team, the one that never appears in the process manual.
Consider a warehouse shift lead whose site has an agent drafting the handover notes between shifts. If she reads its draft aloud at the start of the next shift and corrects it in front of everyone, the team learns two things at once: the tool is allowed, and the tool is checked. If she brings her own handwritten notes and leaves the draft unread in her inbox, the team learns something else, and no training module will undo it.
Most of all, they decide whether using AI is smart or a bit of a cheat. Nobody writes that down. Everybody on the team knows the answer within a fortnight.
Why they hesitate, and why that is reasonable
It would be easy to cast the reluctant manager as the villain here. That would be both unkind and wrong. Middle managers have good reasons to be careful, and most of those reasons were handed to them by the people now asking them to move faster.
Start with the numbers they are judged on. A team learning a new way of working usually gets slower before it gets faster. The executive who announced the strategy sees the long-term curve. The manager sees this quarter's dip, on their own dashboard, under their own name. Asking them to absorb that cost without changing how they are measured is asking them to take a risk on someone else's behalf.
Then there is the uncomfortable part. A great deal of middle management consists of coordination: collecting updates, turning them into reports, chasing status, summarising for the floor above. Agents are rather good at exactly that work. A manager who enthusiastically rolls out a tool that writes the weekly status pack is, in a small way, automating a piece of their own job description. Some are secure enough to do it anyway. Expecting all of them to be is optimistic.
Finally, they carry the quality risk. When an AI-drafted clause causes a problem, the general counsel will not remember the all-hands. She will remember who signed off the contract.
Adoption spreads sideways, through people
Everett Rogers's work on the diffusion of innovations made a point that applies neatly here: new practices spread mainly through local opinion leaders, people others already watch and trust, rather than through the original source. In most teams, the local opinion leader is either the manager or someone the manager visibly backs.
Watch an engineering department for a few months and you can see it happen. One engineering manager starts pairing with an agent on code review and talks openly about where it saves time and where it invents nonsense. Within weeks, her team has a shared set of habits for using it. Two teams over, nobody has heard of those habits, even though they have the same licence and attended the same training. The tool spread exactly as far as one person's credibility carried it.
This is why a company-wide announcement so often produces patchy results. It addresses everyone and persuades the people who were already going to try. Everyone else waits to see what their own manager does, because that is the person who will be judging their work on Monday.
Talk to the people who actually decide
If middle managers are where adoption is won or lost, the sensible move is to stop treating them as a channel for the message and start treating them as the audience for it.
That means a few unglamorous things. Give managers an explicit allowance for the learning dip, agreed in advance with the person who reviews their numbers. Let them pick the first use cases for their own team, because they know which tasks are tedious and which ones carry hidden judgement. Change one or two of the measures they are held to, so that a working AI process counts in their favour, not just on a slide. And get the early-adopting managers in a room with the hesitant ones, without executives present, so the conversation can be candid about what went wrong.
It also means involving them before anything is bought. Deciding what to build or buy with the managers who will live with it, in a hands-on workshop rather than a steering committee, tends to surface the objections early, while they are still cheap to deal with.
Senior leaders can help by changing what they ask for. A licence-utilisation chart measures whether people opened the tool, which says more about curiosity than about adoption. Ask managers instead which piece of their team's work now happens differently, and what they stopped doing to make room for it. A manager who can answer in one sentence has adopted. One who reaches for the usage chart probably has not.
If you want to know how your AI strategy is really going, skip the adoption dashboard for a month. Ask to sit in on three ordinary team meetings, picked at random, and listen for whether anyone mentions the tools at all. The lead who opens with "what did the draft get wrong last week" is running your strategy. The one who never mentions it is running theirs.
