A contracts manager at a mid-sized software company spends a weekend playing with an AI agent and gets it to do a first-pass review of incoming NDAs. It flags the non-standard clauses, suggests the company's fallback wording and writes a two-line summary for the sales rep. It is about as good as the person who does that job now. That person sits two desks away. He has been there eleven years, he covered for her when her father was ill, and he brings in pastries on the last Friday of every month. On Monday she closes the laptop and says nothing.
Multiply that small silence across a company and you have a large part of the explanation for why AI adoption looks so busy and changes so little.
Automation stalls in the middle
Senior leaders are mostly keen on automation, at least in the abstract. They see a cost line and a competitor's press release. People at the edges of the organisation are often keen too, because the tools make their own day easier and nobody else's job is involved.
The trouble is in the middle, where someone can see a whole workflow and knows exactly who does each step. That is where the opportunities are clearest, and where they are least likely to be raised. The person best placed to spot that an agent could take over a colleague's work is, almost by definition, the person who knows that colleague. They've had lunch together. They know about the mortgage.
So the knowledge that would make automation useful sits in precisely the heads least willing to act on it. Leadership never hears about the NDA agent, because the only person who could tell them has decided it isn't worth being that person.
Offices run on unwritten rules
Every team has a set of rules nobody wrote down. You don't make a colleague look slow in front of the boss. You cover for someone having a bad week, and you expect the same in return. You don't go over a peer's head, and you don't volunteer someone else's work for cuts.
These rules exist for good reasons. They are a large part of why teams function at all. They create trust, and trust is what lets people admit mistakes and ask for help. A team where everyone is quietly auditing everyone else's role for automation potential is a team nobody wants to join.
Factory floors had a word for the worker who produced well above the agreed pace: rate-buster. It was not a compliment. The norm protected everyone from a management that would raise the target as soon as anyone proved it could be hit. The office version is subtler, but the logic holds. Proposing an agent that does a colleague's job is the knowledge-worker equivalent of busting the rate, and everyone understands the social price.
Guilt does more work than fear
Most writing about resistance to AI focuses on the fear of being replaced. That fear is real. But a different emotion does more of the blocking: the reluctance to be the one who did the replacing.
Fear of replacement at least has outlets. People can retrain, get ahead of the change, or decide the tool makes them more valuable. Guilt has no such release. If you propose automating a colleague's work and it goes badly for them, you will see the consequences every day, in the kitchen, at the leaving drinks, on their profile when they update it. If it goes well for the company, the credit is diffuse and mostly flows upward.
This is a textbook principal–agent problem with a social twist. The company captures the benefit. The individual who raises the idea carries a private cost, paid in awkward lunches. Any rational employee in that position keeps quiet, and most employees are rational. Then leadership wonders why the internal "AI ideas" channel is full of suggestions about meeting summaries and contains nothing about the work that actually costs money.
What the stall looks like from outside
You rarely see an outright refusal. The resistance is polite, and it tends to take a few recognisable forms:
None of these show up on a dashboard. The licences are used. The workshops were well attended. The work that would move a number is left exactly where it was, protected by people who like each other.
Take the decision off one person's shoulders
If the obstacle is that nobody wants to be the person, the fix is to make sure no single person has to be. That is a management job, and it has to happen before anyone is asked for ideas.
Separate the question of which tasks can be automated from the question of what happens to people. Mixing the two is what creates the guilt. If a team knows, in writing and in advance, that time freed by automation goes to a named backlog of work they've been putting off for a year, the conversation changes. It is no longer "should we get rid of his job" but "which of his tasks should he never have to do again".
Let people automate their own work first. A support specialist who builds the agent that drafts her own replies is threatening nobody. She is also its best possible designer, because she knows exactly where the standard answers fall short. The IKEA effect works in your favour here: people value what they help build, and they defend it.
Talk about tasks rather than roles, and mean it. The moment one automation project turns out to have been a headcount reduction in disguise, every future idea in that department will be withheld. People who watched it happen will remember the colleague who stayed quiet and conclude that she was right.
Managers carry their own version of the problem. A team lead who automates a chunk of her team's work may end up with a smaller team, and in most companies headcount is still how a manager's importance gets measured. Asking her to propose it is asking her to argue for her own shrinking. Change what you reward, or expect her ideas to stay modest.
When a role genuinely will change, say so plainly and early, from the top. That burden shouldn't land on a peer two desks away. Leadership collects the savings, so leadership should carry the conversation.
The colleague belongs in the design
Go back to the NDA agent. One possible outcome is that the contracts manager overcomes her scruples and pitches it over her colleague's head. Plenty of AI programmes quietly hope for that version. It costs a working friendship and teaches everyone watching to keep their next idea to themselves.
The better outcome is that her colleague gets the agent first. He uses it to clear the routine NDAs before lunch and spends the afternoon on the awkward negotiations he's actually good at, the ones where eleven years of knowing the sales team counts for something. She gets credit for the idea without having to be the person.
That version needs someone above them both to say, before any tool is shown, what freed-up time is for. Say it first and the real ideas start arriving. Skip it and you'll keep getting meeting summaries.
