Working with AI agents: what it does to your people, and how to get it right

Microsoft's 2026 Work Trend Index, built on a survey of 20,000 AI users across ten countries, found that teams getting real value from agents do a few specific things differently: they work out together where an agent fits (63% against 32% of other teams), they share tips and mistakes openly (61% against 36%), and they actually talk about what good AI-assisted work looks like (54% against 29%). None of that is about the technology. Most organisations skip it anyway, buying the agents, activating the licences, ticking the training box, and calling it adoption, which is what I call Checkbox Adoption: everything deployed, nothing actually changed in how people work. There's a second shift underneath that's easy to miss, too. An agent that assigns work, flags exceptions, or decides what gets escalated is managing your team, whether you call it that or not, and research in the Academy of Management Annals shows people sense being managed by something opaque that nobody answers for, and trust drains fast when they do.

The evidence for what works is consistent, and it's simpler than most governance frameworks make it sound: let the agent handle the repeatable, well-defined work, and keep the judgement with your people. Jia and colleagues tested this directly in a 2024 Academy of Management Journal field experiment at a telemarketing company, where AI took the routine part of the job, finding and qualifying leads, and employees kept the human part, persuading the customer. Creativity went up, sales went up with it. A 2026 Scientific Reports study led by Wakslak sharpens the mechanism: people who drafted first and used AI to refine kept their confidence, ownership, and sense of meaning intact, with quality matching unaided work, while people who simply copied AI output lost all three. The order matters. Think first, then bring in the agent.

Get the order wrong and the cost is real, not just theoretical. Wu and colleagues, across four experiments with more than 3,500 people published in Scientific Reports in 2025, found that working with generative AI improved immediate output, but the improvement didn't carry over when people later worked alone, and they came out of it less motivated and more bored. The Wakslak study found the same shape, and the dip in confidence outlasted the AI use itself. Worse, the Jia experiment found the gains were skill-biased: higher-skilled employees got more creative and more energised, lower-skilled employees improved only a little and felt worse. Drop agents into a team without support and you can widen the gap between your strongest and weakest people while telling yourself you've levelled the playing field, and when that gap goes unspoken, people don't raise their hand, they go quiet. That's the same Silent Resistance pattern I trace across every AI rollout: the most reliable sign the collaboration is failing while every adoption metric says it's working.

One framing choice does more quiet damage than almost anything else: treating the agent as an employee, with a name, a job title, a slot on the org chart. A randomised experiment published in Harvard Business Review in May 2026 gave 1,261 managers in HR and finance identical flawed documents to review, varying only who supposedly wrote them. Among managers whose organisations already list agents on the org chart, the personal responsibility they took for the work fell by 9 percentage points and the blame they pinned on the AI rose by 8. Managers reviewing an 'AI employee's' work caught 18% fewer errors than those reviewing identical output attributed to an AI tool, and requests to escalate work for another review rose 44%, often replacing the reviewer's own checking rather than adding to it. The framing didn't even buy what leaders hoped for: it did nothing for adoption, and it made managers 13% more likely to feel unsure about their own professional identity. Be honest about what an agent is: software that needs a human accountable for it, which demands more disciplined governance than a human colleague would, not less.

None of this shows up in a licence report. It's the psychological and cultural readiness underneath the rollout, and it comes down to a short list of design choices: make people draft before they delegate, say out loud what the human is for and who owns the agent's output, make it safe to say 'I don't understand this,' and watch who's gone quiet, because agents don't lift everyone equally and rising usage doesn't tell you whether people are learning or hiding. The organisations getting this right aren't the ones with the best agents. They're the ones that treated the human side of the collaboration with the same rigour as the technical rollout, and that's still a design choice, not a technology one.