Why Your Most Experienced People Resist AI Hardest (and Why You Need Them Most)

By Victor Hoang · Updated Oct 11, 2026

Almost every leader who has rolled out AI has met the same person. She's your best account executive, your most trusted controller, your senior engineer who has seen every outage. She's polite in the training session, nods at the demo, and then quietly goes on working the way she always has. When you ask how the tool is going, she says it's "interesting" and "not quite there yet."

Don't read that as stubbornness. And don't plan around her as a late adopter. Our position: senior experts resist AI because it threatens the scarcity of their knowledge, yet they're the people best placed to catch its errors. So an adoption plan should make them the reviewers and designers of AI work, not the last people to get a license.

This piece is about the expert's role in adoption. For how to talk to a whole team about job security, see how leaders should respond to AI job anxiety. For how to price seniority, see whether experience is still worth paying for. Here the question is narrower and more operational: how do you get your deepest expertise working with AI rather than around it?

Key Facts

• Harvard Business School researchers Das Narayandas and Shunyuan Zhang argue that employees often resist AI because it threatens their identity and expertise, not because it performs poorly, and that the most common response is "symbolic adoption": complying on the surface while quietly undermining the tool (HBS Working Knowledge, June 2026).
• In a randomized trial, 16 experienced open-source developers took 19% longer with early-2025 AI tools on 246 tasks, after predicting AI would make them 24% faster and later estimating it had made them 20% faster (METR).
• In a field experiment with 181 professional recruiters reviewing nearly 8,000 resumes, those given higher-quality AI were less accurate than those given lower-quality AI, and the effect was driven by the more experienced recruiters (Dell'Acqua, Harvard Business School).
• On a task just outside the AI's capability, consultants were correct about 84.5% of the time without AI, versus 60% and 70% in the two AI groups (Dell'Acqua and colleagues, 2023).
• Gallup's late-2025 survey found that 69% of leaders, 55% of managers and 40% of individual contributors used AI at work at least a few times a year.

Why the Resistance Is Rational

Start by taking the resistance seriously, because it's more rational than it looks.

Expertise is a scarcity asset. An experienced person is valuable because the thing they know is hard to get. It took ten years of cases, mistakes and repetition, and nobody else in the room has it. That scarcity is what earns them the good projects, the senior title, the deference in meetings, and the salary.

Generative AI attacks the scarcity directly. It can produce a competent answer on demand, in seconds, to anyone. The answer isn't always right, but it's often good enough to make the novice look closer to the veteran than the veteran is comfortable with. So when your senior person hesitates, she's often making a correct reading of her own position. The tool doesn't threaten her job tomorrow. It threatens the premium that makes her knowledge worth something.

The Harvard Business School research gives this a useful structure. Narayandas and Zhang describe AI as a "self-disruptive" technology and name three identity threats: role compression (judgment-based work gets automated, leaving fewer and lower-status tasks), control shift (algorithms take over decisions, so people lose their sense of authorship), and span erosion (managers lose influence over people, budgets and processes). Their warning for leaders is that people facing these threats have three options. They can refuse, they can adopt fully and risk losing their identity, or they can comply while quietly undermining the tool, which the researchers describe as the most common response.

That third option is the one that should worry you. Open refusal is visible and you can deal with it. Symbolic adoption looks like success in your adoption dashboard.

One caution on the evidence. This isn't a simple story about age or tenure. In Gallup's survey, leaders and managers are heavier AI users than individual contributors. And in a survey study of 164 medical students and 42 experienced physicians, the students reported stronger identity threats and resistance than the experienced physicians (Jussupow, Spohrer and Heinzl, 2022). The resistance we're describing isn't "old people dislike technology." It's specific: people whose standing rests on a particular body of hard-won knowledge react when a tool seems to give that knowledge away. That can be a 29-year-old specialist as easily as a 55-year-old director.

Why You Need Them Anyway

If the story ended there, the answer would be to work around the experts. It doesn't, because they hold the one thing the tools lack: a map of where the answers go wrong.

The clearest evidence comes from the "jagged frontier" experiment run with 758 consultants. Researchers built one task to sit just outside what the AI could do well. Consultants without AI got it right about 84.5% of the time. With AI, accuracy fell to 60% and 70%. The AI's answer was fluent and plausible, and many people accepted it. Someone had to know where the frontier lay, and the people who know that are the ones with enough cases behind them to feel when an answer is off.

But here is where I want to be honest, because the evidence also cuts against a lazy version of this argument. Experts aren't immune to AI's failure modes. Two findings should shape how you use them.

First, the METR trial. Sixteen experienced developers, with an average of five years on the mature projects they were working in, completed 246 real tasks. When AI was allowed, they took 19% longer. They had expected a 24% speedup, and even after the slowdown they believed AI had helped by 20%. Experienced people were confidently wrong about their own productivity. That's a sample of sixteen on one kind of work, so treat it as a caution rather than a law.

Second, the recruiter experiment. Dell'Acqua's formal model predicts that as AI gets better, humans have less reason to stay attentive, so the AI substitutes for their effort instead of adding to it. The field experiment bore this out: recruiters with better AI did worse, and the effect was concentrated among experienced recruiters. They stopped checking.

Put the two together and the lesson is not "experts will save you from AI errors." It's sharper. Experts are the only people who can recognize an error, but they will only catch it if the process makes catching it their job. An expert handed a polished tool and told to go be productive will either ignore it or fall asleep at the wheel. An expert handed a defined review role, with failure cases to look for, does what you need.

The Reviewer, Designer, Teacher Model

This is the framework I'd use. Instead of asking where each senior person falls on an adoption curve (early adopter, laggard), ask what role their expertise plays in the AI workflow. There are three, and each answers one of the identity threats above.

Reviewer. The expert samples and approves AI output where errors are costly. This answers control shift: the person isn't being replaced by the algorithm, they're the authority it answers to. It requires an explicit edge map (the known failure cases) so that review is active rather than a rubber stamp.

Designer. The expert shapes how the work flows: what the AI drafts, what a human decides, what the thresholds are, what the prompts and examples look like. This answers role compression and span erosion: the person's judgment is now encoded in a system that many others use, which expands their influence instead of shrinking it.

Teacher. The expert turns tacit knowledge into examples, checklists and coaching for people and for the tools. This answers the scarcity problem head on. Their knowledge stops being a private asset that depreciates and becomes a recognized contribution to the firm. We cover why this matters for the next generation in the broken bottom rung and keeping human judgment sharp.

Matching the role to the expert

The last row matters. Not every senior person's value sits in judgment. Some have been paid mostly for doing standard work well, and AI exposes that. Don't hide it from them. Give a straight answer and a real path, because the alternative is a gifted person who quietly sabotages the rollout.

Where Most Adoption Plans Go Wrong

Three patterns come up again and again.

Training everyone the same way. A generic prompt-writing workshop lands badly with someone who has twenty years of context. It implies the problem is that they don't know how to use a chatbot. The problem is whether they trust it and what it does to their standing. Run a different track: scenario-based sessions where experts try to break the tool on their own hardest cases. They'll enjoy it, and what they find becomes your edge map.

Measuring adoption by usage. Logins and prompts per week reward symbolic adoption and penalize the careful expert who uses the tool rarely and well. Measure outcomes instead: errors caught before they reached a customer, cycle time on redesigned workflows, quality at the exception. Our piece on how performance reviews change with AI goes further on this.

Letting the loudest enthusiasts define the program. Peer champions are valuable, and peer-led AI programs can spread skills quickly. But if the program is built only by enthusiasts, the experts see a project that wasn't designed with their judgment in it, and they stay out. Put a skeptical senior person in the design group early. Their objections are free quality assurance.

The same logic applies at the top. Leaders who use AI themselves are far better placed to ask experts credible questions, a theme we take up in whether CEOs should use AI themselves. And the managers in between decide whether an expert's role gets protected or squeezed, which is why middle management is both the obstacle and the opportunity.

What to Do Next

Here's a sequence that fits inside a quarter.

1. List your ten most expert-dependent workflows. These are the places where one person's judgment prevents costly errors. Name the expert for each.
2. Diagnose the resistance before the remedy. For each expert, ask which threat they feel: role compression, control shift or span erosion. A short, honest one-on-one beats a survey. Listen for the reasoning, not the sentiment.
3. Assign a role, in writing. Use the table above. Give each expert one of reviewer, designer or teacher, with a defined scope and a named workflow.
4. Build the edge map together. For each AI-assisted workflow, have the expert document the failure cases, the checks, and what must never be left to the tool. Review the map monthly.
5. Design the review so it can't be skipped. Use sampling, spot audits and a requirement to record why output was accepted or rejected. The recruiter study shows attention drifts when the tool seems reliable, so don't depend on goodwill.
6. Change what you recognize. Credit experts in reviews and promotions for the errors they catch, the workflows they redesign and the people they teach. If the incentive system still rewards private knowledge, you'll get private knowledge.
7. Tell the truth about the trade. The firm is not asking them to give their expertise away for nothing. It's asking them to turn it into something that compounds, and it's committing to treat that as the most senior contribution there is. The HBS researchers' broader point, that adoption depends on whether people see how it benefits them, is the standard to hold yourself to.

A last test. Ask your most respected expert in any function what the AI gets wrong in her field. If she answers instantly and in detail, she's engaged and you should put her in charge of the review. If she shrugs, you haven't yet given her a reason to care. Either way, she's the person you need to hear from before you scale anything.

Learn More

• How Leaders Should Talk to Their Team About AI and Job Security
• Is Experience Still Worth Paying For When AI Narrows the Skill Gap?
• Peer-Led AI Programs
• Middle Management: AI Obstacle or Opportunity
• The Broken Bottom Rung
• Keeping Human Judgment Sharp