Why People Resist AI at Work: It Threatens What They Are Good At
The most-quoted number about AI projects failing is Gartner’s forecast that at least 30% of generative AI projects would be abandoned after the proof of concept. Its stated causes are poor data quality, inadequate risk controls, escalating costs, and unclear business value.
Read that list again. There are no people on it.
That absence is worth sitting with, because there is a body of research arguing that some of the most durable resistance to automation is not about data or cost at all. It is about what the technology does to the person being asked to approve it.
The research is about the approver, not the team
This is the part that usually gets lost when the finding is repeated, and it is the part managers should care about most.
Das Narayandas and Shunyuan Zhang at Harvard Business School published a Research circulated before, or instead of, peer review. It lets findings reach people early, and it means nobody independent has yet checked the method. Treat it as a serious argument from credible researchers rather than as a settled result, which is roughly how the authors would want it read too. in February 2026 on what they call Their term for tools that improve how the organization performs while eroding the authority, discretion or span of control of the very role that has to approve them. A system that automates scheduling makes the department faster and makes the scheduling manager less necessary. Both things are true at once, and the person who has to sign it off is the person holding the second one. . Their subject is not the reluctant employee at the bottom of the org chart. It is the person with the authority to say yes, and the awkward position they are in when saying yes diminishes the job they hold.
The paper’s practical proposal is aimed at the people selling these systems: design the offer so an approver can endorse it without losing standing. But the diagnosis travels further than the sales problem it was written for. If approval stalls where the technology undercuts the approver, then somewhere in your organization a sensible person is quietly not championing something, and it will not look like that from the outside.
Three ways a tool can threaten a job
Writing up the research, Harvard Business School’s own Working Knowledge names three distinct threats. They are worth separating, because they call for different responses.
- Role compression. Automation takes the high-status judgment work and leaves the residue. The job still exists, but the interesting part of it has gone, and what remains is supervision and cleanup.
- Control shift. The algorithm starts making the call. Even where the outcome is better, the person who used to decide is now reviewing a decision somebody else’s software made, and their sense of authorship goes with it.
- Span erosion. Influence over people, budgets and processes shrinks. This one is the most concrete, because it shows up in headcount and org charts where the other two only show up in how someone feels about Monday.
None of those are irrational, and none of them are solved by another round of training. A person who correctly perceives that a tool makes their expertise less valuable is not confused about the tool. They have understood it.
It rarely arrives as an objection
The reason this gets misread is that nobody says “this threatens my standing.” They say the outputs are not accurate enough, or that it does not fit how the work is actually done, or simply nothing at all.
Gallup’s February 2026 survey of 23,717 US employees found that what most separates non-users from light users is not access. It is belief. Non-users were far more likely to cite ethical concerns, 43% against 25% among infrequent users, and to doubt the tool was useful, 39% against 22%. Infrequent users mostly raised workflow fit and practical risk. Neither group is short of licenses.
Stated reasons are the reasons people are willing to say out loud. That does not make them false. It does mean the list of objections you collect in a meeting is not the same as the list of reasons, and a manager who only works the stated list will keep solving problems that were not the blocker.
What to do about it
You cannot train this away, and you should not try to argue somebody out of a correct observation. What you can do is change what the tool takes.
- Say what the role becomes, not just what the tool does. The unanswered question is not “what will this automate,” it is “what am I, afterwards.” AI Role Redesign covers doing that deliberately rather than letting it happen by default.
- Give away the boring half first. Aim the first use case at work nobody built their reputation on. Adoption that starts by automating somebody’s signature skill is starting at the hardest possible place.
- Move the judgment, do not delete it. If the tool now makes the routine calls, be explicit about which harder calls the person now owns. Control shift is much easier to accept when something arrives to replace what left.
- Check your own position honestly. If you are the approver and the system erodes your span of control, you are inside the finding, not above it. That is uncomfortable and it is better than mistaking it for your team’s problem.
And notice this is a different failure from the one in yesterday’s post, where people wanted to use AI and had not been given the time. Both look identical from a distance: low usage, polite non-adoption. One is fixed by naming a priority, the other by naming what the job becomes. Treating the second as the first is how a manager ends up repeating themselves.
The takeaway
The standard explanations for AI stalling are all about the technology and the budget. The research says a share of it is about identity: what the tool does to the standing of the person being asked to adopt it, including the person signing it off.
Before the next rollout, answer the question nobody asks out loud: after this works, what is this person for? If you cannot answer it, you have found the resistance, and it was never about the tool.
Sources: the self-disruptive technologies concept is from Selling Self-Disruptive Technologies: Identity-Compatible Advantage and the Role-Level Microfoundations of Automation Adoption (opens in a new tab) (Narayandas and Zhang, Harvard Business School Working Paper 26-050, 12 February 2026). Two caveats. It is a working paper rather than peer-reviewed work, and its framing is a B2B selling problem: how a vendor should design an offer an approver can accept. The extension from that to internal adoption is my reading. I also could not extract the full text, so the description rests on the published abstract. The three threats are named in Harvard Business School’s own Working Knowledge (opens in a new tab) write-up of the research rather than in the abstract itself. The barrier figures are from Gallup’s AI in the Workplace: What Separates Adopters and Holdouts (opens in a new tab), 23,717 US employed adults, 4 to 19 February 2026. The 30% abandonment forecast is Gartner’s, from July 2024 (opens in a new tab), and it is quoted here for the causes it lists rather than as evidence for this argument. Some coverage attributes that 30% to employees rejecting tools. Gartner does not, and the two claims should not be merged.
Related: Your Team Is Building AI Tools After Hours is the other reason adoption stalls quietly, and the one to rule out first because it is easier to fix. AI Role Redesign: Every Job Is Changing, Almost No One Is Rewriting Them is the practical answer to “what am I afterwards.”