The doorman fallacy, herd mentality, and why most AI adoption doesn't stick

A striking number of "AI strategies" aren't strategies at all. They're a response to a question a board member asked after reading what a competitor announced. That's not a criticism. It's just worth naming, because it explains a lot about why so much AI adoption produces so little measurable change.

I want to be upfront about where this piece comes from, since most of the interesting part isn't mine. It was prompted by a LinkedIn thread from Ian Harris sharing an idea from Rory Sutherland, which sent me to Gediminas Lipnickas's piece on the "doorman fallacy" in The Conversation, which sent me back to McKinsey's recent organisational research. What follows is my attempt to connect those three things; credit where it's due, throughout.

The doorman fallacy

Sutherland's "doorman fallacy" starts with a simple observation: a hotel doorman's job looks, on paper, like opening a door. Automate the door and you've automated the job. Except the door was never really the point. The doorman also signals status to arriving guests, provides a beat of informal security, remembers regulars, and smooths a hundred small moments a checklist would never capture.

Many companies are falling for what is known as the doorman fallacy: reducing rich and complex human roles to a single task and replacing people with AI.

Lipnickas's examples are pointed. Commonwealth Bank of Australia laid off 45 customer service staff in favour of an AI voice bot, then rehired them once the bot's limits became a customer problem, not just an internal one. Taco Bell rolled out AI at the drive-through and ran into the same wall: some interactions need a human, especially at peak load, and no amount of prompt engineering fixes that on its own. Lipnickas cites an Orgvue finding that 55% of companies that replaced staff with AI now admit they moved too fast.

None of this is an argument against AI. It's an argument against measuring a role by its most visible task and assuming that's the whole job.

Why the fallacy spreads faster than the fix

Here's where herd mentality comes in, and where the LinkedIn thread that started this became genuinely useful. Sutherland's framing (via Ian Harris) reaches for an older analogy: early factories that switched from steam power to electric motors saw almost no productivity gain at first, because they kept the same physical layout: machinery still clustered around a central drive shaft, exactly as steam power had required. The gains only showed up once factories were redesigned around what electricity actually made possible: distributed motors, flexible layout, work organised around flow rather than around a shaft.

Most organisations adopting AI right now are running electric motors through steam-era floor plans. They've bought the technology. They haven't redesigned anything it was meant to change.

That's the herd mentality problem in one image. Adoption is visible and comparable: you can announce it, benchmark it, put it in a deck. Redesign is slow, internal, and doesn't photograph well. So the incentive is to be seen adopting, not to actually restructure around what's been adopted. Everyone ends up doing the same visible thing (buying the tool, announcing the pilot, replacing the task) because that's the behaviour that gets noticed and copied, while the harder, less visible structural work quietly doesn't happen anywhere.

What the numbers actually say

This is where McKinsey's recent organisational research is worth pulling in, because it puts a number on the gap between "adopting" and "transforming." Across its State of AI and State of Organizations work, McKinsey finds that roughly 88% of organisations now use AI in some form, but only around a third report scaling it across the enterprise, and just 6% see it contribute meaningfully to profit (5%+ EBIT impact). That's a wide, consistent gap between visible adoption and actual value: exactly the gap the doorman fallacy and the steam-to-electric analogy both predict.

The more interesting finding is what separates the small number of companies that do see real impact. It isn't the sophistication of their AI. It's leadership ownership and structural redesign: organisations with strong AI-focused change leadership report roughly 3.2x higher profit margins and are about 2.7x more likely to see AI adoption succeed across departments. And only 21% of companies have redesigned workflows end-to-end around what AI makes possible (the electric-motor step, in Sutherland's terms), even though that's the step high performers are three times more likely to have taken.

Put plainly: the technology is not the differentiator. Almost everyone has it. The differentiator is whether anyone was willing to do the unglamorous, unphotographable work of redesigning around it, and whether leadership treated that as the actual job rather than the announcement.

The uncomfortable implication

If herd mentality is doing most of the work in AI adoption decisions, then the standard "are we behind on AI?" question is close to useless. The more honest question is closer to: if we removed the social pressure to be seen adopting this, would we still choose to redesign this particular workflow, this particular role, around what AI can now do? Sometimes the answer will still be yes. Often, I suspect, the honest answer is that the redesign was never really the plan. The announcement was.

I don't think that makes anyone foolish. Herd behaviour is a genuinely rational response to uncertainty; copying what confident-looking others are doing is a reasonable strategy when you can't evaluate the underlying decision yourself. But it's worth at least knowing that's what's happening, rather than mistaking the herd's direction for a considered one.

A confession this piece forces on me

Writing this, I couldn't avoid asking it about myself: I used AI to help build this site and draft this article. Research, structure, a fair amount of the sentence-level work. I'm not going to pretend otherwise, not least because "be open about your sources" is the whole premise of the closing section below.

So here's the harder question underneath that: would you, reading this, know straight away if it wasn't really me? Tone alone, honestly, probably not any more. A model can hold a register close enough that voice stops being the tell. What's harder to fake is the specific, slightly inconvenient detail: the actual York hotel internship, the actual reasons a framework does or doesn't match what I've seen in the field, the willingness to say "I think this, but I'm not fully sure." Judgement over what's true and what's worth keeping is still mine to exercise or abdicate. AI changes the cost of producing the sentence, not who's accountable for what's in it. That's the same distinction I made in an earlier piece on AI and judgement: the mechanical part compresses; what's left, uncomfortably exposed, is whether a person was actually paying attention.

Which is exactly the doorman fallacy again, aimed at me instead of a hotel or a call centre. The visible task, producing readable paragraphs, is now cheap. The invisible one, deciding what's true, what's mine to claim, what I'd defend if pushed, isn't, and no amount of fluent output substitutes for it.

The same question, aimed at sales

That reframing is what makes this personally uncomfortable rather than just theoretical, because I sit close to exactly this trade in my day job. A lot of sales organisations are now using AI to draft outreach, personalise at scale, even run first-line conversations, for the same herd reasons covered above: everyone else is doing it, so not doing it starts to feel like the risk.

But sales has always been priced, implicitly, on a doorman-fallacy bet: the buyer isn't just paying for information (that part is genuinely commoditised now); they're paying for attention, judgement, and the sense that someone on the other end actually understood their specific problem and is accountable for the recommendation. That's the invisible part of the role, exactly like the doorman's status-signalling and quiet vigilance. When outreach, discovery, even negotiation start running through AI with the visible task automated but the price and the implied relationship left untouched, buyers eventually notice the mismatch, not because the words are wrong but because the care the price was supposed to be buying isn't actually there. People are quite good at sensing when they're being processed rather than heard, even when the processing is fluent.

I don't think the answer is "never use AI in sales", that's as naïve as automating the doorman down to the door. I think the answer is closer to the framework above: be honest about which parts of the interaction are still genuinely yours (your judgement, your accountability, your attention), and don't let the price of the relationship stay the same once that part has quietly been hollowed out. If the human touch is what's being charged for, it has to still be there. Otherwise the buyer isn't wrong to feel like they're talking to a machine at human prices, and eventually they'll act on that feeling, whether or not anyone in the room said it out loud.

---

Sources and further reading, since I've leaned on all three throughout this piece:

• Ian Harris's LinkedIn post sharing Rory Sutherland's take on AI, agencies and the steam-to-electric analogy, the piece that started this train of thought.
• Gediminas Lipnickas, "The doorman fallacy: why careless adoption of AI backfires so easily", The Conversation: the doorman fallacy itself, and the Commonwealth Bank/Taco Bell/Orgvue examples.
• McKinsey's State of AI and State of Organizations research (2025/2026): the adoption-vs-scaling gap and the leadership/redesign findings. Figures paraphrased from published summaries of that research rather than a single report; treat them as directional rather than exact.
• This site itself, and this article's drafting (research, structure and a good share of the sentence-level writing), were done with help from Claude (Anthropic). I edited, fact-checked and decided what to keep; see the confession above for why I don't think that distinction is cosmetic.