The quiet reason CIOs are slowing AI down

Start with a number that should ruin your week.

In MIT’s GenAI Divide study of enterprise adoption, only about 40 per cent of organisations had bought official large language model subscriptions. Workers at more than 90 per cent of those same organisations were already using personal AI tools for work. The finding that got the headlines was that roughly 95 per cent of formal pilots produced no measurable return. The finding that matters are the other one.

Your people have already adopted AI. They did not wait for the architecture review. They did not raise a ticket. They did not tell you.

This is not a security failure, though it is that too. It is a verdict. Every one of those employees made a private judgement that the tool’s value exceeded that of your process and acted on it. The governance function did not prevent adoption. It only prevented visibility, measurement and control of adoption. That is the worst of all possible outcomes.

Nobody is ever fired for the opportunity they declined

The stated reason for the slow lane is always risk. Data sovereignty. Model drift. Vendor lock-in. IP leakage. Each is a legitimate concern, yet none of them explains the behaviour, because the same organisations happily accept far greater risks when the upside is legible to the board.

Philip Tetlock’s research on accountability gets closer. When people know what their audience thinks, they do not analyse. They conform, using what Tetlock called the low-effort acceptability heuristic. When the audience’s views are unknown, the same people reason far more carefully and criticise their own position before anyone else can.

Read that as a description of your executive committee. A CIO who can read the board’s mood is not incentivised to be right. They are incentivised to be defensible. And caution is unfalsifiable in the short term: the cost of the deal you did not do never appears in a variance report, never gets attributed to a meeting held eighteen months earlier, and never has a name on it.

The part that does not get said out loud

Here is the uncomfortable version.

For thirty years, the authority of the technology function has been collateralised by scarcity. You knew things the business could not know and could not check. Budget followed that asymmetry. Headcount followed it. A seat at the table followed it.

AI attacks the collateral, not the job. When a finance manager can build a working prototype over a weekend, when a category buyer can interrogate a data set without a BI queue, when a graduate can produce in an afternoon what used to require a specialist and a fortnight, the scarcity that underwrote the CIO’s authority quietly ceases to be scarce.

Chris Argyris spent a career on what organisations do next. He found that defensive routines are learned behaviours for handling threat and embarrassment, that they are taught by culture rather than by individuals, and that they outlive everyone who created them. Their signature is what he called the “undiscussability of the undiscussable” the topic cannot be raised, and the fact that it cannot be raised also cannot be raised.

No CIO says, “I am slowing this down because it devalues my expertise.” Almost nobody thinks it consciously. That is precisely the point. Defensive routines do not feel like defensiveness. They feel like diligence.

Domain expertise is depreciating, and it is not coming back

Daniel Kahneman and Gary Klein spent years arguing about whether expert intuition can be trusted, and eventually agreed on the answer. It can, but only under two conditions: the environment has to be regular enough to be predictable, and the expert has to have had enough practice and feedback to learn those regularities.

Enterprise technology just failed the first condition. Build versus buy, minimum viable team size, time from idea to working prototype, the cost of a custom integration, what counts as a hard problem: the rules that generated twenty years of accumulated pattern recognition have moved. Robert Merton had a phrase for what happens next, written in 1940. He called it trained incapacity, the state in which one’s own abilities function as blind spots.

The evidence that the gradient has flipped is now hard to argue with. Brynjolfsson, Li and Raymond studied 5,179 customer support agents and found a 14 per cent average productivity gain from an AI assistant, with a 34 per cent improvement for novices and almost none for the experienced. Dell’Acqua and colleagues ran a field experiment with 758 Boston Consulting Group consultants and found that below-average performers improved by 43 per cent, against 17 per cent for those above average.

Read those numbers as a compression of the skill distribution. Deep specialisation used to be the moat. AI is filling the moat in, from the bottom. The same BCG study contains the warning: on tasks outside the model’s capability, consultants using AI were 19 percentage points less likely to get the right answer. Judgement still decides outcomes. But the judgement that pays now is knowing where that capability boundary sits this month, and that is learned by using the tools daily, not by having survived three ERP migrations.

What thinking differently actually looks like

Four shifts, none of which require permission.

1. Make governance proportional to reversibility. Irreversible commitments deserve every gate you own. Reversible experiments deserve almost none. Treating both the same way is the most expensive habit in enterprise IT, and it is entirely self-inflicted.
2. Stop hiring for accumulated years. Start hiring, promoting and listening for rate of learning. The person who was excellent in 2019 and the person who was excellent last quarter are no longer the same bet.
3. Use the tools yourself, daily, badly at first. A CIO who has not personally shipped something with AI in the last month is making capital allocation decisions about a technology they know only through vendor decks. That was survivable with ERP. It is not survivable here because the capability boundary moves faster than the reporting cycle.
4. Count something uncomfortable. How many times this year has a junior person overturned a senior person’s decision on evidence? If the answer is zero, that is not harmony. That is a measure of how much intelligence your organisation is throwing away.

Why now, and not next planning cycle

Because shadow adoption compounds. Every month you wait, more of the organisation routes around you, more institutional knowledge leaves through an unmanaged endpoint, and more of your remaining authority is spent on enforcement rather than direction. There is a point where the technology function stops being the place where decisions are made and becomes the place where decisions are notified.

The irony is complete. The instinct to protect your position by slowing things down is the single most reliable way to lose it. The CIOs who come out of this decade with more authority, not less, will be the ones who spent the next twelve months looking conspicuously less certain, less procedural and less senior than they were comfortable being.

Harmony is cheap. Outcomes are expensive. Only one of them is on the scorecard.