Everybody is an AI expert now, and nobody carries the can

There is a particular slide I have learned to brace for. It turns up around eight minutes into the pitch, it shows a row of boxes joined by arrows, and one of those boxes says AI. Nobody asks what is inside it. Asking would mean admitting you did not already know, and by that point in the meeting everyone has quietly agreed that they did.

I should say straight away that I have been the confident one in that room. I have wanted something to work and let the wanting do my reasoning for me, which is the same failure wearing a better suit. So this is not a piece about other people being stupid. It is a piece about a market that has lost the ability to tell confidence from competence, and about what that costs the people who were never in the meeting.

Driving a car is not automotive engineering#

The thing that changed is the interface. You talk to these systems in English, they answer in English, and the whole exchange feels like understanding. It is not. It is fluency, yours and theirs, meeting in the middle and producing a very pleasant sensation of having grasped something.

I have watched people recommend putting a language model in charge of a process that decides whether an invoice gets paid, without any working grasp of temperature, context limits, or the fact that the thing is a prediction engine rather than a store of facts. They were not lying. They genuinely could not see the gap, because the gap does not show up in the conversation, it shows up eleven weeks later in a reconciliation.

You can drive a car beautifully and know nothing about how the engine works. That is fine, right up until you are the one selling engines.

Bad advice travels faster, and always has#

What makes this worse than ordinary incompetence is the distribution mechanism. Somebody who can compress AI strategy into a confident post gets in front of more buyers than somebody carefully explaining why data quality determines the outcome. The medium rewards certainty, and certainty is the one thing nobody honest should be selling here.

I keep coming back to what the industry did with the phrase "you never get sacked for buying IBM". That instinct, fear dressed as prudence, never went away. It moved. It became the C-suite buying a consultancy as insurance, and it has now become buying the biggest AI name available so that if it goes wrong the decision looks defensible. The technology is new. The purchasing behaviour is a rerun.

And notice who these pitches are aimed at. Not the practitioner who would ask an awkward question about integration, but finance looking at perceived cost and a board measuring efficiency in headcount. Once those people are convinced, the thing happens whether or not it works.

Demonstration debt#

Here is where the money actually goes. Every AI implementation I have watched succeed spent most of its effort on things nobody demos: data quality, integration with the systems that already exist, redesigning a workflow that grew by accident over eleven years, working out who is accountable when the output is wrong. The AI part is real and it is often the smallest part of the bill.

The pitch skips all of it, because none of it looks good on screen. So you get an impressive proof of concept that cannot reach production, and I call that demonstration debt, because a debt is exactly what it is. Someone repays it later, with interest, and it is almost never the person who took it out.

MIT's Project NANDA put numbers on this in their 2025 report on the state of AI in business, finding that around 95% of enterprise generative AI pilots delivered no measurable return. That figure got passed around as evidence the technology does not work. I read it the other way. It is evidence that we are buying demonstrations and expecting systems, which are different products at very different prices.

The recursion nobody wants to talk about#

The part I find genuinely funny, in the way that makes you put your coffee down, is that a fair number of these proposals are themselves written by AI. Advice about implementing AI, generated by AI, sold by someone who cannot evaluate either half of that sentence. The snake has found its tail and is chewing thoughtfully.

Then the governance question arrives, usually about four months after it would have been useful. When the system makes a decision that costs somebody money or dignity, who answers for it? The consultant who recommended it has invoiced and gone. The executive who approved it was assured it was proven. The engineer who deployed it built what they were asked to build. Everyone points at everyone else, and the harm just sits there, unowned.

What actually works is slower and harder to sell#

The alternative is not complicated, it is just commercially inconvenient. You start with a conversation instead of a conclusion. You sit down with the people doing the work and find out what actually eats their week, before anybody says the word model. You educate honestly about what these systems are dreadful at, and you do it before money changes hands rather than after.

This does not scale. You cannot systematise a genuine conversation or automate honest education, and it demands practitioners who are comfortable saying "I do not know" and "that will not work", which is a difficult look in a market flooded with people promising the opposite. That is precisely why it is uncommon, and precisely why it works.

If you are buying, three questions will tell you most of what you need. Ask what proportion of the effort in their last comparable project went on data preparation and integration rather than the AI itself; anyone who says it was mostly the AI has either not done one or is not telling you about it. Ask them to describe a situation where they told a client not to use AI, and listen for whether the example has any texture to it. Then ask who is accountable when the system is confidently wrong, and watch how quickly the answer becomes a process rather than a person.

None of these are gotchas. Someone who has done the work will enjoy them, because they finally get to talk about the interesting part.

The choice, and the shrug#

I am not trying to gatekeep this. I object to consequences, not credentials, and some of the best people I have worked with came into technology sideways from somewhere else entirely. A new breed is arriving from other backgrounds who will take this seriously, and good.

But if someone would rather sell certainty than learn the trade, that is their choice and they are welcome to it. The market will sort them out eventually, though not before a few operations teams have carried something heavy on their behalf. Join the party or go home, and I would much rather people joined.

What I want from anyone reading this is not suspicion. It is better questions, asked earlier, of people who should be delighted to answer them. The technology deserves that, and so do you.