Thoughts. What I notice while building.
I get this call every week now, and it is almost word for word the same every time. We had something built. It cost real money. It does not really run any more. And nobody here says the word automation without smirking.
I sell into that same market. That belongs at the top, otherwise the rest reads as a swipe at the competition. It is not. It is the most honest look I have at my own business.
So before I claim anything, take a look at this market. Scroll slowly.
Nobody in there disappears for being bad. The expensive ones disappear.
The market rewarded exactly what is broken now
Something happened between 2023 and 2025 that almost nobody registered as an event: building a convincing demo has cost close to nothing ever since. A workflow in n8n, Make or Zapier that holds up in a meeting is assembled in an afternoon. An agent that answers cleverly in a demo takes a morning.
Keeping the same thing alive for two years did not get cheaper. Not by one percent.
That put this market in exactly the position George Akerlof described in 1970, using used cars and long before any workflow existed. When the buyer cannot verify quality up front, they will only pay the average price. For the person who works carefully, that price sits below their cost. So they leave. The average drops, the price drops, and in the end every buyer gets exactly what they were afraid of.
That is the drawing above. Not one villain in it, just a falling price and offers that cannot be produced at that price.
Before buying, you could not tell a workflow built to run from a workflow built to convince. In the meeting they look identical. The difference arrives in month nine, when the eleventh edge case shows up and nobody remembers what branch seven was for.
There is exactly one reliable clue about what you actually bought, and it is not in the deck, it is in the contract. Did the engagement end at handover? Then handover was the product.
Anyone working in a market where only the demo gets judged will optimise the demo. That is not fraud, it is adaptation. I feel the pull myself: the meeting rewards the fast picture, not the question of who maintains the thing in eighteen months. The mistake sat in the buying criterion. And the buying criterion is the only part of this story you fully hold the second time round.
Two curves pulling apart
What convinced me this is not an anecdote problem: it happens everywhere at once, in the same order, with the same sentences.
Anything that happens everywhere at once does not have an individual cause. Anyone reading their own failure as a personal failing simply has the statistics against them. The trouble is that tens of thousands of companies are drawing the wrong lesson from those same statistics. And that lesson costs more than the project did.
The point where I had it wrong too
Here comes the part where I had to correct myself. Because even if you had bought cleanly back then, the result would probably still have felt wrong after nine months.
Scroll again. And stop exactly where you would normally write the post-mortem.
Erik Brynjolfsson, Daniel Rock and Chad Syverson supplied the explanation you cannot unsee once you have seen it. New general purpose technologies demand investments nobody books as investments: recutting processes, cleaning data, changing roles, retraining people, actually switching the old route off. That work produces no receipt. It sits in the accounts as expense and never as an asset. So measured productivity falls before it rises.
That is the hatched area. And something uncomfortable follows from it: there is a window in which a good project and a bad project produce exactly the same numbers. Anyone judging inside that window judges negatively. However right everything was.
I used that as an excuse for a long time and it was nonsense. Because the curve does not turn up on its own. It only turns up if the hatched work actually happened. Buy the tool and leave the process as it was, and you get the trough without the climb, permanently. That is the difference between measured too early and rightly failed, and it is testable:
| Measured too early if … | Rightly failed if … |
|---|---|
| the old route still runs alongside the new one | the process runs three times a month and takes ten minutes |
| only the tool was swapped, not the process | the bottleneck was never capacity, it was demand |
| the data was as messy on day one as before | the rules change faster than anyone can model them |
| there is an owner for the build and none for the operation | the case hung on one person who has since left |
| it was measured in weeks rather than quarters | the saving was smaller than the running cost from day one |
Three hits on the left means too early. Three hits on the right means drop it. Both are usable answers. What is worth nothing is the third option, and it is the one I hear most: we do not know, and we do not talk about it any more.
What appears in no post-mortem
There is one number nobody writes down, because it shows up in no set of accounts.
Dan Andrews, Chiara Criscuolo and Peter Gal analysed millions of company accounts for the OECD. Their finding puts the productivity slowdown of the last decades in a different light: there was never a slowdown at the frontier. The most productive firms keep improving at their usual pace. What stalls is everyone else. And the gap widens, most sharply where software is involved.
That sounds like statistics and it is a threat.
The lost amount is on your balance sheet. The lost rate is nowhere.
A written-off project budget is a level. It hurts once and then it is over. A company that touches nothing for two years while three competitors push unit costs down a few percent a year does not lose a sum. It loses a rate, and rates compound.
The cruel part is the delay. While the order book is full, nobody notices. It becomes visible exactly when the market tightens and decisions get made on price. Then you have a company whose offer was competitive three years ago, and the only explanation it has left is that the market has gone mad.
Both reactions lead to the same place
At this point most people believe the danger is behind them. There are two.
At bankruptbyai.com I collect documented cases from both directions, which is why the line at the top of that site reads the way it does: there are two ways to go bankrupt with AI. Overheating and freezing.
Overheating looks like courage. The first attempt was rough, so the dose gets raised. More tools, more promises, and eventually somebody says the headcount could come down now. The roles go before the new route carries the load. Reversing then costs more than the whole project, and the most experienced people have already left.
Freezing looks like prudence, and that is precisely what makes it more dangerous. Nobody decides anything. Things simply stop happening. "It did nothing for us" becomes company policy without anyone ever having agreed to it.
Two reactions to the same stimulus, the same register, only at different speeds.
What I tell people who got burned once
Not: just try again. That is cheap, and it is not honest either, because the first result was real.
Instead: you did not test automation, you tested a purchase. Under conditions that made the bad purchase more likely than the good one. The result says a lot about the market of 2024 and little about your company.
So the second attempt is not a repeat, it is a different transaction. I change four things about it, and three of them cost nothing.
The autopsy first. The failed project was the most expensive market research your company has ever bought. Clear it away without dissecting it and you throw out the last return it had. One hour, three questions: what worked, which concrete edge case did it tear on, who should have noticed.
Then the contract, and this is the part that actually defeats the lemon market. As long as the engagement ends at handover, suppliers compete on the demo. Put twelve months of operation with an agreed response time in it, and a badly built workflow can no longer win, because the supplier pays for their own mess. You then do not need to be able to judge quality. You only need to cut the contract correctly. That is Akerlof's own answer, more than fifty years old: guarantees instead of trust.
Then the scope. One process taken to the end is worth more than ten at eighty percent, because a company runs every half-finished route twice. The curve only turns up where the old route is genuinely gone. Take the one process where you dare to do that.
And finally the ending, before the beginning. Write down how you will know in twelve weeks that it is not working, and attach a number that has to move. Not hours saved, a cost line. A project you cannot stop is not a project, it is a bet. And if the stop was agreed in advance, nobody has to sell it afterwards as a defeat. That single fear keeps companies still for two years.
When I say don't
Otherwise the rest reads like a sales pitch: there are cases where I say leave it.
If the process runs three times a month and takes ten minutes, nothing adds up, not over five years either. If the bottleneck is demand rather than capacity, automation only moves the problem into the warehouse. If the company is being sold in eighteen months, the curve is longer than the horizon. If you are two people short of holding the place together, a software project is the wrong building site.
Stopping a project is a decision, it can be the right one, and it can be taken back.
What cannot be taken back is something else. A company where nobody has been able to tell a good offer from a bad one for two years, because nobody engages with the subject any more, will not recognise the right opportunity either. And it does not arrive with a warning.
The three drawings above are the same thought at three resolutions. The market explains why you bought what you bought. The curve explains why it feels wrong even when it was right. The fork explains why the most obvious reaction to both is the most expensive one.
Stop doing projects when the maths says no. Never stop being able to judge.
Related
- Why AI is not interesting right now: the same pricing question, one level up
- A tool, not a miracle: what the boom costs and when less AI is smarter
- Technology does not repeat itself. Our stories do.: why the explanation always arrives after the winner
- Over budget, over time, off target: the same question, one industry earlier
- bankruptbyai.com: the documented cases from both directions
