Everyone expected AI to come from the big companies. They had the money, the data, the researchers and a twenty year head start. The research came out of those labs. The products did not, and the reason has very little to do with talent or budget.
The question, from somewhere near the back of the room.
So why did your organization not lead the AI revolution?
What the leader said.
Yeah. Look. So I would actually push back on the premise slightly, and I say that with respect. I think if you zoom out to the macro level, what you are really seeing is a deliberate posture rather than an absence. Because at our scale, any capability we operationalise has to be evaluated holistically across a genuinely complex set of stakeholder considerations, the regulatory surface area we sit within, and candidly the trust equity we have built with customers over decades, which is not something we are willing to put at risk for a news cycle. So we have been very intentional about building the muscle first. We have stood up a cross functional working group reporting into the Transformation Office, we have aligned the org around a single north star, and we are already seeing really encouraging signal from a number of internal proofs of value. Now, is there more texture there? Absolutely. Happy to double click on that offline. But directionally, what we are focused on is one platform, one roadmap, one set of differentiated, enterprise grade, responsible by design capabilities that meet our customers where they are on their journey. That is the art of the possible for us. And I think you will see us lean into that pretty aggressively in the fullness of time.
What I heard.
We did not.
That was two hundred and nineteen words on their side and three on mine. I would like it on the record that nobody in that room looked confused. Everybody understood the answer perfectly. Several people wrote it down.
The same paragraph, with three nouns changed, was deployed for mobile. Then for cloud. Then for blockchain. Sit through all four and you develop an ear for it, the way a piano tuner does, and after that you can hear the shape of the answer before the sentence finishes.
For a long time I thought this was evasion. I have stopped thinking that. The person saying it is describing a machine they are standing inside and cannot see the outline of, and everything below is an attempt to draw it.
So start with the proposal.
It is nineteen pages and it is good. Somebody spent a weekend on it, which you can tell, because nobody produces a diagram like the one on page four on a Wednesday afternoon between meetings. It explains the whole thing in one picture. I have met perhaps six people in twenty years who can do that.
It enters the system on a Tuesday.
What happens over the next eleven months involves twelve people, four calendars, one reorganization, and at no point does anybody do anything wrong.
There is no villain in this story. I have been looking for one my whole career and they keep not being there.
Nobody in the Room Is the Problem
Legal has a concern, and Legal is right.
Security has a standard, and the standard exists because of an incident that nobody discusses and everybody remembers. The platform team would prefer you used the platform, which is what platform teams are for. Finance wants the number smaller, which is the whole of the job and not a personality defect. Somebody wants the launch moved so it does not collide with the other launch, and they are correct, because two launches in one week is one launch and a rumour.
Every note is reasonable. Most of them make the thing slightly safer. Several of them are genuine improvements.
And the person who wrote the nineteen pages is in every one of those meetings, nodding, because the notes are fair. They say “that is a fair point” perhaps eleven times over eleven months and they mean it every single time.
Nobody is doing anything to them. They are helping.
Eleven months later it ships. It is fine. Nobody uses it very much and nobody is entirely sure why.
Meanwhile, in a smaller building, four people you have never heard of shipped the same idea in March.
The Arithmetic of Twelve Approvals
I want to stop and do some multiplication here, because I had this completely backwards for years and the numbers are what fixed it.
Take a proposal that needs eight people to sign off. Say each of them says yes nine times out of ten.
Nine out of ten is a good manager. Nine out of ten is a person you would recommend to a friend. There is nothing wrong with nine out of ten.
Now multiply. Nine tenths, eight times over.
Forty three percent.
More than half of the proposals die, and every single person in that chain behaved well. Add four more stakeholders and you are at twenty eight percent. Nobody became difficult. You added four more chances to stop.
Those percentages are mine, not measured, and real approvals are not independent the way this multiplication pretends. People negotiate, they revisit, they take their cue from the person who went before them. Treat it as a model that shows the shape of the thing rather than a calculation of your odds.
Now the part I did not see coming, and it took me an embarrassingly long time.
Safe ideas and bold ideas do not get the same nine out of ten. A safe idea is one everybody can already picture, so call it ninety five percent each. A bold idea is the one where two people in the room honestly cannot see it yet, so call it seventy.
| Chance each person says yes | Survives 8 approvals | Survives 12 approvals | |
|---|---|---|---|
| The safe idea | 95 percent | 66 percent | 54 percent |
| The ordinary idea | 90 percent | 43 percent | 28 percent |
| The bold idea | 70 percent | 6 percent | 1 percent |
At eight approvals, the safe idea reaches a customer about eleven times more often than the bold one. At twelve, it is thirty nine times.
Nobody voted for that. Nobody would defend it out loud. It falls out of the multiplication, the way water falls downhill, and it does it every single time.
The filter selects against bold ideas quietly, while everybody in the chain behaves impeccably. That is why it has survived so long. There is nothing to point at.
Twelve approvers and a genuinely bold idea gives you about one percent. Run it a hundred times and you should expect one brave thing to get through, and your organization will tell the story of that one thing at conferences for a decade.
Now put the thing everyone is actually asking about through that filter. In 2019 a serious bet on generative models was precisely the idea two people in the room could not picture yet. There was no revenue attached to it. It embarrassed itself in demos. It cut across a roadmap twelve people had already agreed to. It never needed a skeptic to kill it. It needed a room of reasonable people, and every one of those companies had one.
Why Committees Build Smaller Things
When twelve people must agree, you do not get the best of their twelve visions. You get the part all twelve can live with, and that is a smaller thing than any one of them was imagining.
Somebody at the small company shipped one person’s conviction, whole, including the parts that were wrong. Some of those companies died of the wrong parts. The survivors shipped something with an opinion in it.
A product designed to offend nobody attracts nobody. Everything you personally love is against something, and a committee is extremely good at removing the part that was against something. It does this gently, over several meetings, and everyone feels the thing is getting better.
Why Failure Has a Name and Inaction Does Not
Ask why intelligent, ambitious people keep choosing the smaller idea and somebody will tell you it is culture. Or courage. Or appetite for risk.
The answer is duller than any of those, and it sits in the payoff table.
A bold bet that fails has a name attached. There is a review. There is a slightly careful conversation in March. Everybody remembers who ran it, including people who were not there.
A missed opportunity has no name attached at all. There is no post-mortem for the product you did not build. In twenty years I have never once watched somebody stand up and account for the market they quietly declined to enter.
Courage has nothing to do with it. The incentives are doing exactly what they were built to do.
Every individual in the chain is behaving correctly. The organization is behaving stupidly. Both true at once, which is why asking people to be braver so rarely changes anything.
What You Lose a Year Later
Losing the idea is the cheap loss.
The expensive one arrives about a year afterwards and makes no sound at all.
Once you have watched two of your own ideas improved to death, you learn. You just get accurate. Laziness never enters into it, and neither does loyalty. You start bringing the idea that will survive the room instead of the one you believe in. The second one costs nine months and a little of your reputation.
Nobody decides to do this. It is what happens to anybody who has burned their hand on the same door twice.
You will never see this happen. The proposals keep arriving, on time, well argued, and slightly smaller every year.
By the time you notice, it has stopped being a process problem. Your best people have quietly recalculated what is worth asking for, and the annoying thing is that their arithmetic is correct.
Forty Initiatives Is Not a Strategy
“We have forty AI initiatives” is a confession, delivered cheerfully as a boast.
Constraint forces the choice that strategy is supposed to make. A company with eleven months of money in the bank cannot fund its second best idea. You can, so you do, and then you do it thirty eight more times, and everybody gets a workstream and a slide and a monthly.
Those forty add up to a single bet, spread so thin that none of them can fail visibly and none of them can win either.
Spreading risk is the correct instinct for a portfolio of investments. Conviction does not average.
Why Your Best People Actually Leave
The good ones I have watched leave were not chasing money. A raise does not usually move somebody who likes their team and their manager.
What moves them is latency. The time between having an idea and watching a human being use it.
Somewhere else that number is a week. In your organization it is nine months, and six of those months are calendar rather than work. Waiting for the board that sits monthly. Waiting for the slot. Waiting for the one person who has to approve it and is on leave until the eleventh.
I want to be careful here, because it is easy to romanticize the people who leave and unfair to the ones who stay. Plenty of excellent engineers stay for twenty years and do the best work of their lives. Still.
Nobody writes “the approvals were slow” on an exit interview. It sounds like an excuse, and “a better offer” sounds like a reason, so you will never be told the real number.
The cruel detail is what the process selects on. It does not test for judgment, or for the ability to see the whole thing at once. It tests for willingness to wait, and that is not the same quality, and it is not correlated with the one you actually needed. Your approval cycle does more than lose people. It chooses which ones, on a criterion nobody chose.
Five Questions for Leaders
If you run something, these are worth ten minutes on a Friday. The answers matter more than the questions.
| The question | What the answer tells you | |
|---|---|---|
| 1 | How many people can say no to this? Not approve. Just say no. | Count them honestly, then look at the table above. |
| 2 | When did we last ship something an important stakeholder actively disliked? | No answer from the past year means the filter is too tight. |
| 3 | What is our median time from idea to something a customer touched? | Median, not best case, and count the waiting. |
| 4 | Name the last person penalized for a missed opportunity rather than a failed attempt. | No name means you have found the incentive problem. |
| 5 | If we could fund only one current initiative, which survives? | Then ask why it is not already the only one funded. |
Question four is the one that changes rooms. Most leaders have never once watched somebody’s year go badly for something they did not do.
In Defense of the Thing I Have Been Attacking
Bureaucracy is scar tissue. Every gate exists because something once went badly. Ask somebody who has been there long enough and they can tell you the story, including the name of the customer and roughly what it cost. A bank that ships like a startup gets fined. A hospital that ships like a startup harms somebody.
Every one of those controls is somebody’s memory of a bad year.
Running one process at one speed for everything is where it goes wrong. Same gates, same committee, same calendar, applied to a prototype that three people would have tried and nobody would have missed. Two speeds would fix a great deal of it, along with the honesty to say out loud which lane a thing is in before it enters.
Why Big Companies Are Overfitted
There is a word for this in machine learning, and it fits uncomfortably well.
Train a model too thoroughly on the past and it performs beautifully on the past. Show it something new and it falls over, confidently, because it memorised rather than understood. The word is overfitted.
Your organization spent twenty years learning what works. That learning was earned and most of it was expensive. Every one of those approval gates encodes a real lesson from a real failure.
Then the world it was fitted to moved.
The incumbent is overfitted. It scores brilliantly on the last decade and generalizes poorly to this one.
The smaller company is not smarter and does not have better people. Some of them are your people, from eighteen months ago. It simply has less to unlearn and fewer humans who must agree before anybody is allowed to try.
The Part I Cannot Answer
I have been chewing on this one for a while and I still do not have it.
Every one of those small companies is hiring. They are adding a legal function, a security review, a platform standard. In four years they will have twelve approvers of their own and somebody will write this exact essay about them. So either this is a cycle and there is no lesson in it, or there is a way to grow without acquiring the filter. I have never watched anybody manage the second one.
If you have, I would genuinely like to hear about it.
The One Who Stayed
None of this needed a villain. Everybody did their job. The gates worked exactly as designed. The concerns were raised and they were fair and several of them were genuinely good catches.
The nineteen page proposal is still in a folder somewhere, with the diagram on page four that explained everything.
The person who wrote it is still with you.
They are good at their job and they are well liked. Their last three proposals all shipped, roughly on time and roughly as written, and you approved every one of them without much discussion, because there was not much to discuss.
Worry about that one. Not about the people who left, because the people who left are busy and doing fine. Worry about the one who stayed, and got quieter, and now brings you exactly what you will say yes to.
They do not know it happened either.
One more thing and then I will leave it alone. The idea underneath this essay, that a room full of people behaving well can produce a result not one of them wanted, is what runs through all thirty essays in my book AI: Nobody’s in There. But we’re still in here. Every essay is free to read at pinaldave.com, and there is a paperback on Amazon if you would rather hold something real.
Nobody said no. The idea died anyway, on schedule, with everyone’s support.
Reference: Pinal Dave (https://blog.sqlauthority.com/), Leadership and Bureaucracy, X
