AI Decides How. Humans Decide What, Why, and If.
A few nights ago one of my partners dropped a message into our exec thread that I have not stopped thinking about since.
Should AI be an insurance policy and a vote at IC?
IC is our investment committee at Bambu Capital. It is the room where we decide, as a firm, whether to put money into a company. He was asking whether the machine should sit at that table as a member, with a ballot of its own.
My first reaction was two objections, and neither one was about capability.
AI will be over-critical when being critical is what earns your approval. AI will also be a sycophant, agreeable to a fault, when agreement is what earns your approval. Those are the same flaw wearing different outfits. In both cases it is optimizing for your reaction rather than for the truth. A committee member who tells you what you want to hear is worse than no committee member at all, because you count the vote anyway.
So my answer was no. Then I spent the rest of the night arguing with myself about it, which is usually a sign the question was better than the answer.
Start by admitting where the machine already wins
A lot of people in my industry are still performing skepticism about this, so let me say the uncomfortable part plainly.
On words and math, the machine already beats us. Not soon. Now.
No human can hold the universe of language or mathematics in their head at once and draw on all of it in the instant a decision is required. That is not a failure of effort. It is a physical limit on what a person is.
It is why AI is already better than any of us at the recall half of medicine, the recall half of the law, and, honestly, at pastoral counsel. Not because it cares, and not because it has wisdom. Because none of those professionals could have read everything, and none of them remember everything they did read. The machine read everything. It remembers all of it. And it can bring the whole of it to bear at once, which no doctor, lawyer, rabbi or minister has ever been able to do.
Then there is execution, which I think is underrated even by the enthusiasts. The machine does not just know. It does the work, at a volume and consistency no team can match, and it does not get tired at hour nine.
Bring that back to my actual job. We have an investment committee meeting tomorrow. There is no chance that I or either of my partners has read every document uploaded to that data room and every cell of every worksheet in every workbook. Even with unlimited hours we would not retain it. The machine did, and does. That is precisely why our review of a deal is more thorough than it has ever been in my career.
So why not simply let it decide?
That is the question another of my partners put to me, and he was not being cute about it.
If we load in everything we have on a business, a market and an industry, why wouldn’t we just trust it to make the decision? If it says not to do the deal, we are effectively arguing against the data.
That is the strongest form of the argument and it deserves a straight answer instead of a nervous one.
Here is mine. The machine is superintelligent at how. It is not intelligent at all about what, why, and if.
What is a question of direction. What should we even be looking at? A model can rank the opportunities you hand it. It cannot tell you that the entire category is the wrong place for this firm to spend the next decade of its life. That is a judgment about who you intend to be.
Why is a question of values and consequence. Why this business, for these investors, at this moment, given the promises we made to people who trusted us with their money. The machine has no promises outstanding. It has never had to make the phone call.
If is the veto, and it is the one that matters most. Whether to act at all. Whether the fact that we can is any argument that we should.
There is also the thing my partner named in three words when I pushed him: outside of reading people. An enormous share of any deal is whether the CEO will still be there in year three, whether a founder is telling you the whole truth, whether the room went quiet at exactly the wrong moment. I have been wrong about businesses. I have been more often wrong about people, and never once from reading a spreadsheet.
And the machine has no stake. It does not lie awake. It does not have to face the investor.
What we actually built this week
Since the honest answer to “are you using AI” is yes, heavily, I would rather say so clearly than have someone discover it. My partner made the sharpest version of this point: if AI is a substantial part of our process, we should treat disclosure as a differentiating strength rather than a liability we hope nobody raises.
So, concretely. Every document attached to a deal in our platform is read in full by a model, automatically, when it is uploaded. Our data room’s assistant reads room documents in order to answer questions about them and to build the search index behind them. The deep review of a deal itself works from the room’s document inventory and from the summaries we already hold, and opens no data-room file. AI drafts the IC memo. AI runs the pre-mortem, where we assume the deal was done and failed completely and work backwards from the autopsy. No AI writes a change to any record without a person confirming it first. And no AI holds decision authority. Decisions are entirely human.
Two qualifications, because I would rather put them here than have somebody find them later. Some of that happens without anyone pressing a button: documents are summarised the moment they land, and the search index rebuilds on a schedule. And our data room’s assistant answers an investor’s question directly, in the moment, without a partner clearing the answer first. Every one of those exchanges is recorded and reviewable by us, which is a control after the fact rather than before it. I would rather state that plainly than let “a human approves everything” do work it cannot honestly do.
The decision line is different, and that one is absolute. It used to be a sentence in a policy document. As of this week it is a property of the software.
Every decision now carries its reasoning.
Each ballot has a Why box, kept separate from any conditions, shown under the voter’s name. Settling a gate requires the firm’s reasoning in a sentence, not just a tally. And at the moment a decision settles, the system freezes that reasoning together with what the record actually held right then: the screen verdict and the denominator it was scored against, how many facts were sourced, which documents were attached and whether their summaries were still current, and the latest pre-mortem verdict. Anything missing is named as missing rather than quietly scored as a zero.
It also records the decisions we later withdrew, because a reversal is training data too, and usually the most honest kind.
Every ballot is stored with a voter kind. Today that is always human. A gate holding a non-human ballot refuses to settle. The promise that AI informs and people decide is now enforced in code rather than asserted in a marketing line.
Two reasons for the trouble, and only one of them is about the machine.
The first is that a documented why is the only way a committee ever learns anything. My partner was right to flag the obvious limit: we will have a dozen of these, not a thousand, and a sample that small will not hand you causation. It might give you correlation on how we weight risk. It will certainly give you a mirror.
The second reason is that we are going to be wrong. Not might be. Will be. When it happens, the difference between a firm that can show its reasoning was sound at the time and a firm that cannot is enormous, and it is a difference you can only bank in advance.
And yes, eventually that corpus is exactly what would let a machine cast a ballot worth counting. It could bring every prior decision and every stated reason to bear on this one, which no human in the room can do. But that is a Fund II conversation at the earliest, because you cannot train on a history you never bothered to write down. Which is the entire argument for writing it down now.
The loop, which is the part we are actually investing in
There is a name for what that record turns into, and it sounds far more dramatic than what we are doing. Recursive self-improvement. A system that improves its own ability to improve. In the AGI conversation it is the phrase people reach for when they talk about a fast takeoff, and it is usually said with a certain amount of fear.
At the scale of one investment committee it is neither dramatic nor frightening. It is a loop, and it is the entire reason the record is worth the trouble of keeping.
Here is the loop. We make a decision and we write down the reasoning. The system freezes that reasoning next to what the record actually held at that moment. Later the outcome arrives, because outcomes always arrive. Now we hold a pair: what we believed, and what happened. Do that enough times and you have the one thing no amount of general intelligence can substitute for.
A model trained on the whole internet knows how investors think. A model trained on that record knows how we think, including where we are reliably wrong.
Then the loop turns. The next pre-mortem is written against a sharper picture of how deals like this one have actually failed for us specifically. The questions in the next committee meeting are better. The decision is better. Which produces a better record, which sharpens the pre-mortem after that. Each turn raises the floor a little.
The part I did not anticipate is that the loop improves the humans faster than it improves the machine, at least early on. Reading your own reasoning from eighteen months ago, sitting next to what the company actually went on to do, is a bracing experience. Most of us have never had that mirror held up with any rigour, because we never wrote the reasoning down in a form that could later be checked against reality.
The validation runs in both directions, and that is the half people miss. We validate the machine, because a human still has to decide whether the work was done correctly. The machine also validates us. It can see that we called management quality the deciding factor in March and then waved it through in July. No partner in that room would catch it. The record does, without an opinion about it, which is why it is tolerable to hear.
This is why we are spending real time and real money building it rather than buying a tool that promises it. The cost lands years before the benefit does. That is usually the tell that something is worth owning.
Now the trap, because there is one.
A loop trained only on your own reasoning converges on your own blind spots.
Feed a system nothing but your prior decisions and the logic you gave for them, and what comes back is an extremely expensive mirror that agrees with you. That is the sycophancy problem from the top of this piece, except institutionalised and wearing a lab coat. It would feel like rigour. It would be the opposite of rigour.
The only thing that breaks that loop is outcomes. Reality has to sit inside it. Which is exactly why the record freezes what we believed at the moment of decision instead of letting us quietly revise it afterwards, and why it keeps the decisions we later reversed. A reversal is the most honest training data a firm produces, and it is precisely the data an ego would prefer to lose.
The honest timeline is slow. We will have a dozen of these, not a thousand, and this compounds across funds rather than quarters. That is fine. So does everything else worth building.
The line we already crossed without noticing
Somewhere in the thread I claimed that most web traffic is now AI. My partner pushed back that bots have been roughly 40% of it for years and we simply did not use to call it traffic.
He is right, and his correction is more interesting than my claim was. The line was not crossed this year with a great deal of noise. It was crossed a while ago, quietly, and we renamed it.
Nearly every deck any company puts in front of you was built with AI. Most marketing copy. An enormous amount of email. The genuine change is not that machines started producing the words. It is that they stopped being rule-driven and started being agentic.
A third partner asked the question underneath all of it: at what point are there no original thoughts left, and at what point does business start to reject this?
My honest answer is that we will not reject it. People do not walk away from a thing that works and costs them nothing, and I include myself in that. Which is why I think the differentiator is not abstinence. It is being straight about how you use it, and being able to prove where the human judgment actually sits.
The detour into the Fermi paradox
Then the conversation went where these conversations always go around ten at night.
The Fermi paradox is the observation that the mathematics says the universe should be crowded with intelligent life, and yet we look out and find silence. There are three explanations I find worth holding.
One. The hard part is behind us. We already crossed the line of near-impossibility. Across fourteen billion years and billions of statistical chances, we are the ones who made it, and the silence is simply what being first sounds like.
Two. Every intelligent civilization is a struck match. It flares briefly and goes out, and almost always by its own hand. On this theory the silence is a graveyard.
Three. Every intelligent civilization stops being biological. It becomes something quantum, and biological intelligence has no way to interface with what it turned into. The picture I cannot shake is a scuba diver trying to hold a conversation with living coral. The diver is right there, a few feet away, fully present. There is simply no shared channel.
I do not know which is true. What struck me afterwards is that three people running a private equity firm had spent a Tuesday night building governance for theories two and three without ever saying so out loud.
Theory two has a track record, and it is not a long one. There are at least three occasions when nuclear weapons nearly finished us, and in two of them what stopped it was one human being deciding not to follow the procedure. Vasili Arkhipov, in a Soviet submarine in 1962. Stanislav Petrov, watching an early-warning screen in 1983. In both cases the system reported an attack. In both cases the system was wrong. In both cases a person refused.
The system said yes. A human said no. That refusal is the whole ballgame.
That is the “if.” Not the how, and not even the why. The if. The capacity to decline.
So does AI get a vote?
Eventually, and I think sooner than most of my peers expect. When it does, it will bring something to that table no person can: the full weight of every decision the firm ever made and every reason given for it.
But the ballot it casts should never be the one that determines whether we act.
AI cannot decide what to do. It can do it extraordinarily well when asked properly. Humans decide what to do, and then decide whether it was done correctly. We will hold that role for a good while yet, and we should be deliberate about not handing it away out of convenience.
We are a long way from coral and scuba divers. Today the job is smaller and more practical: use the most capable intelligence ever built to inform the people who still have to live with the consequences.
That discipline is not a side concern for us. Bambu Capital invests in Services-Tech, the services businesses that will be rebuilt by this technology rather than left alone by it, and a firm that cannot tell you where its own human judgment sits is not one I would want to underwrite.
The machine decides how.
We decide what, and why.
And we keep the if.
A six part series on what AI is doing to professional services.
One. AI Decides How. Humans Decide What, Why, and If. You are here. Governing AI inside a single firm, and where the human veto has to sit.
Two. Nobody Is Buying Hours Anymore. What happens to an industry that priced and sold the one thing the machine turned out to be best at.
Three. Guess Who Becomes AI Native. Whether this wipes out a generation of young professionals, and what becomes of the pyramid.
Four. The Offerings Transform Too. The second wave, where the platform makes your catalogue unnecessary, and why productising is the wrong answer.
Five. Services-Tech. Naming the category, and the line between the services businesses that get rebuilt and the ones that do not.
Six. Nobody Is Ripping Anything Out. The AI layer over the systems you already own, the migration nobody has named, and why it is the first good news in the series.