It is eleven at night. The board pack is due at nine. He has read the same two pages four times and the numbers still will not resolve into a decision he can defend out loud. So he does the thing he would never write down. He opens a new tab, pastes the problem in, and asks the model what he should do.
The answer comes back in seconds. Clean. Structured. Three options, a recommended one, a confident closing line that sounds like something a very expensive advisor would say. He reads it twice. It is better than what he had. Not by a little. He lifts the shape of it into his own words, closes the tab, and goes to bed slightly relieved and slightly ashamed, though he could not tell you which feeling is larger.
The next morning he walks into the room and presents the recommendation as his own. Someone senior asks him why. He gives a reason. It is a good reason. It is also not quite the reason the model gave, because he cannot fully reconstruct the reason the model gave, only that it landed cleanly at the time. Half of what he is now defending is his. Half of it came from a machine he has already closed and cannot reopen without everyone seeing him do it. If you asked him to draw the line between the two halves, he could not. He is not sure the line exists.
Nobody in the room asks. Nobody in the room knows there is a line to ask about. There is a third participant in the decision, and it has no chair, no name in the minutes, and no fingerprints on the page.
The Third Participant Nobody Introduced.
This is not a story about one anxious leader at midnight. It is the quiet operating reality of most senior teams right now, and it arrived without a memo. Somewhere between last year and this one, the model stopped being a tool people talked about and became a colleague people consult in private. The consultation is real. The colleague is uncredited.
Watch how it actually enters the building. Not through a transformation programme. Not through a licence the CIO negotiated. It enters through the individual leader, alone, under deadline, doing the sensible thing any competent person would do when a machine can hold more of the problem in its head at once than they can. Nobody signed off on it becoming part of the judgment. It simply did, one late night at a time, across every function at once, and now the organisation runs on a set of decisions that are quietly co-authored and universally presented as solo work.
The tell is the silence. Ask a senior team, plainly, how many of the recommendations that reached the last board meeting passed through a model on the way. You will get careful faces. Some genuinely do not know, because a colleague two layers down did the pasting. Some know exactly and would rather not say, because saying it feels like confessing that the judgment they are paid for was outsourced for the price of a subscription. The discomfort is the data. When a practice is healthy, people describe it. When a practice is load-bearing and unspoken, people manage the optics of it instead.
And here is what the silence costs, which no dashboard is measuring. The organisation has acquired a new and enormous input to its most important decisions, and it has no idea when that input is good. It cannot know, because it has agreed not to look. The model is in every serious room in the building, and its reliability is being assessed one exhausted individual at a time, at eleven at night, by the light of whatever bias that individual walked in with.
Deference and Reflex Are the Same Mistake.
Once you know to look, leaders sort themselves into two camps, and both camps believe they are being rigorous.
The first camp defers. The output is coherent, it is fast, it cites things, and it does not have a bad day or an agenda, so they take it. They will tell you they are simply being efficient, letting the machine do what it is good at. What they are actually doing is treating fluency as if it were correctness. The model is superb at sounding right. It is a separate and much harder question whether it is right, and the two feel identical on the screen at midnight. Deference mistakes a confident sentence for a verified one, and the confidence is free.
The second camp overrides. They pride themselves on it. They read the model's answer, feel a flicker of resistance, and reach for their gut, because their gut is what got them the corner office and no algorithm is going to tell them how their business runs. This looks like the opposite of deference. It is the same error wearing a braver face. The reflex override treats the model's disagreement as noise to be dismissed rather than a signal to be checked, which means it throws away the one moment the machine was most useful: the moment it was about to tell them something they did not already believe.
Notice what neither camp has. Neither has a rule. The deferrer takes the answer whether or not the answer earned it. The overrider rejects the answer whether or not the rejection earned it. Both have replaced judgment with a posture, and a posture is not a decision procedure. It is a mood you adopt to avoid having to build one.
The uncomfortable part is that the model does not create this problem. It reveals it. It hands a leader a confident answer, and what the leader does next exposes whether they have ever actually known how they decide. Most have not had to know. They have run on instinct refined by experience, which works beautifully right up until a machine starts producing instinct-shaped output at scale, cheaply, on demand, and they discover they have no way to tell their own considered judgment apart from a bias the machine has just handed back to them with a citation attached. The model is a mirror. Very few leaders like what is in it.
A Doctrine for Deciding With the Machine.
The leaders who will run the next decade are not the ones who use the model most, and they are not the ones who proudly refuse it. They are the ones who build a doctrine for deciding with it. Not a policy document. A doctrine: a personal and organisational answer to the only question that matters here, which is when the machine's confidence is earned and when it is just your own bias handed back to you at speed.
A doctrine starts by naming the seam. The leader who can say, in the room, which part of a recommendation is his reasoning and which part came from the model has already done something ninety percent of his peers cannot. He has made the third participant visible. He can now be challenged on the right thing. The board can interrogate the human judgment as human judgment and the machine input as machine input, instead of swallowing a blended thing that hides its own composition. Naming the seam is not weakness. It is the first act of control.
A doctrine also decides, in advance, where the machine's confidence counts and where it does not. There are questions where the model has seen more patterns than any human ever could and its answer deserves real weight. There are questions loaded with context it cannot have, politics it cannot read, and consequences it will never carry, where its fluency is precisely the thing to distrust. Knowing which question is in front of you, before you paste it in, is the whole discipline. The leader without that distinction is not deciding with the machine. He is being decided by it and calling the result his instinct.
This is the architecture SSUNDAR builds, and it is why the work has moved. For years the question was how to sharpen a leader's judgment. The question now is how to structure judgment that is already, quietly, being made with a machine in the room. That means building the doctrine into how decisions are actually taken: where the model enters, who names its contribution, which classes of decision earn its weight and which forfeit it, and how a team pressure-tests a confident answer instead of being flattered by it. It is unglamorous. It leaves no lanyard. What it produces is an organisation that can finally see the third participant it has been consulting in the dark, and can therefore govern it.
Because the alternative is already running. Right now, across the building, the most important decisions of the year are being co-authored by an input nobody has agreed to trust, assessed for reliability by tired people at midnight, and presented in daylight as pure human judgment. That is not integration. It is integration that happened to you while you were pretending it hadn't.
The model is already in every serious room in your organisation. The only choice left is whether it sits at the table with a name, or under it without one.
You are already deciding with the machine. The question is whether you have a doctrine for it, or just a posture you adopted so you would not have to build one.