July 2026 · 10 min read
Venture Kills Its Losers. Your Org Can't.
Parallel AI bets are the right instinct under uncertainty. But exploration only works when bets are cheap, independent, and allowed to die. Inside most organizations they're none of the three, and the refusal to converge is what kills the AI mandate.
Fifty threads. Or two hundred, pick your number. Every organization I’ve been near this year is running some version of the count, and the count is the culture. It looks like agility right up until you ask what shipped.
In most rooms I’m in, the number of active AI initiatives is out of proportion to the number that have reached anyone outside the building. Every function has a pilot. Every leader has a demo. The standups are full of movement. Little of it converges on anything a customer or the P&L ever touches. Last year MIT put a number on where that ends: 95 percent of enterprise generative AI pilots deliver no measurable P&L impact. The velocity is real. The direction is not.
Starting fifty things is the right instinct
Starting a lot of bets is not the mistake. When nobody knows what will work, and in AI nobody does, spreading small wagers across many ideas is the correct response to uncertainty. It’s how research works, how venture works, how you find the thing worth building when you can’t reason your way to it in advance. If I told you to pick the one AI initiative that matters and pour everything into it, I’d be asking you to be certain about the exact thing the whole situation is uncertain about.
But the portfolio logic holds on two conditions. The bets have to be cheap, and the bets have to be independent. Venture meets both. Each startup has its own team, its own runway, and a default of death when the runway ends. Most organizations meet neither, and that difference is the rest of this essay.
The portfolio is fake
Threads are free to start inside a company. Nobody prices them. A leader spins one up with a kickoff and a channel, and from that moment it exists, consuming attention and waiting on resources. What threads are not is free to run, because they all draw on the same scarce thing: the handful of people who can move AI work end to end.
Watch where the work routes. Interns drift off their own projects and onto those few people’s work, because that work is the only work going somewhere. The few are resourcing parallel projects while trying to keep core delivery alive. Everything is close to shipping, but the business is impatient, so instead of finishing one thing and moving to the next, they swirl across a dozen. It’s a meat-grinder approach to delivery: throughput instead of intelligence.
Which means the fifty bets were never independent. They share a bottleneck, so they are one overloaded effort wearing fifty names. The portfolio was fake from the first kickoff, and the funnel is the tell.
The other tell is that nothing gets to done. Done requires someone to say this one first, that one later, and the rest not at all. In a funnel no one says it. Work keeps arriving at the same few people, who keep all of it half-alive because dropping any of it means a conversation with whoever asked. The queue never resolves. It gets churned. And churn reads as progress right up until someone asks what shipped this quarter and the room goes quiet.
Convergence is built into venture
The venture comparison cuts deeper than the people making it intend. Venture converges without a boss. No one walks into a startup and ends it. It starves, because the money is finite, the bet has to pay its own way, and death is the default every founder is racing. Exploration works there because convergence is built into the economics.
Inside a company the economics run backward. Threads cost their sponsors nothing, survival is the default, and there is no market signal to starve the weak ones. You can try to rebuild the scarcity, and pricing threads honestly is a real lever: make every initiative carry the loaded cost of the people it draws on, and watch half the portfolio become indefensible at the next budget review. But internal scarcity is administered by the same politics that created the sprawl. The leaders who started threads to show they’re doing AI will defend those budget lines the way they defend headcount, and the review becomes theater. Scarcity helps. It doesn’t decide. Someone still has to.
The real problem sits higher
That someone is missing, and the absence is a leadership vacuum. Nobody is willing to have the hard conversation about how AI gets organized, which means nobody is willing to tell four teams their bet is over.
Everyone has an AI title now. Nobody has real authority. That combination is the whole problem. You can staff a Head of AI, a VP of AI, a Center of Excellence, and three AI product leads, and still have no one who can walk into a room and end an initiative. Titles got handed out faster than mandates did, because a title is cheap and a mandate means someone else loses the ability to spin up their own thread. The org fills with people accountable for AI and empowered to decide nothing.
Alignment is impossible when power dynamics matter more than problem-solving. A room full of people with the word AI in their title and no authority to concentrate resources will start threads forever and converge on none of them, because starting a thread is the only move available to someone who wants to show they’re doing AI and can’t direct anyone. The missing person is not another strategist. It’s someone who can kill a thread, in writing, and make it stick.
This is harder than naming it. The power dynamics that stopped anyone from having authority don’t dissolve because I’ve pointed at them. Convergence rights have to be granted from above, by a sponsor willing to spend the capital to take them away from a dozen people who currently enjoy not having a boss. If that sponsor doesn’t exist, none of what follows happens, and the fifty threads are the equilibrium the organization prefers. Naming the vacuum is the easy part. Someone with real standing has to decide to fill it.
Brooks saw the mechanism
The cost of all this was named fifty years ago. In The Mythical Man-Month, Fred Brooks observed that adding people to a late project makes it later, and he was careful about why. Part of it is communication overhead. Two people talking is one conversation. Four people is six. Eight is twenty-eight. The channels grow with the square of the headcount, so each new person’s context has to be explained by the people who already understand the system. The cost of help is paid by exactly the people who had no spare capacity to begin with.
But Brooks named a second reason that matters more here: some problems don’t partition. As he put it, some tasks are like bearing a child, nine months no matter how many people you assign. When work can be cut into clean pieces with defined boundaries, more people helps, because the interface does the talking. When it can’t, adding people multiplies the coordination and moves nothing. A late project with a clear spec can absorb people. A fuzzy problem cannot, no matter how much time is on the clock.
That is the connection to AI, and it is Brooks’s point, not a correction of it. An un-converged AI bet is the least partitionable work there is. The goal shifts as you build, the tasks are tightly coupled, and no one can hand off a chunk because the chunk depends on four decisions nobody has made yet. Add a body to a bet like that and you add drag.
The fifty-thread version of the cost is one Brooks didn’t need to name, because he was writing about one project at a time. When fifty threads pull on the same few heads, the price is paid in switching. Every context change dumps one half-loaded problem to page in another, and the deepest work gets the most fragmented attention. Within a bet, communication tax. Across the bets, switching tax. Converge and both fall at once. The survivors become definable enough to divide up, and the people carrying them get to hold one problem long enough to finish it. Definition is not the thing you do before you staff. It is what convergence buys you.
You’re not adding capacity. You’re adding tax.
What structure looks like
Structure here is not more process, and it is not a list of parallel fixes. It’s one decision with consequences: someone gets real authority over the AI portfolio, and everything else follows from whether that happened.
The someone is the role I’ve argued the AI-era org is still missing: the person who holds the map, sees the collisions before they ship, and decides what gets built and where. Not a title over the portfolio. The power to concentrate resources and end the bets that aren’t paying, publicly, in a place the decision survives.
Once that person exists, the rest has teeth. Threads get priced, so the portfolio faces the loaded cost of the attention it consumes and internal scarcity has an enforcer instead of a theater. Bets get gates: a date, named in advance, when the bet is re-underwritten. Not a mechanical checklist, because the goal of a nebulous bet shifts as you build, and criteria written a quarter ago can be aimed at a stale outcome. The question at the gate is the underwriting question, whether you would fund this today knowing what you now know, and the portfolio owner asks it, because gates written by thread owners come out soft. That is the honest version of the prioritization work: committing in advance to look, and tying the survivors to a number the business already watches. And the work gets organized around clear roles and ownership, so that once a bet survives, adding a person means adding a defined responsibility instead of a twenty-ninth channel.
The order is the point. Explore many, converge hard, staff the survivors. The fifty-thread culture does the first step forever and never reaches the second, so the third lands on work that was never made partitionable and turns into drag.
If you’re the funnel person, with no authority and no sponsor in sight, you still have one move, and it works the same way a real no works: specific and public. Publish the queue. Every thread you’re carrying, in one place leadership reads, with what each one is displacing and a forced choice: this one first, that one later, or overrule me in writing. You can’t converge the portfolio from below. You can make the refusal to converge visible, and visible refusals have a shorter shelf life than invisible ones.
The bill for skipping convergence never arrives as a failure. Nothing crashes, nothing is formally cancelled, the demos land and the standups stay busy. It arrives as a year spent, fifty things at eighty percent, and nothing compounding. Then it arrives a second time, worse. Leadership doesn’t conclude the organization failed to converge. It concludes AI doesn’t work here, and the budget moves on. That is what never converging is fatal to. Not the company. The mandate.
None of that is a framework. It’s the willingness to let most things stop so a few can finish, and someone with the standing to enforce it. That willingness is exactly what the leadership vacuum is missing, which is why the fix is a person before it’s a process.
Speed without structure doesn’t deliver faster. It delivers slower, noisier, and at a higher human cost.