Betting on AI as THE Intelligence. Here's What That Misses.

I keep watching CEOs make the same bet, loudly, with layoffs. The bet is that AI is the intelligence and the people are the trim around the edges. I think they are about to find out how expensive that is.

Two pieces of news landed in my inbox the same week. Read them next to each other and the whole shape of the moment shows up.

Brian Armstrong sent an internal memo at Coinbase. He’s “rebuilding Coinbase as an intelligence, with humans around the edge aligning it.” He laid people off to prove he meant it. Same week, Anthropic stood up an enterprise services arm. So did OpenAI. $1.5B and $10B respectively, both PE-backed. They are hiring senior operators. Not to write the code. To install the work inside companies that already bought the tech and don’t know what to do with it.

Two bets, opposite directions.

One says intelligence is the center of gravity. The other says intelligence is the input, and the work is what someone does with it inside a real company with real people, real politics, real data nobody cleaned.

Guess which one I think is right.

Humans first.

Armstrong wasn’t alone. Mark Zuckerberg told Meta employees that teams of 50 or 100 are now counterproductive when 10 can do the work.

Two F100 CEOs in one week. Same bet, different industry. If you only listened to the supply side of this market, you’d think the question was already settled.

It isn’t.

Look at how companies are actually hiring and the picture inverts. The qualities the layoffs cut, depth of research, response when challenged, the ability to defend a position, are exactly what hiring managers are now actively screening for. Polish is dead signal. Judgment is the differentiator. Some companies are doubling their product teams while others trim theirs.

Two F100 CEOs are betting the work is intelligence. The market is hiring like the work is judgment.

Both can’t be right.

What the labs just told us by what they did

The thing I cannot stop thinking about is what Anthropic and OpenAI did the same week as the Coinbase memo. They both stood up consulting arms. Same conclusion, same day, on what Sakeeb Rahman called a $111 billion industry that just became the target.

If the labs believed AI was the consulting, they would not need a services arm. They are betting the consulting still needs people, and that the people who install the work are now worth more than the work itself.

a16z put the question in print in their Speedrun newsletter that same week: which categories of work should be automated, and which categories require a person who can be looked in the eye?

Operators have been asking that question for two years. Now frontier capital is asking it on its own blog.

There was a line in that same Speedrun piece from Simo Rachidi at Safeworld that stopped me cold:

“AI gives every engineer more hands. It does not replace ownership.”

That is the entire thesis of this article in eleven words. From a YC-class founder, on a16z’s platform, in the week Coinbase shrank itself to prove the opposite. The disagreement isn’t between operators and the hype cycle anymore. It’s between the labs and the CEOs running the companies that bought from them.

Even the companies furthest along on AI deployment don’t yet know how to actually run with it. The install is the bottleneck, not the model.

What this looks like on the ground

I’ve been in conversations with operators making this bet for months now. CPOs, founders, heads of product, peer operators. Not on stage. Over coffee, on video, in the kind of call where people actually say what they think.

The pattern is the same one. Every time.

The build side got fast. The deployment side got expensive. Not in dollars. In clarity. What a senior PM does that a junior PM and an agent together can’t. The teams getting somewhere stopped pretending the model was the answer and started treating the people layer as the actual work.

Rishin Banker at Maven is one of the clearest examples I’ve seen. Same 25-person team, went from 2–3 concurrent projects to 5–6, by shrinking pods and blurring roles. The designer shipped a marketing page to production. The PMs were building real product, not writing about it. Same headcount. Different shape.

I’m running the same bet on myself. I’ve shipped three AI-native products this year, by myself, in evenings and weekends. Three different categories, same architecture underneath. None of them are the model. Every one of them is what I decided the model should do, where I drew the line, what I refused to automate.

The apps aren’t the architecture. They sit on top of one. A Chief of Staff system that does ambient context work across all my projects. Agents that handle bounded tasks, research, summarization, file ops. Skills that codify the workflows I run over and over. The apps are the visible part. The install layer underneath is what makes them all work, and what took the actual time.

The teams that compound are doing that work out loud. The teams that are stuck are deploying AI on top of last year’s org and wondering why nothing moves.

The strongest version of the other side

The honest pushback to all of this: what if AI capability outruns us so fast that Armstrong is right, just six months too early? What if the model gets good enough that the install layer collapses and the senior person becomes the trim around the edges after all?

I take that seriously. I’m just not sure it works the way the bet assumes.

The faster the capability curve moves, the more pressure lands on integration, judgment, and install. A more capable model in a company that hasn’t figured out whose job changes, who owns the output, what to refuse to automate, generates more confident wrong answers, faster. The bottleneck doesn’t disappear. It moves up the stack and gets more expensive.

If the model can do every analytical task by Q4, the question of which tasks were worth doing in what order gets bigger, not smaller.

If anything, Armstrong’s bet looks better in slow-capability worlds, where the existing org has time to absorb the tools. In fast-capability worlds the install layer is the only thing that compounds.

What the bet is actually about

When someone says they are rebuilding their company as an intelligence, what they are saying is that the inputs (the model, the data, the agents) are the company. The people are friction to be reduced.

Treating the model as the company is the same shape as treating the spreadsheet as the analyst. The tool isn’t the work.

The model is a coworker. A very fast, very confident coworker who needs a senior person around to read the room and push back when the answer is plausible but wrong. That coworker doesn’t replace the senior person. It changes the shape of what the senior person spends their time on. Less time on the formatting. More time on the call that matters.

It’s also the part the labs just bet $11.5B is worth solving.

Same question for the CPO running the product org. The CEO who keeps the senior people doesn’t win. The CEO who frees them to do the install work wins. If the work is judgment, then the question right now isn’t how many seats got cut. It’s whether the people who stayed have the room to do the part that compounds. That’s a different conversation than the one Armstrong is having on stage.

If you’re a CEO or CPO reading this, the move isn’t another lay-off announcement. It’s a meeting with the senior people you kept, asking what they could actually do if you took the artifact-theater off their plate.

The intelligence isn’t the center of gravity. The person aiming it is.