AI is killing enterprise sales.

Twenty a morning, every morning, and all of them want fifteen minutes about your Q4 priorities. Some open on the funding round. Others quote the same press release. One notes that it is slightly overcast and 66 in Chicago and hopes you avoided the rain. It was, and you did. Every fact checks out, and every sender reached for the same three gambits, the same harmless closing question, the same studied enthusiasm. They read like the same person wrote every one of them, because in a sense the same thing did.

Everyone has noticed. AI made outreach infinite and identical. Fine. It is real, and it is irritating.

It is also not the damage.

The email was never the craft

Cold email was never what made a top-tier enterprise seller. It was the cheapest part of the job, handed to the newest person on the team.

The craft began when someone answered.

A customer is about to spend six figures. The rep worth their number knows the deal is not won on a side note. It is won by getting the room to answer one question: what are the three or four things driving your business that we have to crush? The rest of the requirements list is noise, some of it written by people who will have forgotten they asked by spring.

The seller who transcribes the whole list has not sold anything. They have taken dictation.

The veto nobody designed

What used to back that refusal was not virtue.

A promise made on a sales call had to survive a gauntlet. Engineering had a backlog. Product had a roadmap fight. Someone had to explain what would get bumped to build it. Scarcity never made good decisions. It forced decisions. A one-off request had to be worth more than whatever it displaced.

That constraint is loosening. In randomized field experiments with 4,867 developers, access to an AI coding assistant raised completed tasks by 26% (‘The Effects of Generative AI on High-Skilled Work’, Management Science, February 2026). Spotify reports a 76% increase in pull request frequency and says where that leads: “As coding velocity increases, the constraints shift toward human decisions” (‘Coding Is No Longer the Constraint’, Spotify Engineering, June 2026).

The implementation cost of saying yes is falling. Scarcity supplies less resistance every quarter.

Grant the objection

The usual next move is to say AI-built software rots. Duplicated code, thin tests, a maintenance bill due in year three.

Assume that is wrong. Assume agents maintain a sprawling codebase indefinitely, better and more cheaply than your best engineers ever maintained an elegant one. Concede all of it. The argument does not need it, because the cost was never in the repository.

What the agent does not supply is the objection. David Heinemeier Hansson built a new 37signals product with AI agents and described what went missing. The old check on scope was a person: “I would’ve said no… I’d say like, ‘That’s probably not the best way to do things.’” What replaced that person does not argue. “AI is not going to ask you that question. It’s not going to push back” (‘AI challenges in software development’, REWORK, 37signals, July 1, 2026). He describes a cycle board holding about twenty items where a year earlier it would have held six.

We have written before about the model’s own missing stopping rule. This is the version one level up: the company is the thing with no stopping rule.

The bill lands on the screen

Feature #117 did not just get written. It arrived.

It is in the left nav now. Or in settings, which has three tabs where it used to have one. The admin who opens your product to change one thing walks past it every time, along with everything else that other customers asked for. Perfectly maintained, every one of them. Immaculate test coverage.

Nobody experiences your codebase. They experience the distance between what they came to do and where you decided to put it.

Capability sells. Usability is what they live in.

There is research on exactly this gap. Thompson, Hamilton and Rust found that consumers “give more weight to capability and less weight to usability before use than after use,” so they choose products too complex to satisfy them once they have to operate them (‘Feature Fatigue: When Product Capabilities Become Too Much of a Good Thing’, Journal of Marketing Research, November 2005).

That was consumer research. Enterprise buying makes the gap worse rather than better.

The person scoring the capability list is not the person who will hunt for a setting at four on a Friday for the next five years.

What gets added matters too. Fürst, Pecornik and Hoyer found that de-emphasizing feature heterogeneity – capabilities drawn from unrelated domains – avoids low expected usability, while features built to work together raise expected capability (‘How product complexity affects consumer adoption of new products’, Journal of the Academy of Marketing Science, 2023). Deal-by-deal accumulation tends toward exactly that heterogeneity, because each request gets judged against one account instead of the product.

“We’ll put it behind a flag”

The real objection is better. If building is cheap, so is reversing. Flag it, ship it to 5%, measure, kill it if it does not work.

That is not what a flag usually is. A flag is how a product argument ends without anyone conceding. Fine, we’ll flag it and test it. Then the build runs two weeks late. Someone points out the thing looks fine. It goes to everyone, and the switch that justified shipping is still sitting there in the dashboard.

A 2026 study tracked over 4,000 toggle events across Kubernetes and GitLab and found toggles “often persist beyond their intended lifecycle, accumulating as technical debt” – removals lagged additions by roughly 35% and 13%, “leading to growing toggle inventories.” Median lifespans were 734 days and 185 days (‘Feature Toggle Dynamics in Large-Scale Systems’, arXiv, April 2026). Those are large, mature codebases run by disciplined maintainers. Even there, flags linger.

Uber’s answer was to build a tool that deletes stale flag code and assigns the diff to whoever created the flag. It generated cleanups for 1,381 flags, 17% of the total, and 65% landed without changes (‘Piranha: Reducing Feature Flag Debt at Uber’, ICSE-SEIP, 2020). Retirement happened at scale when a machine opened the pull request.

So the exit is real. It is just exercised a year later, long after the thing became furniture. The hardest thing to remove was never the code. It is the promise.

Once a customer has bought on a capability, deleting it is a commercial negotiation, not a pull request.

What is actually being lost

The AI-written email is the visible degradation. It ruined a task that was already the least valuable thing in the funnel.

The consequential loss is the seller’s ability to force a decision – to tell an important customer no. What they asked for is not what they need. The deal turns on the three capabilities that decide whether that customer wins, not the eighty that decide nothing.

Scarcity was never that craft. It was the backing for it. Limited engineering capacity gave a disciplined seller leverage: if only three things could be built, the buyer had to say which three. It also capped the damage an undisciplined one could do. Cheap implementation removes both the leverage and the cap.

Five excellent things become a thousand mediocre ones, one reasonable-sounding yes at a time.

Engineering used to say no for you. Now someone has to do it on purpose.