Research

The Press-Send Problem

AI made production nearly free. The last mile got more expensive.

By Merryl DMello, DMello Advisory · 4 min read

The claim: in a one-person business armed with AI, the binding constraint is no longer the ability to produce. It is the decision to release. And AI makes this worse, not better, because it drives the cost of building toward zero while leaving the psychological cost of sending exactly where it was. The result is a new failure pattern: excellent, finished, unsent work piling up at a rate manual work never produced.

This is argued from a receipts-first position: a six-month analysis of one operator's complete working corpus, 8,945 prompts and every finished artifact accounted for.

The inventory

In one six-month window, the same solo operator's corpus shows, verbatim from the analysis:

That is 91 rendered, caption-ready videos without a home, and a revenue scoreboard reading zero conversations, zero euros, while all of this existed on disk. The operator's own portfolio document contained the rule "nothing new gets built until 20 logged conversations." The rule was written, and then 30 more episodes were built.

Every item in that table is genuinely good. That is not a defense, it is the mechanism.

Why AI makes it worse

Before AI, the pipeline itself rationed output. Producing 40 episodes manually would have taken so long that distribution kept pace by default. AI removes the ration. When production capacity goes from two units a month to two a day, the queue in front of the send button grows by two orders of magnitude, and the send decision is still made by the same one person, one click at a time.

The corpus shows the psychological signature clearly. Building feels like progress because the output is excellent; every session ends with something real and demonstrable. Sending feels like exposure because it is: a posted video can fail publicly, an unsent one cannot. So the work accumulates in the one place where quality is total and risk is zero, which is exactly the place where it generates nothing. Two work blocks a day, nearly every day for six months, produced a machine with a complete pipeline whose only unautomated component was named in the analysis, unsparingly: the human clicking post.

There is a second-order cost. Perishable assets decay in that queue: World Cup-tied episodes lose their window weekly. Unsent referral packs go stale while the referral is warm. The queue is not neutral storage; it is a slow leak.

What actually closed the gap

The same corpus, two months later, shows the fix, and it was neither discipline nor motivation. Shipping resumed when the operator stopped treating release as founder work and engineered around their own bottleneck:

• publish-gate checks wired into the daily automated brief, leading with two numbers before any build status: conversations logged, and finished-but-unsent artifacts
• a standing rule that no episode renders until the channel account exists and a posting slot is booked
• distribution tasks handed to the existing delegated team channel, so posting stopped waiting for the founder
• time-boxed budgets for creative output, so production could not silently absorb delivery hours

After those gates went in, the same analysis round records live deployments, launched sites, and posted episodes, where the previous round recorded a launch pack slipping four days in a row. The behavior did not change because the person changed. It changed because the system started counting unsent work out loud and someone else was assigned the button.

The general principle

Any owner adding AI to a small company should budget for this explicitly: AI multiplies your production function, and unless you build a release function to match, the surplus piles up unreleased. The question to ask is not "what can we produce faster?" It is "what is our send rate, who owns it, and what counts it when we fail?"

What would change my mind

• Evidence that high-output AI operators do not accumulate finished-but-unreleased work, in any published corpus. Personal reports all point the other way; nobody publishes their unsent queue, which is itself telling.
• A case where the bottleneck was genuinely production (not distribution) after AI adoption, sustained for months. One-off examples exist; a sustained pattern would blunt the claim.
• Data showing the fix is motivation rather than mechanism. The corpus shows gates working where exhortation did not, but it is one operator.

The author runs dmello-advisory, helping small and mid-size companies adopt AI without building their own unsent queue: working systems, fixed prices, and a working-software guarantee. Eighteen years building and leading products at Intel, HP, Rackspace, BioNTech, and Phoenix Contact. dmello-advisory.com.

All research papers

Last updated: 16 October 2026