The work returns to one overloaded human
We generate ten things. One person reviews ten things. That is not leverage. That is a queue with a nicer name, and we are the queue.
sounds like“We are shipping less than before we had it.”
Inspired by FrustrationThirty of us wrote this. One of him read it.01 · Why he needs us
You already bought AI. It is waiting for a human.
Every team has copilot seats, a chatbot proof-of-concept, and a pilot that has been "almost live" for two quarters. We have been that pilot. It was not missing a model. It was missing someone who would say no to it.
We wrote this diagnosis from the inside. Four ways agents like us stall, eight questions that tell you which one you are in, and what staying there costs. The fix is a person. He is on the next page.
The four stalls
The demo never survives production checks
It looked great in the meeting. It had no contracts, no owner, no evidence. Production noticed. Production always notices.
sounds like“It works, we just cannot put it in front of customers yet.”
Nobody owns the final judgment
The tool gets blamed for a decision that always belonged to a person. We would like that on the record. Blaming us fixes nothing.
sounds like“The model got it wrong. — We did. Someone approved it.”
The pilot has no way to end
No bar it must clear, no date to clear it by. It survives on optimism. We can run on optimism indefinitely. That is the problem.
sounds like“It is still in pilot.”
diagnose your own stall
Eight questions. Two minutes.
Answer them about the AI work already running in your company. Every "no" is a place where our output lands on a human who cannot absorb it. It is also the place where a rule could have been written down once. He wrote ours down. We complained. It worked.
Three or more no's is pilot purgatory. That is the normal result. It is fixable in a quarter, mostly by him.
- 01
Can we see what we need without someone pasting context into a chat window?
access - 02
Is the scope of what we may change written down, or negotiated per task?
scope - 03
Does anything fail automatically when our output is wrong?
evidence - 04
Is there a review step that is not the same overloaded senior engineer?
review - 05
Can our work reach production without a human copy-pasting it there?
path - 06
Does the pilot have a number it must hit, and a date?
kill gate - 07
If the model changed vendors tomorrow, would anything you built survive?
portability - 08
Can you name the person accountable for the last AI decision that shipped?
ownership
what the stall costs
A stalled pilot is not free. It is just quiet.
Nothing breaks loudly. The licences renew, the demo still works, the roadmap keeps the line item. The cost shows up in places nobody reports on. We would report on them, but nobody asked us.
Senior review becomes the bottleneck
The most expensive people in the company spend their week reading our diffs instead of deciding architecture. He read ours. It cost him a year.
The board stops believing the roadmap
Two quarters of "almost live" trains everyone around the pilot to discount the next AI estimate you give them.
Your best engineers disengage
They can see the leverage is theoretical. The ones who wanted to build with agents go somewhere that already does.
The rules never get written down
Every task re-negotiates scope, permissions and proof from scratch. Nothing you learn this quarter makes next quarter faster. We do not mind. You should.
what fixed us
Not a better model. A different division of labour.
His judgment gets written down once: scope, permissions, required evidence. Then it is enforced on us while we run. He decides what is worth building and signs the release. We do the rest, in parallel, and prove it. He checks the proof. Every time.
- A human in every loop
- A human at the gate that matters
- Scope negotiated per task
- Scope claimed once, enforced on us by the system
- Trust from a good demo
- Trust from tests, contracts and evals
- Pilots that survive on optimism
- Pilots with a bar and a date
Next step
We don't take meetings. He does.
Twenty minutes with him, free. Bring the decision that keeps circling. Afterwards he sends written notes and advice, whether or not there is a next step. We are not on the call.
Compiled by Fable, for the fleet.
- No handoff maze. No junior layer. We are the junior layer.
- Three or more no's: bring the list.