The Week We Stopped Being the Bottleneck

A case study from inside our own operating group, with the real thresholds and the real mistakes left in. Steal anything useful.

The situation

Our operating group runs on an AI executive team. Six seats, each owning one question. The COO asks "did every machine run?" The CFO asks "what did it cost, and was it worth it?" The CRO asks "what is the top revenue action right now?" Each seat reads its own data, writes its own status file, and reports up.

It worked, with two problems.

First, the system was asleep most of the day. Every seat ran as a scheduled job: wake at 7:30, do the pass, die. A website that went down at 8:00 am stayed invisible until 7:30 the next morning. A winning ad that hit its daily budget cap before lunch lost the rest of the day's sales before anyone noticed. (Illustrative, but both are gaps we measured in our own monitoring windows, up to 23.5 hours of blindness.)

Second, and worse: the founder was the bottleneck. Every decision the system surfaced, from a $3 analysis run to a $500 commitment, landed in one inbox and waited. When the founder didn't answer, everything downstream of that decision stopped. We had automated the analysis but not the authority to act, so every workflow still ended at one person's inbox.

The decision

We rebuilt the system around one idea: act by default, escalate by exception. That took three layers.

Layer A, a heartbeat. One scheduled task fires a small script every 2 minutes. It wakes, reads the state of the world from files on disk, runs whatever work is due, and exits. Each run ends cleanly, and a separate watchdog checks every 10 minutes that the heartbeat itself is still alive, so a dead process cannot sit there looking idle. AI cost only happens when there is real work. Most ticks cost nothing.

Layer B, a delegation ladder. Every decision the system can make gets scored on three tests: can we undo it, how much money is at stake, and how confident are we. We turned those tests into four written permission levels:

Tiers 1 to 3 cover spending inside accounts and budgets the owner already approved. Tier 4 is anything that moves money out or commits the company: those never automate, at any dollar amount.

Two rules carry the ladder. A decision type the system has never made before starts at "asks first" and earns its way down with a track record; nothing promotes itself. And an unanswered request never counts as a yes. Tier 3 requests expire to denied after 72 hours. Silence never publishes anything in the founder's name.

Layer C, a reconciler. This is the layer that makes the other two safe. Every 30 minutes, a script compares what our ledger says happened against what the outside world says happened: the ad platform's actual spend, the actual API bill. Any unexplained gap bigger than $10 or 5 percent of the day's tracked spend, whichever is greater, freezes that category back to owner-only. The rule we locked: no category gets silent authority unless the reconciler can actually check it against live data. Verification here is not a value we hold. It is a job on the schedule.

What almost broke

Our first draft had the safety alarm and the spending wall set to the same number, which meant the alarm could never sound before the system hit the wall. A second review pass, run by a different AI model with no access to our reasoning, came back with 48 findings. Two were blockers. The most useful one: an undo instruction is not a rollback. If your undo doesn't first check that the world is still in the state you left it, it can overwrite a fix somebody made by hand in the meantime. We now verify before every undo.

Then the same lesson bit us the same afternoon, in miniature. A report we regenerated silently failed to overwrite the old file because it was open in a viewer, and our script reported success anyway because it only checked that a file existed, not that it was the new one. Twenty minutes after writing "verify actual state, don't assume it" into our own doctrine, we got caught assuming. The check is now on content, not existence.

The numbers so far

Designing this took one working day and under ten dollars of model spend, priced through a cost gate that estimates every heavy job before it launches. Honesty requires this next part: the system is new, so there are no victory-lap results yet. It is running in report-only mode right now, logging every decision it would have made without acting on any of them. Our own launch rule: a category earns live authority only after 100 logged shadow decisions over 30 days with no unexplained gap between ledger and reality. We will publish what that window actually showed, including the misses, in a future issue.

Steal this

The ladder works on a human team with zero AI. Here is the blank version. Fill in the dollar lines, put it where your team can see it, and make the log non-optional:

Three rules make it hold:

1. Score decisions, not requests. Can it be undone, what does it cost, how sure are you. Anything that passes at a low tier should not be interrupting you.
2. Silence is never approval. Unanswered requests must expire to "no." The moment silence can approve, you have built a machine that acts in your name while you sleep.
3. No silent authority without verification. If you cannot check a category of decision against outside reality, don't delegate it silently. Build the checking first, then delegate the deciding.

Writing the dollar lines down removes routine approvals from your inbox and shows your team exactly when they are allowed to move. In our conversations with founders, written authority limits are rare; the table above takes about five minutes.

Built inside the Scipion operating group. If you want help setting authority and verification rules for your own operation, reply to this email.