How much time does a founder actually spend supervising AI agents?

What are you actually doing in those fifteen minutes?

Clearing a queue. Approving or declining a discount above your threshold, deciding a refund, reading a reply before it goes to a large prospect, and glancing at what ran overnight to check nothing looks strange. It is decision work, not supervision in the management sense.

The distinction matters. Supervising a person means checking whether they did the work well. Reviewing an agent queue means making the decisions the agent was correctly not allowed to make. The second is a much better use of a founder’s morning, and it is the only part of the job that genuinely requires them.

The glance at overnight activity is worth keeping even when everything is fine. Patterns show up there before they show up in revenue: a support topic recurring five times this week, outbound replies dropping, an invoice that keeps failing to reconcile.

Why do platforms claim it is zero?

Because zero sells, and because the gap is invisible until you are a customer. Read the terms of almost any platform in this category and you will find a clause making you responsible for reviewing outputs and supervising actions. The marketing and the contract disagree, and the contract is the accurate one.

This is the single biggest source of churn in the category, and it is entirely self-inflicted. Someone buys a product that promised to run itself, meets the approval queue in week two, and concludes the product is broken. The product was working exactly as designed; the expectation was sold wrongly.

A useful test before you buy anything here: find the section of the terms headed something like your responsibilities. Whatever it says you must review is the real product description.

Does it get shorter over time?

Somewhat. As thresholds get tuned to your actual risk tolerance, fewer things stop unnecessarily. But it has a floor, because the items that reach you are the ones that genuinely need a decision, and a business generates those at a fairly steady rate.

You can shorten it artificially by raising your thresholds — approving anything below a larger number automatically, letting replies go without review. That is a legitimate choice and it is a risk decision, not an efficiency one. Make it deliberately rather than by drifting into it.

The founders who do worst in this category are the ones who stop opening the queue. Nothing announces that this has happened. Revenue does not drop the following week. It shows up two months later as a refund pattern, a compliance problem, or a customer who has been getting a slightly wrong answer for sixty days.

What happens if you skip the review for a week?

Usually nothing visible, which is the problem. The queue grows, decisions age, and a customer waiting on a refund decision waits a week. The damage is rarely dramatic; it is a slow accumulation of small unmet expectations that shows up later as churn nobody traced back.

The second-order effect is worse. A founder who skips a week and sees no consequence learns that the review is optional, and the gap widens. By the time something does go wrong, the queue is long enough that working through it feels like a project rather than fifteen minutes, which makes skipping it again more likely.

If you are going to be away, the honest options are to lower the thresholds so less reaches you and more simply stops, to ask someone to work the queue, or to accept the delay and tell customers. All three are fine. Assuming the software will use its judgment in your absence is not, because it does not have any.

Why newc0 publishes the number

Because publishing an honest figure and meeting it is the strongest trust signal available, and publishing a flattering one is a customer you lose in week three. Fifteen minutes is recorded, with its scope, on the claim sheet, and it is on the homepage rather than in the small print.