Time
What I measure is time on my turn: the stretches in which the machine is waiting for me. They come from the assistant's session logs. Each break counts for at most 30 minutes, even if I was away longer.
What no log captures is looking: watching clips, listening to cuts. For that I add a flat 50 percent. For the audio feature, that is listening time.
Machine time is listed separately. It covers every stretch in which the system is working, including waiting for render jobs. When several agents run in parallel, the time counts once.
AI cost
The costs in the reports are AI costs only: credits, generations, plan shares. Staff costs are not included. My work appears as time, not in euros. The one exception is the “In everyday business” calculation in the report on the Kroll quarterly meeting.
I convert image and video credits at list price: 1,000 Higgsfield credits are 49 US dollars. ElevenLabs counts at the stated price per generation.
The assistant runs on a Claude plan at 200 US dollars a month. Its share is calibrated against the measured weekly allowance. Next to it stands what the same tokens would have cost at the API list price of Claude Opus 5.5.
Subscriptions without a unit price, such as Suno, appear as “in the plan”. All amounts are in euros, at a rate of €0.86 per US dollar.
Waste
Waste is the share of credits that does not end up in the finished result: discarded clips, model tests, rework. In THE AGENCY, only 37 percent of the image and video credits made it into the film. At Kroll, 48 percent went into the search for good German.
It is the most uncomfortable number in every report. It shows where sorting out costs money.
Division of labor
Every decision is assigned to one side: human or machine. At the end of each report, both columns stand side by side. That shows where the judgment lies.
Checks
Before I sign anything off, a script checks it. Speech recognition compares every line with the script; phoneme recognition checks names. Loudness, frame boundaries and the offset between lips and sound are measured as well.
The check runs before the human. It cannot replace them.
Why this way
Every number comes from a log, none from memory. What has not been measured is not shown. Where I estimate, it says so.