Signal Archive — What I See Inside AI Deployments

Field notes

A support deployment priced at $0.04 per ticket carried $1.80 of human review per ticket in its first month.
Review time fell 70% once the team logged which answer types were actually being corrected, instead of sampling at random.
No finance team in five engagements had the review cost in the business case.

Analysis

Put the human cost of being wrong in the model before you sign anything. A workflow with a cheap token bill and an expensive correction loop is a worse deal than the reverse, and it is the one that quietly stalls at month three.

What I see inside AI deployments

All issues

Get the next issue on Sunday.

What I see inside AI deployments

All issues

Issue No. 6 · 20 September 2026The cost line nobody forecastsInference is rarely the expensive part. Review, correction and escalation are — and they land on a team that never signed up for them.−

Issue No. 5 · 13 September 2026Why the pilot passed and production failedPilots are run by the people who want them to work. Production is run by everybody else.+

Issue No. 4 · 6 September 2026One named owner, or nothingThe fastest predictor of a deployment surviving its first bad week is whose phone rings at 2am.+

Issue No. 3 · 30 August 2026Six weeks to twoA teardown of a cycle-time reduction that had almost nothing to do with the model choice.+

Issue No. 2 · 23 August 2026What a growth system actually doesNot campaigns. A growth system is the set of decisions you no longer have to make by hand each week.+

Issue No. 1 · 16 August 2026What I see inside AI deploymentsWhy this briefing exists, and the one question I ask in the first meeting of every engagement.+

Get the next issue on Sunday.