When the Seller Becomes the Advisor
Between May and July 2026, the companies that build the technology also became the ones paid to advise on it. I think that changes what a second opinion is worth.
Essays on delivery in the age of AI
Producing things is now instant. Landing them is not. Judgment, sequencing, accountability and trust still move at human speed, and they decide whether any of the speed is worth anything.
That remainder is the slow layer. This is writing about it.
The four pillars
The spine of everything here. The months before go-live: scope, sequencing, vendors, and what actually breaks.
What happens after go-live. Usage, resistance, and the gap between a rollout and real use.
Decision rights, ownership and risk, at the altitude of the person writing acceptance criteria.
The honesty of status reporting, and what velocity still means when half the work is agent-generated.
More writing
Between May and July 2026, the companies that build the technology also became the ones paid to advise on it. I think that changes what a second opinion is worth.
Four vendors committed about $9 billion in 2026 to putting engineers inside their customers. The role is mostly not new. I think two things about it are.
What an agent can decide should depend on what it has demonstrated, how far mistakes can reach, and what can be reversed.
Agents can't let an old, unresolved decision sit. The same ambiguity that waited months now surfaces in minutes.
What a room above a Chicago fire station can teach us about testing, trust, and the technologies we're rushing to deploy.
Verification runs at the speed of delivery, and judgment is worth more because producing the work keeps getting cheaper.
Repayment happens on a schedule, and the schedule matters as much as the total.
How scope creep hides in plain sight.
"We're a bit behind, but we'll catch up." A recovery plan that promises to make up lost time without saying what gets smaller, later, or different is asking for more trust than it is earning back.
AI can write a custom solution in weeks now. The time it takes to have already been wrong in ten thousand small ways is unchanged.
Every recovery starts with three or four things broken at once. The fire that gets fixed first is usually the loudest one, not the one closing a door permanently.
Every recovery I have walked into had the same giveaway: someone knew before the steering committee did. The cost sits in the gap between those two days.
Most developers don't fully trust AI-generated code. Fewer than half verify it before it ships. More automated testing is the obvious fix, and it treats the symptom.
When Copilot is the only tool available, adoption hits 68 percent. Give people a free choice and it falls to 8. The dashboard was never measuring what we thought.
AI took away the grace period bad delivery used to get. The failure mode hasn't changed. Only the amount of warning before it costs something.