AI advisory · partner level · US and worldwide
AI into production. With the cost and the output measured.
I take AI pilots and agents into production, or kill them. I prove what an AI system costs per outcome and whether the vendor's number holds. And I measure whether it raised output, with numbers a CFO accepts. Fixed scope, fixed price, the prices on this page.
Three concrete things.
Each one ends in a written decision with its numbers, not a deck. Most engagements start with the first.
The production verdict
Your pilot or agent, two weeks, one answer: ship it, fix it, or kill it — with what it costs per outcome today and at volume.
$35,000fixed · two weeks
- The eval set built from your own cases, run against the system as it is
- Cost per outcome: tokens, inference, agents, the vendor fees, at your volume
- The fix list in priority order, or the kill memo and what to keep
- If a vendor builds it: the exit, IP and handover terms to put in the contract
30% of the fee is due only when you act on the verdict.
Cost validation
What the system really costs per result, and whether the vendor’s cost claim survives your volume.
$15,000fixed · one to two weeks
- Every cost line traced: model, agent runs, retrieval, the people around it
- The vendor’s claim rebuilt from their own pricing, at your load
- The three changes that move the number most, with the saving each is worth
- A unit-cost model finance can keep
Or 15% of the saving verified over twelve months, your choice.
Productivity proof
Did AI raise output? A measurement a CFO signs, not a survey of how people feel.
$50,000fixed · eight to twelve weeks
- A staggered rollout with a control group, on the teams you choose
- The baseline signed by finance before anything changes
- Output read from the logs, never self-reported time savings
- The result with its confidence, and what it is worth a year
Self-reported savings do not survive scrutiny; a control group does.
Architect of record
While a vendor or a forward-deployed team builds, I read the design, the costs and the contract for you and sign off each gate. $30,000 for one to two weeks; $50,000 with a code review.
Operating model and roadmap
Which work goes to agents, which stays with people, what it costs, in what order, with the governance that keeps it. $90,000 to $150,000, four to six weeks.
Advisor on call
Two days a month, a second opinion within a day, and my seat in your decisions. $15,000 a month. Hands-on four days a month: $20,000. Six to eight days: $30,000.
A day, a briefing
$5,000 a day, remote; travel days at half. A half-day executive briefing on where AI stands and what to do first: $9,000, credited to a verdict.
How I work.
- Tell me the problem, not the solution. The first meeting is about what is costing you, not about which model.
- Fixed scope, fixed price. What is on this page is what it costs. I never discount the rate; I change the scope.
- A decision, not a presentation. Every engagement ends in a meeting where something is decided and written down.
- Prototype instead of argue. When two paths look equal, I build the small one on real data within the two weeks.
- Independent of the tools. Our advice and our verdicts do not depend on platform usage; any financial interest in a product we recommend is disclosed in the proposal. A verdict never recommends my own platform.
- Remote by default. US and worldwide; on site in the US, the UK and Switzerland when the work needs the room.
Builds run on Plunge AI. When a verdict says ship and you want it built, the system lives on your own Plunge AI account and platform usage is paid there, never re-invoiced by me. Our advice and our verdicts do not depend on platform usage; any financial interest in a product we recommend is disclosed in the proposal.
Plunge AI Consulting →Who is telling you this.
I founded GoLive Systems and sold it to Adobe in January 1999 (an asset purchase of about $31.0 million plus $8.0 million contingent, in Adobe’s 10-Q); GoLive was Macworld’s 1998 Editors’ Choice for web authoring, and I stayed on as an Internet-strategy consultant to Adobe. Before that P.INK (1987), after it AMANGO (2003). Today I build the Plunge AI platform, an agent runtime on Cloudflare, which is where the verdicts, the cost models and the measurements on this page come from.
Seven US provisional patent applications, 210 claims, filed August to September 2026 (nos. 64/131,227 · 64/156,763 · 64/156,775 · 64/158,642 · 64/158,644 · 64/158,648 · 64/158,650), on the speed, cost and determinism of agent runs. Provisional applications are not published, so this is a stated fact with its numbers, not something to look up yet.
Start with the problem.
Write me three lines: what the system is meant to do, what it costs you today, and when you need the answer. I reply within a working day.