The method · business first, technology last
Size the problem. Find the value. Pick the lightest tool.
Most AI programs fail because technology was the first decision instead of the last. My method runs the other direction: price the problem, locate the value, then walk down a ladder that starts with deleting the task and ends, only when earned, at agents. Sometimes the honest answer is a spreadsheet and a deleted step.
1 · Size the problem
Decompose the work to activity level and price the baseline: volume times frequency times unit time times loaded cost, plus error rate times error cost. If the number is small, stop here and do it manually.
2 · Find the value
Locate the value pool before picking any technology. Count beyond labor: error reduction, cycle time, and redeployed capacity are often a third to half of the true return. Sized and realized stay separate columns forever.
3 · Design the solution
Only now choose the intervention, and choose the lightest one that clears the bar. Compare against the baseline on accuracy, speed, and total cost of operation including the maintenance nobody budgets.
The ladder · six rungs, cheapest first
Every rung must fail before the next one is allowed.
Eliminate
Should this work exist at all? Question the requirement, delete the step, add back only what reality demands.
Canon: Musk's algorithm via Isaacson; lean waste doctrine; the ESSA sequence Shell ran in finance
The best automation I ever shipped was a deleted report.
Redesign
Fix the process before the tooling. Automation applied to an inefficient operation magnifies the inefficiency.
Canon: Gates' first rule of technology; Bain's simplify-eliminate-automate lens; McKinsey's redesign-first doctrine
The $20M turnaround was an operating cadence before it was any system.
Rules
Deterministic work gets deterministic tools. Codifiable logic with stable inputs wants scripts and workflows: cheap, fast, auditable, no hallucinations.
Canon: Google's Rules of ML, rule one: don't be afraid to launch without machine learning
My weekly client deck pipeline runs with zero AI calls and has not missed a Monday.
Analytics
If the pain is knowing rather than doing, instrument the process and expose the data. The dashboard is often the whole fix, and it audits your data for any future model.
Canon: Measurement-first doctrine from the same Google canon; every baseline requirement in the AWS ML lens
Cost and quality per run sit on one scoreboard across my whole platform.
Model
Structured data predicting a defined outcome, with history to learn from and patterns too complex for maintainable rules: classic supervised learning, not GenAI.
Canon: Google Cloud's published decision tree; AWS guidance to baseline with rules first; the hidden-technical-debt literature on true ML cost
The buy-fit ranker across 157,000 companies is an embedding model, because ranking is prediction.
Agent
Language, judgment, and open-ended paths, last and governed. Escalate single call to workflow to agent only as the task demands it, with budgets, evals, and a human on irreversible actions.
Canon: Anthropic's Building Effective Agents: the simplest pattern that works; agents trade cost and latency for flexibility
Twenty agents run my four business units, and autonomy is a dial I turn up with evidence.
Bring a process. Walk the ladder yourself.
Six questions, and the honest output is sometimes that you should buy nothing from anyone. That is the point.
The advisor · six questions, honest answers
Question 1 of 6
If this task vanished tomorrow, what breaks?
The deletion test. Most processes carry steps nobody can defend.
Fifteen minutes with your hardest process beats any framework page.
What does the machine this method built look like?
The architecture →