The step zero of AI adoption

Last week Boris Cherny, the creator of Claude Code, posted a table called Steps of AI Adoption: how an engineering team goes from zero to running agents at scale, with the agent count multiplying by ten at each step

.

Step 0, nothing runs. Step 1, one agent and you check every change. Step 2, ten agents and you review the diffs. Step 3, a hundred and Claude writes nearly all of it. Step 4, a thousand run and you steer by intent, stepping in only on the exceptions. What rises across the steps is how much judgment you hand over, and you hand it over once the agent can check its own work.

Engineering teams get there fast for one reason: the agent inherits a record it can read. The code, the history, the tests, and the review comments already describe how the software gets built, so the agent reads it and knows the job.

The same week, Replit’s CEO Amjad Masad published The Self-Driving Company. Over six months his engineers nearly tripled their code output (yes, lines of code is a vanity metric).

What matters is how they got there. Before the agents were useful, Replit gave them context across every system it runs: the data team put a semantic layer over the warehouse so the agent knew which tables were the source of truth, support gave it playbooks and escalation rules, and sales gave it internal account knowledge. Amjad’s line: “With context across systems, we saw a leap forward in productivity.” And what it did to his people, in his words: they “don’t feel like they’ve been automated. They feel like they’ve been promoted.”

Now look at most large enterprises. Copilot seats bought and idle, pilots that stall at the demo, a leader who reads Replit’s numbers in the morning and stares at a flat dashboard in the afternoon. That experience is real, and the AI really is underperforming. The model is fine and the budget is fine. What is missing is the record.

Ask an agent to clear failed payments in your ERP. Where does it read how? The SOP covers the happy path. The system of record logs the outcome and leaves out the reasoning. The exceptions that take real judgment live in the heads of the two or three people who handle them every day. In issue #1 I put a number on it: about 70% of operations cases are edge cases. The agent has nothing to read, so every edge case is a surprise, and every surprise reads as “the AI doesn’t work.”

So the enterprise version of Boris’s table needs a step engineering never had to take. Every step he lists assumes a codebase the agent can already read, and operations has nothing like it. Before any of it can start, someone has to capture how the work actually happens: every process, every exception, every escalation, every unwritten rule, somewhere an agent can read it.

There are two ways to get one. By hand: interview every operator, workshop every process, document every edge case, then keep it current as tariffs reroute a shipment or IT swaps a tool. That gives you a clean snapshot that ages from the day it is written.

Or capture it from the work: software that watches how the job is actually done, writes the record from what it sees, and keeps it current as the work changes. That is what we build at WIQ, on top of the stack you already run. We call each entry a blueprint, and a blueprint is to operations what a commit is to code.

None of this is new. When factories first electrified, the winners rebuilt the whole operation around the new power source, while the ones that stalled bolted an electric motor where the steam engine used to sit and wondered why output barely moved. Replit rebuilt. Most enterprises are still plugging an agent into a process it cannot read, and the flat dashboard that follows just means nothing behind it changed.

Your record’s first entry takes about ten minutes. Record one task with the app from issue #1 at getwiq.ai/try. Do one a week, and by the end of the quarter your agents have something to read. Or book a demo at getwiq.ai/demo and we will show you what a full record of your operations looks like.