How does customer zero work for AI adoption?
The technology function deploys AI into its own work first and scales only what survived contact with reality. BCG recommends exactly this: the tech function acts as the company’s first AI adopter, embedding AI in its own delivery before rolling it outward (BCG, 13 October 2026, read 19 October 2026). The output is not a slide of learnings, it is a working playbook: which tasks the tools absorb, what oversight they need, and what they cost, measured before the AI adoption program bets the company on the answers.
Why is customer zero credible in a way pilots are not?
Because the deployer eats the consequences. A pilot can be curated into success; a team that depends on the system cannot hide its failures. You cannot sell a transformation you refused to live through. Microsoft and NetApp both run named customer zero programs, and for a consultancy or vendor the logic doubles as proof: the practice you show a client should be the practice you run.
What does an honest customer zero program require?
Three disciplines. Real workloads, not sandboxes: the AI handles the team’s actual tickets, code and documents, exceptions included. Measured baselines, so before and after are numbers, feeding the scorecard an AI readiness assessment would. And published failures: a program that reports only wins has stopped being customer zero and started being marketing.
What has being our own customer zero taught Soba Labs?
We run the company on the kind of agents we build for clients. Every change an agent makes to this site ships as a pull request that a gate tests, and the writing agent cannot merge its own work, the line we recommend in should employees build their own AI agents. Tools earn a recommendation on our machines first: filtering our coding agents’ command output saved 49.6% of tokens across 1,456 commands, measured in how we halved Claude Code token usage.
The failures arrived on us too. Our first system for capturing lessons from agent sessions used a separate model that forgot past captures; its inbox reached about 172 entries, with one bug recorded five times. A reminder the agent could ignore came next, and a person still had to ask after every task whether everything was captured. Then a blocking check fired on a clock and stopped one 564-turn session about 140 times; now it fires only when a session changed something. Each version failed on our own work before any client saw the pattern.
Where does customer zero fit in the operating model?
As the standing first stage of every AI rollout, not a one-off stunt. New tools, agents and policies enter internally, graduate on evidence, and carry their runbooks into the wider AI operating model and the AI transformation behind it. The sequencing discipline is the whole method: nothing reaches a customer or a company-wide rollout that has not already survived the people who built it.