𝐉𝐚𝐧𝐞 𝐒𝐭𝐫𝐞𝐞𝐭 𝐟𝐞𝐝 𝐢𝐭𝐬 𝐀𝐈 𝐦𝐢𝐥𝐥𝐢𝐨𝐧𝐬 𝐨𝐟 𝐥𝐢𝐧𝐞𝐬 𝐨𝐟 𝐢𝐭𝐬 𝐨𝐰𝐧 𝐜𝐨𝐝𝐞. 𝐈𝐭 𝐠𝐨𝐭 𝐰𝐨𝐫𝐬𝐞.

That is not a failure story. It may actually be the most useful thing I have heard about AI adoption all year.

Jane Street is a trading firm built on unusual software and an unusual programming language, so nothing commercially available understood their environment.

They built their own assistant, and the obvious first move was to train it on everything they had ever written. Millions of lines. The crown jewels.

The results were disappointing.

What worked was smaller and stranger. A few thousand 𝐩𝐚𝐢𝐫𝐞𝐝 𝐞𝐱𝐚𝐦𝐩𝐥𝐞𝐬: here is what broke, here is what somebody did about it, here is the thing working again.

The machine learned far more effectively from problems joined to their solutions than from information delivered in bulk.

Your organization generates that material every single day.

A customer complaint got escalated last month, and somebody resolved it. Both halves were written down, in two different systems, by two different people, and nobody has ever placed them side by side.

The link between the problem and the fix is the only valuable part, and it was never recorded anywhere.

𝐄𝐧𝐨𝐫𝐦𝐨𝐮𝐬 𝐟𝐥𝐨𝐰. 𝐍𝐨 𝐚𝐜𝐜𝐮𝐦𝐮𝐥𝐚𝐭𝐢𝐨𝐧.

One detail in the process is worth pointing at. Jane Street never asked a single developer to document a mistake. Instead, they directly captured what was being edited, what temporarily broke, and how it got repaired. Automatic, ambient, and blameless.

That design decision is not a footnote, it is the entire mechanism. Ask people to log their own errors and you will receive a beautifully curated file of things that went well.

They also trained an AI to review code. One day it replied: "I'll do it tomorrow."
It had learned that from the human reviewers.

What you capture teaches your habits alongside your expertise. Uncomfortable, for sure, yet also the most honest mirror your organization will ever get.

This works wherever the work already leaves a trail: complaints, corrective action reports, contract redlines, returns, purchase order corrections. Most of that paperwork exists for compliance reasons, so nobody has to volunteer anything.

Where the correction happens in someone's head, in a room, in real time, and never touches a system, ambient capture fails. That knowledge is real and this method does not reach it.

Recognize which is which.

The good news is that you are not missing an asset. 𝐘𝐨𝐮 𝐚𝐫𝐞 𝐟𝐚𝐢𝐥𝐢𝐧𝐠 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐨𝐧𝐞.

Here is a test you can run this month. Pick one recurring task and try to assemble fifty real problem-to-solution pairs from the last four weeks, without asking anybody to write anything new. If you cannot, the finding is not that the pairs are missing. It is that your work leaves no trace.

No model you rent will fix that.

#AIStrategy #KnowledgeManagement #OrganizationalLearning #AIAdoption #Leadership

Contact

bruno.gentil@sherpaconsultingasia.com

© 2026. All rights reserved.