Your competitors are using the same models you are. Probably the same tools too. If you are waiting for a better AI to give you an edge, you are waiting for something that will arrive for everyone simultaneously.
The organisations that are pulling ahead with AI are not doing it because they chose a different model or hired more engineers. They are doing it because they gave their AI something better to work with. That something has a name.
It is codification: one of the least glamorous practices in technology leadership.
What the level playing field actually means
When AI capabilities were scarce, the advantage went to whoever got access first. That window has largely closed. GPT-5, Claude, Gemini, GLM: these tools are available to your team and to every competitor in your space. HBR wrote earlier this year: “Context is demonstrated execution: the accumulated body of decisions, standards, and institutional knowledge that shapes how an organisation actually works.” When every company can use the same AI models, that organisational context becomes the differentiator.
Most organisations have almost none of this in usable form.
Their best judgment lives in the heads of their most experienced people. Their decision logic is handled informally: “ask Sarah, she’ll know.” Their standards exist, but only as shared intuition among the people who have been around long enough to absorb them. This works well enough when humans are doing the work, because humans are good at filling gaps. When AI enters the picture, those gaps become problems. The AI has nothing to check itself against. It produces output that looks right but misses the edge cases your best people would have caught. You end up reviewing everything carefully, which means the AI is saving you much less time than you hoped.
The fix is not a better AI. It is codification.
What codification is, and what it is not
Codification is not writing a wiki page that nobody reads, or a documentation project that sits in Notion and goes stale. Those things describe what a process is. Codification makes the decision logic explicit: what triggers each step, what the rules are at every fork, what good output looks like, and what you would override and why.
The distinction matters because AI doesn’t need description: it needs criteria. “Write a customer email in our brand voice” produces mediocre results. “Write a customer email in our brand voice: direct, not corporate; one idea per sentence; never apologise for the product; escalate if the customer mentions a refund” produces something you can actually use.
That second version is codification. Someone sat down and made the tacit judgment explicit. That is hard work. It requires time with the people who actually do the job, patience with ambiguity, and the discipline to keep asking “what does that mean in practice?” until you have something specific enough to act on. Most teams skip it because it is slow and because the pressure to show AI results fast runs in the opposite direction.
I argued earlier this year that you need to codify before you automate: rushing to deploy AI onto undocumented processes is why many AI projects hit a wall. That argument is about sequencing. This one is about something different: what happens after you have deployed, and why the organisations that keep codifying are building an advantage that compounds.
It compounds
When you codify judgment once, you get a better AI output today. When you keep codifying, adding the edge cases your first pass missed, updating the standards as your understanding improves, layering in the exceptions that turned into incidents, something different starts to happen.
New team members get up to speed faster. Your AI outputs get more trustworthy because there is something concrete to check them against. And the institutional knowledge that used to walk out the door when your best people left? It stays.
Barry O’Reilly argues that the scarcest resource in the AI era is not information but codified judgment: visible, specific, and up to date. Information is everywhere. Judgment encoded in a form that can actually be acted on is rare, and it holds its value.
This is what Sarah Guo calls the untrainable: the assets that AI cannot absorb because they are specific to how your organisation actually works, the decisions you have made, the standards you hold. These things are competitive advantages precisely because they are hard to copy. A competitor can licence the same AI. They cannot licence the body of codified judgment you have built over time.
The uncomfortable part
None of this is fast.
Codifying judgment well requires working sessions with the people who carry it, the discipline to keep writing things down when delivery pressure is pushing you to move on, and the patience to update things when they change. When the expectation is AI productivity gains in weeks, investing in foundational work is a hard thing to justify in a meeting.
That is precisely the opportunity.
Most organisations will not do this. They will adopt the tools, measure token spend and PR velocity, and wonder why their AI outputs are not as good as they expected. The ones that invest in codification will be operating at a different level within six to twelve months, not because they had a better AI, but because their AI has better material to work with.
The sequencing problem in AI adoption is real, but the sequencing does not end at launch. The organisations winning with AI are not the ones who got there first. They are the ones who kept doing the unglamorous work after everyone else stopped.
The next post in this series is about the specific challenge of getting judgment, not just process, into codified form: Judgment at Scale Is a Codification Problem (2 July 2026).
And the third covers what a living standard looks like in practice: how you maintain codification as the organisation changes, rather than letting it go stale: The Living Standard (4 July 2026).
Leave a Reply