Some of the most critical evidence about how an organisation really operates is not missing from the record because nobody thought to capture it. It is missing because it lives in conversations, relationships, and informal agreements that happen outside of any system and were never meant to be logged.
I worked with a team that had a formal review process. It was documented, logical, and well-designed on paper. An item was submitted via a template, entered into the system, scheduled for the appropriate governance discussion, and a decision was made to move forward, pause, or cancel. It was clean, auditable, and traceable.
Thanks for reading The Operating Model Dispatch! Subscribe for free to receive new posts and support my work.
Except that was not how it worked in practice.
In practice, teams were having a preliminary conversation before anything went near the system. They talked through what to put in the template. They agreed on the governance timing informally. Only then did they submit. By the time the governance meeting happened, the meaningful conversation had already occurred. The formal session was abbreviated, because everything substantive had been resolved upstream. Teams had learned that going through the official process without that preliminary alignment usually resulted in resubmission and a second pass through governance. The workaround was more efficient than the designed process. So the workaround became the real process.
Now here is the question I want you to consider: what would an AI audit of that process have found?
High template completion rates. Strong governance adherence. Decisions being made. All the right signals. It would have concluded the process was working.
It would have been wrong about almost everything that mattered.
The data problem is not a solvable data problem
There is a version of the argument that says AI cannot yet diagnose operating models because it does not have access to the right data, and that this will be resolved as organisations get better at capturing organisational behaviour. I want to challenge that framing, because I think it misunderstands the nature of what is missing.
The preliminary conversation in the example above was not missing from the system because nobody had thought to record it. It was absent because it was the kind of thing that happens between trusted colleagues who have worked out, through experience, how to get things done. The agreement on governance timing was not a process step that needed to be logged. It was a relationship artefact, built over time, that made the formal process functional. The reason the governance meeting was abbreviated was not documented anywhere, because the people in the room did not think of it as a deviation. They thought of it as how work actually gets done.
This is the category of information that sits outside every system an AI can access. Not because of a data capture gap. Because it is, by nature, informal, relational, and contextual.
Let me be specific about what falls into this category.
Informal networks. Every organisation has a formal structure and an informal one. The informal one is where a significant proportion of decisions, influence, and problem-solving actually live. Who gets called before a meeting to shape the agenda. Whose support is needed before something gets tabled. Which relationships carry weight that is not reflected in any reporting line. An AI can analyse org charts, communication metadata, and collaboration patterns. It cannot tell you that the person three levels down in the hierarchy is the one whose read of a situation determines whether an initiative moves or stalls.
Decision rationale. Systems capture decisions. They do not capture why those decisions were made, what was discussed and discarded, what the informal trade-offs were, or what was left unsaid. The documented rationale and the actual rationale are frequently different things, not because people are dishonest, but because the real reasoning includes context, history, and relationship dynamics that have no natural home in a system. An AI reading the decision log is reading the official version of events.
Employee sentiment as it operates in practice. Surveys and engagement data give you an aggregated signal with a lag. They do not tell you what the team that is critical to your AI deployment actually thinks about it, whether the person managing that team is managing it in a way that creates or suppresses honest feedback, or whether the stated support for a transformation is genuine or performative. The gap between declared sentiment and operational reality is one of the most common sources of AI programme failure I am seeing right now.
Judgement about what the signals mean. Even with good data, the interpretation of operating model evidence requires contextual judgement that AI does not yet possess. The same pattern, in two different organisations, can mean completely different things depending on history, culture, leadership style, and what has been tried before. The governance meeting being abbreviated might be a sign of mature, efficient process. It might be a sign of a decision-making culture that avoids scrutiny. Knowing which requires knowing the organisation, not just the data.
Real-life application of any model. Operating model design is not a pattern-matching exercise. The right model for a specific organisation depends on things that are specific to that organisation: what its leadership team is actually capable of executing, where the informal power sits, what has failed before and why, which structural tensions are productive and which are destructive. These are not generalisable. They are the product of diagnosis. Proper, human-led, contextually informed diagnosis.
Why this matters for AI deployment
I want to be clear about what I am not arguing. I am not arguing that AI has no role in operating model work. It has a significant role, and I am actively working on where and how it applies. I am not arguing that AI tools are overhyped in general. Some are; most are not.
What I am arguing is that organisations that deploy AI onto an operating model that has not been properly diagnosed are building on foundations that AI itself could not have assessed. The informal networks that will determine whether adoption happens are invisible to it. The decision rationale that explains why previous initiatives failed is not in any system. The employee sentiment that will shape whether the programme lands is not captured in any data source the AI can read. The workarounds that reveal how work actually gets done look, in the data, like process compliance.
This is not a reason to delay AI investment. It is a reason to do the diagnostic work first.
The operating model problems that cause AI to fail are exactly the category of problem that AI, properly deployed, is well-suited to help solve. But only once the groundwork has been laid. You cannot automate your way to clarity about a system you have not yet understood.
The organisations that will get the most from AI are the ones that know what they are deploying it into.