AI has an unusual habit. It exposes reality.
Point it at a process, and the gaps appear quickly. The documented procedure says one thing, experienced people do another, and the system records something else. A hand-off that looked simple turns out to depend on a spreadsheet, an inbox and the memory of one person. A performance measure exists, but nobody can explain whether it represents a good customer outcome. A process has many participants, yet no obvious owner.
None of these problems were created by AI. They were already there. AI simply raises the cost of pretending they are not.
I have become increasingly interested in whether an organisation can tell itself the truth. Can people describe how work actually moves, rather than how it was designed to move? Can leaders acknowledge what they do not know? Can teams explain where work stops, where decisions wait, and where people compensate for systems that do not fit the job? Can we discuss weak performance without immediately turning the discussion into blame?
This matters because agentic AI needs more than a set of instructions. It needs intent, context, boundaries and a way to recognise what good looks like. If those things are unclear to the humans responsible for the work, they will not become clearer simply because we place a model in the middle.
The temptation is to treat documentation as the answer. I do not think it is. A polished process map can still describe a fiction. The deeper requirement is a culture in which people are willing to expose variation, exceptions, workarounds and uncertainty. Documentation becomes valuable when it captures operational truth, not when it provides comfort that a process has been mapped.
This is where psychological safety and operating model design meet. Candour reveals the real work. Business architecture helps us connect it. Ownership gives someone responsibility for improving it. Measures give us evidence that the change was worthwhile.
The organisations that safely scale AI may not be those with the cleanest starting point. They may be those most willing to see their starting point clearly.
I began this journey asking how we could make AI safe for the organisation. Increasingly, I am asking whether the organisation is safe enough for people to reveal what AI needs to know.