Notes on deploying.
Or: on what change management knows about the AI race.
The capital is moving. The conversation is loud. For three years the question that absorbed the field was which model is best. It is starting to be which deployment is real, and against which existing systems. Companies that call themselves AI-native are not exempt. Companies that don't are even less so. The next two years will involve rethinking work that has not been touched in a long time.
That part of the work has a name. Practitioners have been refining it for twenty years.
The race moved
The model layer has not finished moving, but it is no longer the constraint that decides outcomes. Capability is, increasingly, a given. What an organization will trust and reuse is the scarce thing.
On a November evening in 2022 I was sitting next to a team that had spent months inside the pre-chat versions of these models. I have written elsewhere on this site about what the following week did to our estimate of the ceiling. What matters here is where it left the constraint: not capability, but access.
The constraint has moved again. It is now neither capability nor access. It is what happens once both are in place.
The discipline has a name
The book Rewired, by Rodney Zemmel and his co-authors at McKinsey, names the building blocks of digital and AI transformation in a way that is useful to anyone who has worked inside one. There are six: strategy and roadmap; talent; operating model; technology; data; adoption and change management. None of them are new. None of them are AI problems.
The reason Zemmel and his co-authors take the time to name them is that practitioners keep skipping the ones they find hardest. Anyone who has done a platform replacement in financial services has lived through some version of each. The terminology is McKinsey's. The work is older.
The two everyone names
Of the six, talent and technology are the easiest to put into a slide deck, a board update, a CEO town hall. A certification cohort is countable. Enterprise model access can be committed to in writing. A center of excellence is photogenic.
These are real, and useful. They are also the two of the six that the conversation has been built to carry. The market right now names them well. It mostly stops there.
The four it doesn't
The other four are strategy and roadmap, operating model, data, and adoption and change management. They photograph badly, and they are where the outcome gets decided.
There was a State Street accounting platform replacement I worked on at a global asset manager, twelve business groups deep, whose report inventory I have written about elsewhere on this site. The piece of work that decided whether the replacement was honest was a piece nobody had asked for: a capabilities assessment that took the prework question seriously. It found two things the plan had not accounted for. The back office was already struggling with the data and reports it had. And the platform we were buying sat at the leading edge of what the vendor had built rather than safely behind it, which meant we would be living on that edge with them.
That went into the contract. The price stayed roughly where it was; what came with the price did not. The vendor committed dedicated support through the migration and access to products on their roadmap as they shipped, on the understanding that this was a partnership in something still being built rather than the purchase of something finished.
Years earlier, a Cambrian capabilities assessment for an insurance asset management business surfaced the same lesson. The assessment phase is the underpaid phase. It is also the phase that decides the rest.
The conversation right now describes deployment. It mostly does not describe the discipline that converts capital and talent into outcomes. No deal or vendor contract does that conversion. People do it, on the inside, over a longer horizon than a news cycle.
Where the stakes are
I was covering emerging-markets debt during the 1998 Russian financial crisis, the one that took the ruble down by more than two-thirds in three weeks and a sovereign into default. For two or three weeks of it I was up at four, reading the overnight news or calling the junior traders in London to find out how much had moved while I slept. Most of those mornings I went in. The note had to be on the desk by six, because clients read it on the commute, and the internal meeting at seven-thirty was hours after they needed it. You could do the analysis, set the strategy, build the book, and watch all of it get rewritten overnight.
I think about that period whenever someone hands me a finished AI governance plan or a locked rollout sequence. The plan is worth making. It is also more provisional than the deck admits: change the conditions enough and you are back at step one, holding a roadmap for a world that no longer exists. The discipline is being ready to write it again.
The new tools don't change the disciplines. They raise the stakes on them.
None of this surfaces in a market briefing or a quarterly letter. It decides whether the headline numbers ever happen.
On purpose
The capital is arriving and the capability is here. The discipline that decides outcomes is old, known to anyone who has lived through enough transformations to know what gets skipped and what that costs. The firms that do that work on purpose this time will not have to discover it in the post-mortem.
// DG