Every few years a new operating model arrives with the confidence of a discovery. This year it is the forward-deployed engineer — put the builders next to the customer, collapse the distance between the problem and the person who can fix it. The instinct it triggers in most leadership teams is a reorganisation.
Before you sign it off, one inconvenient question: where is the evidence that it works better?
The reorg instinct and the missing test
I went looking for the head-to-head comparison — embedded or forward-deployed teams versus a centralised centre of excellence, measured on delivery outcomes. In the reachable software-delivery literature, that test does not appear to have been run. There is plenty of advocacy. There is very little measurement. (In healthcare, where more head-to-head outcome studies exist, they tend to favour centralisation — which if anything cuts against the fashionable direction, not for it.)
What the strongest general research does say is more interesting than either camp's position. MIT CISR's late-2025 study found that innovation was more effective in enterprises with strong IT leadership, modularity and reuse — and that this held regardless of the environment's clockspeed, whether the firm competed in a low- or high-innovation-velocity market. Note what it does not say: it explicitly allows that leadership can sit at the enterprise or business-unit level, or both. The winning variable is the quality of platform and standards leadership, not its location on the org chart.
That is an awkward result for the reorg. It says the axis everyone is arguing about is close to the wrong axis.
What the economics actually says: complements
The framing of "structure versus skills" turns out to be a false choice, and this is the best-established finding in the area.
The organisational-economics literature on why IT investment pays off for some firms and not others is consistent across decades: the returns are largest when technology, organisational redesign and skill investment move together. They are complements — firms that buy the technology without the organisational change, or restructure without building capability, tend to underperform the ones that do both. (That's "work best together," not "neither works alone": the same literature does find independent positive effects for skills or for technology on their own — the point is that the combination compounds.) The productivity-paradox work makes the same case at economy scale: the missing ingredient is accumulated intangible capital, of which both process and skills are components.
Two useful qualifications sit underneath. Training programmes evaluated rigorously tend to show near-zero effects in the short run, with modest gains appearing two to three years out — so training is real but slow, and shouldn't be sold as this year's fix. And different technologies push in different directions: research finds information technologies and communication technologies have opposite effects on how far decision authority gets pushed down. "Decentralise because technology" is not a general law.
The counter-case for the thing you were about to dismantle
Centres of excellence have become the strawman of AI operating-model discussions, so state their position precisely. A scoping review of the centre-of-excellence model catalogues a set of proposed value criteria — concentrated expertise, standards-setting, leadership, structural coherence — but its central, honest finding is that these criteria have not been validated against outcomes: no studies were found correlating them with results.
Read that carefully, because it doesn't rescue the CoE so much as confirm the theme of this piece: on the specific question of which operating model delivers, the measurement mostly isn't there — for embedded models and for centralised ones alike. What we have is theory. And the relevant theory — the economics of knowledge hierarchies — predicts a hybrid rather than a winner: push routine problems down to where they arise, and concentrate the rare, hard ones where scarce expertise sits. Any real organisation needs both, and the live argument is about the boundary, not the victor.
What does predict outcomes
If not the org chart, then what?
Management-practice quality — measured across thousands of firms — independently predicts productivity. That is a claim about how well you run the operating model, not which one you picked. It's worth noting the same research finds that adopting modern management practices often increases decentralisation, so this isn't a pure "structure doesn't matter" argument — it's that execution quality carries weight independent of the model you choose. Alongside it sits the CISR finding: platform and standards leadership, with genuine reuse.
Both are things you can improve without reorganising around this year's fashion.
What to do instead
Stop treating the operating model as the decision. The defensible moves are narrower and duller than a reorg, and better supported:
Invest in structure and skills together, because the returns compound when they move as a pair. Put real leadership behind platforms, standards and reuse, because that is what actually correlated with outcomes. Keep a concentrated capability for the hard, rare problems, and push the routine ones to where they occur. And if you do restructure, be honest that you're acting on judgement — because on the specific claim that forward-deployed beats centralised, the evidence in this domain simply isn't there yet.
The reorg is the most visible thing a leadership team can do about AI delivery. It is not, on the available record, the most effective.