A senior CAIO at a global financial services firm told me recently that she had finally landed her operating model. After eighteen months of internal politics she had moved her organisation from a centralised AI CoE, through a messy federated phase, into embedded squads with a thin central platform team. The board approved the structure. The org chart was elegant. The senior engineers had stopped quitting.
Two weeks later her CFO asked the question every CAIO knows is coming. "Show me the unit economics on these agents." She could not. Eighteen months of operating model work had not produced the one thing that mattered in that meeting: a way to see what the agents cost against what they were worth.
That gap has a name. Call it the operating tax. It is the cost of running agents whose economics nobody can see, and it compounds quietly, because invisible cost does not announce itself. It waits for someone to ask for the number.
This issue is about why the operating tax is the dominant gap in the 2026 AI leadership conversation, and what to do about it before your own CFO asks.
· · ·
The operating model conversation has matured. We now have real frameworks for where AI capability should sit inside an enterprise: centralised CoE, federated, embedded squads, evolving across four stages from centralised by necessity to institutional. Five years ago leaders argued about whether to have a CoE at all. Today they argue about how to evolve out of one. That is genuine progress, and I do not want to wave it away.
But every one of these frameworks answers a single question. Where does AI work live? None of them answers the question the CFO actually asks. Is that work profitable?
The phrase that matters here is cost-per-outcome. I do not mean what an agent costs to run. I mean the all-in cost of one useful unit of work: the tokens, the compute, the human review, the rework, set against the value that unit produces. The cost of a result worth having, not the cost of the engine that produced it.
Here is the idea I want you to leave with, because it reframes the whole conversation. Your operating model does not decide whether you have a unit economics problem. It decides whether you can see it. The org chart is a visibility instrument before it is anything else. Get it wrong and the problem still exists. You just meet it later, in a board meeting, with no data in your hand.
· · ·
Watch what happens to visibility as an organisation matures through the four stages, and the pattern is hard to unsee.
When everything is centralised, the CAIO can see every agent, because every cost line and every value claim runs through one team. The constraint here is throughput, not sight. You know what your agents cost. You simply cannot build them fast enough.
Then the pressure arrives. Business units start buying their own tools, hiring their own engineers, shipping agents the central team never hears about. This federated-by-pressure stage is where the operating tax is born. Shadow AI spreads, the footprint outgrows anyone's ability to inventory it, and cost-per-outcome goes dark at the exact moment spend starts to climb. The board hears that AI capability is scaling fast. What nobody has the data to say is that AI cost is scaling faster, and the value side of the ledger has quietly stopped being measured.
The fix leaders reach for next is to distribute delivery while centralising the platform. Embedded squads, central rails. It is the right instinct. But the rails that get built first are almost always availability and security. Cost attribution and economic telemetry get filed under Phase 2. And here is what I keep seeing: by the time Phase 2 arrives, the damage is already done, because the agents that shipped in Phase 1 have been running blind for a year.
Which is why the timing matters so much. Every enterprise agent has an Inversion Point: the month its cost-per-outcome climbs past its value-per-outcome. After that month the agent still works. It still produces volume. It just costs more than the result is worth, and every additional outcome it generates makes you marginally worse off than if you had switched it off. Copilot-style agents that assist a human invert late, sometimes never. Workflow agents acting inside fixed boundaries tend to invert between approximately month twelve and month eighteen. Autonomous agents that chain reasoning across steps invert earliest, often inside the first year.
The Inversion Point exists in every operating model. It does not care about your org chart. The only thing your org chart changes is whether you watch it coming or get told about it afterwards. A centralised team sees it because cost and value sit in the same place. A federated organisation sees almost none of them, because the two halves of the equation sit in different business units. An embedded model sees them only if someone insisted that cost-per-outcome telemetry was a non-negotiable rail and not a Phase 2 nicety.
· · ·
You do not need a maturity assessment to find out where you stand. You need one question.
Given your operating model as it exists today, can you produce, inside 48 hours, a list of every AI agent in production with its current cost-per-outcome, its projected Inversion Point, and the threshold at which you would switch it off?
If the answer is no, your operating model is producing invisibility, whatever stage it has reached. A polished embedded model with no economic telemetry is not safer than a federated mess. Both deliver the same surprise to the same finance committee about eighteen months in. The elegance of the structure is not the protection. The visibility is.
So the sequence is not operating model first and economics later. It is both, in the same breath. Decide where AI work should live, yes. But decide at the same time that no agent ships without three numbers attached: what one outcome costs, when it is projected to invert, and what would make you retire it. Then build the capability to actually retire it, because a kill criterion that lives in a slide deck and not in the runtime is a wish, not a control. And put the whole picture on a dashboard the board sees quarterly and you see weekly.
The operating model is the container. The unit economics discipline is what tells you whether anything worth keeping is inside it. Build one without the other and you arrive at a beautifully evolved org chart full of agents nobody can defend.
The leaders who run both will define this field in 2027. The ones who run only the org chart will be back in front of the board, explaining the operating tax, wishing they had built the visibility a year earlier.
· · ·
One thing before you close this issue.
Run the question on your own program, not a hypothetical one. Could you produce that 48-hour list right now? Every agent in production, its cost-per-outcome, its Inversion Point, the threshold to switch it off. Not eventually. This week.
Reply (vijayan@aideliverydiscipline) with one word. Yes or no.
I read every reply myself. I am building a picture of how many enterprise AI programs can actually answer the question their CFO is about to ask, and when it is complete, everyone who replied gets the aggregate.
· · ·
The operating model is where AI work lives. The unit economics discipline is whether that work pays. Both decisions compound.
Vijayan Seenisamy is the creator of the AI Role Operating Framework (AI ROF™), the author of The AI Delivery Manager Blueprint and The Pilot Trap, and the publisher of The AI Delivery Discipline newsletter. He currently leads group transformation at a top-30 ASX-listed company.
New issues arrive by email first. Subscribe here.
