The deck is finished. The number hasn’t moved.

Sit in enough media boardrooms and you start to hear the same meeting twice a year. The Head of Streaming presents the AI roadmap: personalization, generative production, an assistant in search, a pilot in ad ops. The slides are good. The heads nod. Then the CFO asks the only question that matters, which is what any of it did to revenue, and the room goes quiet.

That silence is the signal, and it is not local. McKinsey calls it the gen AI paradox: nearly 8 in 10 companies have deployed generative AI in some form, and roughly the same share report no material impact on earnings. More than 80% see no tangible enterprise-level EBIT effect. At the same time Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on escalating cost, unclear business value and inadequate risk controls.

The instinct in the room is to reach for a better model. That instinct is wrong. The strategy was never the bottleneck. The wiring was.

Two different objects wearing the same word

We use one phrase, “AI,” to describe two things that behave nothing alike. The first is capability: the model, the demo, the roadmap slide. The second is an operating change: a signal that fires inside a live workflow, produces an action a human can approve or reject, and moves a number a controller can see. Most media companies have bought the first and quietly assumed it delivers the second. It does not.

The capability layer is getting cheaper by the month. As Foundation Capital put it this year, “in a world where AI capabilities rapidly commoditize, implementation expertise becomes the lasting differentiator.” When the primitives are near-free and the demos interchangeable, durable advantage shifts to implementation depth: taming messy data, wiring edge-case workflows, integrating the product into the customer’s actual world rather than a slide about it.

Grade AI on how far it has travelled into the revenue system, not on how good the model is. On the left sits ambition, a deck and a pilot. On the right sits an operating model, a governed signal that a board can inspect. The distance between them is where the money lives, and it is almost never a modelling problem. It is an integration problem, and integration has four specific failure points.

Where AI actually dies in the commercial layer

The failures cluster in four places, and none of them is the model. Use the toggle below to see each one from both sides: what the AI deck claims, and what would actually have to be true to change the number.

Interactive view toggle requires JavaScript. The four failure points are: no single revenue owner, no governance or human-approval loop, no baseline to measure lift against, and no wiring into the revenue system.

No owner: a working group is not an owner

Gartner’s Anushree Verma is blunt: “Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.” Hype-driven experiments have committees. Revenue outcomes have owners with a target. When no single P&L leader is accountable for the number the AI is meant to move, it stays an experiment forever.

No governance: the human-approval loop is missing

Gartner also flags “agent-washing,” estimating only around 130 of the thousands of self-described agentic vendors are the real thing. In media, the risk is not just a bad vendor; it is an autonomous action against a brand or an advertiser with no approval gate. Without a governed signal-to-action loop, legal and trust teams veto deployment, and rightly so.

No baseline: “engagement is up” is not a lift

McKinsey finds fewer than 10% of deployed use cases ever make it past the pilot stage. A pilot with no pre-registered baseline cannot prove lift, so it cannot earn budget, so it dies in the pilot graveyard by default.

Not wired in: the tool sits beside the workflow, not inside it

This is the quiet killer. Tubi’s Rabbit AI, a ChatGPT-4 discovery assistant, launched strong in September 2023 and was shut down in 2024 on low adoption. The model was fine. It simply lived beside the viewing decision instead of inside it.

The integration gap, up close

The same three patterns surface whenever you trace a media AI programme from slide to statement. Each is cheap to say and expensive to do.

The self-assessment requires JavaScript. Ask four questions of any AI initiative: does it have a named revenue owner, a human-approval loop, a pre-registered baseline, and is it wired inside a revenue workflow? Any “no” is where it will die.

From AI strategy to AI operating model

Almost every media organisation now sits somewhere on a single line. On the left, an AI strategy: real ambition, real slides, no change to the number. On the right, an AI operating model: a governed signal, a human-approved action, a measured lift that a board can inspect. The gap between the two is not talent or budget. It is integration maturity. Tap any stage to see the operator’s tell.

The spectrum runs left to right: (1) pilot launched, (2) tool purchased but workflows unchanged, (3) AI wired into one revenue workflow, (4) a governed signal to approved action to measured lift loop a board can inspect.

What this means Monday morning

If you operate the business (CEO, CRO, CPO, Head of Streaming)

Stop asking your team which model they are using. Ask which revenue workflow the AI now runs inside, who owns the number, what the baseline was, and who approves the action. If those four answers are not immediate, you have an AI strategy, not an AI operating model, and you are somewhere on the left of the spectrum. The good news is that this is fixable without a single new model. Netflix’s own pattern is instructive: the efficiency wins ship because they are unblocked, while the commercial-layer AI moves slowly precisely because it touches the number. Treat that friction as the work, not the obstacle. Pick one revenue workflow, wire AI inside it, and measure a real lift before you scale to the next.

If you back the business (PE operating partners, value-creation leads, board directors)

In diligence and value creation, “we have an AI strategy” should now read as a neutral fact, not an asset. With Gartner forecasting a 40%+ cancellation wave by 2027, the durable question is integration maturity: can management show you one governed signal-to-action-to-lift loop, with an owner and a baseline? If they can only show pilots and a deck, you are underwriting optionality, not performance. The upside is that integration is a repeatable discipline, so a portfolio company at stage two on the spectrum can reach stage three inside a single planning cycle, which is exactly the kind of value creation that compounds.

Signals over the next 18 months

The next year and a half will sort the operating models from the roadmaps. Six things to track.

The winners don’t have the best AI

The companies winning with AI in media rarely have the best AI. They have the best integration, and that is a discipline, not a download.

The pattern is oddly reassuring once you see it. If the moat were the model, it would already be gone, because the models are commoditising into a utility. The moat is the unglamorous part: naming an owner, building the approval loop, capturing the baseline, and wiring the signal into the workflow that actually bills. That work does not demo well. It compounds instead. Closing that gap, from strategy to governed execution, is the work we spend our time on.

So run the honest test on your own flagship initiative: if your CFO asked today what your AI did to a specific revenue number, would you have an answer, or would the room go quiet?