Will this streaming business survive AI? The diligence question nobody asks

The question that quiets the boardroom

Ask a streaming CEO how AI helps and you will hear about cost. Faster localisation, cheaper marketing assets, leaner post-production, a smaller line in the content budget. The answer is fluent because it has been rehearsed. Then ask a second question: is the business more or less defensible after AI? The room goes quiet.

That silence is the real story.

It matters because the money is already moving. McKinsey estimates that AI could influence roughly 20% of original-content spend within five years and, after mass adoption, redistribute up to $60 billion of annual industry revenue. Redistribute is the operative word. Some of that value stays with incumbents. Some of it lands with smaller studios, creators, and platforms that barely featured in the last diligence cycle.

Distribution moved first. S&P Global Market Intelligence forecast that CTV, social, and online video ad spend would together reach $72.4 billion in 2025, up 13.5%, surpassing linear for a second year. The audience, and the data trail it leaves, now sits in places the old bundle never measured.

So the question for anyone buying, holding, or running a streaming business is not whether AI saves money. It will. The question is what remains that a competitor cannot copy once everyone has the same tools.

Diligence still reads the rear-view mirror

Commercial diligence on media assets has a settled method. You rebuild the revenue history, cut the cohorts, test churn and ARPU, and check that the growth story in the management deck survives contact with the data room. It is forensic work, and it answers one question well: did this business earn what it says it earned?

AI changes the question. A revenue line can be entirely real and still be substitutable. If a rival, a platform, or the customer itself can reproduce a motion with AI at a fraction of the cost, historical accuracy tells you little about forward value. The forensic answer is right. It is also beside the point.

The shift is from revenue forensics to substitution testing. Instead of asking whether a motion produced revenue, you ask what happens to that revenue when the work behind it becomes cheap for everyone.

Two lenses make the test usable, and the rest of this report relies on both.

The first separates the production engine from the commercial engine. The production engine makes the content: development, production, post, and localisation. AI hits it loudly, mostly as cost. The commercial engine wins and keeps the money: packaging, pricing, distribution deals, ad sales, partner management, and retention. AI hits it quietly, and this is where defensibility is decided.

The second separates the operator from the investor. The operator has to run the transition without breaking the engine that funds it. The investor has to price whether the transition is possible at all, inside a hold period, with this team. Same facts, different risk. A business can look cheap precisely because nobody has asked the operator the hard question yet.

Four numbers that shift the burden of proof

Content supply is about to get cheaper and larger

McKinsey’s scenarios include more personalised content, a larger role for user-generated content and smaller studios, and a potential increase in total content supply. Its 20% estimate for AI-influenced original-content spend is a five-year view, not a distant one. When supply expands and production cost falls, a catalogue’s scarcity value falls with it. A library is still an asset. It is no longer a moat on its own.

Distribution has already changed hands

The S&P Global figure describes advertisers following audiences out of the bundle. Growth of 13.5% on a $72.4 billion base rewards whoever owns the screen, the login, and the behavioural signal behind them. A content business that rents its distribution also rents its data.

Most AI projects still do not pay

Bain’s 2025 survey of investors representing $3.2 trillion of assets under management found that nearly 20% of portfolio companies had operationalised generative AI use cases with concrete results. PitchBook reports that adoption remains uneven and estimates only about 20 to 25% of AI projects produce meaningful returns. Read together, the numbers say the same thing. Owning AI is common. Converting it into results is rare, and that conversion capability is exactly what diligence should test.

The bottleneck is plumbing, not models

Deloitte’s 2026 outlook finds that many media companies still lack a unified view of audience behaviour across services and platforms. Pilots commonly stall on fragmented metadata, inconsistent rights records, and legacy contracts, not on model quality. Gartner makes the structural point: AI services built on common models and public datasets risk commoditisation, and durable advantage requires proprietary operational data the provider owns and can reuse. The model is a commodity. The governed data underneath it is not.

Put the four together and the burden of proof moves. Supply is rising, distribution has migrated, most AI spend has yet to pay, and the real constraint sits in data that few media companies have organised. A seller who claims AI makes the business stronger now has to show it, motion by motion. A buyer who accepts that claim on the strength of a pilot list is underwriting hope. Toggle the chart below to see both sides of the ledger.

Where the moat actually sits, layer by layer

Four tests turn the evidence into a diligence method. Each one is cheap to run. Most deal teams skip all four.

Four public bets on where value will settle

The past year produced a natural experiment. Four large moves gave four different answers to the layer question, and each one leaves a substitution test unanswered.

Notice the pattern. Each move buys, separates, or combines layers. None of them, on its own, delivers a governed process running across those layers. That process is the part no purchase price includes, and it is the part the 40% shock tests hardest.

Read the table as an investor and you see four valuation theses. Read it as an operator and you see four integration programmes, each with a rights database, an audience view, and a commercial team that must now work as one system.

Defensibility is engineered in the hold, not staged for the exit

For operators

Start where Deloitte says pilots stall: metadata, rights records, and legacy contracts. The NBCUniversal separation shows why. When a group splits, audience data, rights, and shared services must be divided and rebuilt. Every split is a chance to build the unified audience view on purpose, or to inherit two fragmented ones by accident. Pick the two or three commercial motions that carry the business. Run the 40% shock on each. Fund the process changes that protect the survivors, and let the rest go. Bain’s advice to tie AI to a small set of priorities is the discipline here.

For investors

The Fox and Roku deal shows the shape of both the opportunity and the risk. A $22 billion price can buy a platform and its data. It cannot buy the integration that turns platform data into better ad pricing and retention inside the acquirer. That work happens after close, across the hold period, or it does not happen at all.

Timing is the trap. Defensibility cannot be retrofitted in the six months before a sale, because the next buyer will run the substitution test you should have run at entry. Build the value creation plan around the moats that survive the shock. Measure progress on those moats, not on the number of AI pilots launched. Pilots are activity. Moats are evidence.

Six signals that will separate builders from buyers

• One rights record at Paramount SkydanceWatch whether the Warner Bros Discovery integration produces a single governed rights and audience view, or two libraries under one roof.
• Roku signals inside Fox ad salesThe real test of the $22 billion deal is whether platform data shows up in Fox’s pricing and packaging.
• Peacock as a standalone engineOnce NBCUniversal separates, its disclosures will show whether streaming economics hold without the broadband parent.
• Versant’s use of cashVSNT is valued on linear cash flow. Watch whether that cash funds a transition or simply runs off.
• Creator and small-studio shareMcKinsey’s scenarios point to more supply from smaller producers. A rising share signals faster commoditisation of mid-tier content.
• AI return rates in portfoliosIf PitchBook’s 20 to 25% figure moves, so should the discount you apply to the AI plan in the management deck.

None of these signals is about model capability. Every one is about whether a company can turn assets it already owns into a process a competitor cannot copy.

The question to take back to your own engine

A moat that management cannot draw on a whiteboard is a moat no buyer will pay for.

Across commercial diligence for private equity and partnership deals closed in more than 70 countries, the pattern we see is consistent. The businesses that hold value through a technology shift are rarely the ones with the best tools. They are the ones that turn strategy into governed execution: clear ownership of each commercial motion, clean data and rights beneath it, and a team that can explain plainly where AI helps and where it threatens. Tools are bought in a quarter. That operating system takes a hold period to build. That gap, strategy to governed execution, is the work we spend our time on at nGülam.

So run the test on yourself. Given a marker and 20 minutes, could you draw your commercial engine and mark exactly where AI makes it stronger, and where it makes it optional?