Customer data is becoming the AI battleground

Every AI vendor is now fighting over one thing, and they are not going to tell you which thing.

It is your customer data.

Two events last week made this obvious. On July 6, TechCrunch reported that Google had quietly expanded what it retains from your uploaded media (Lens images, Search Live voice recordings, Translate audio, all of it) and is now training on that pile by default. Not opt-in. Opt-out. A few days earlier, Amazon Web Services announced it was closing Mechanical Turk to new customers on July 30, the crowdsourced task marketplace it launched in 2005 under the winking name "Artificial Artificial Intelligence." And in the background of both stories is SAP's April 2026 API policy, which prohibits third-party AI agents from touching customer data through documented APIs unless SAP has personally blessed the integration.

If you read those as three unrelated tech news items, you miss the actual story.

The AI industry is done pretending the model is the product. The moat that matters now is who gets to hold, gate, and retain the customer's context. And every vendor with a system of record, the databases that hold the authoritative version of your customers, your invoices, your CRM records, your inventory, has spent the last twelve months figuring out how to close ranks.

I know the objection forming in your head. "Sure, but this always feels alarmist and it always resolves back to open."

Let me steal George Fraser's argument, because he actually has receipts.

Fraser is the CEO of Fivetran, the company that has spent thirteen years running a pipe from every SaaS tool a company uses (Salesforce, NetSuite, whatever your finance team lives in) into a data warehouse the company controls. In a June a16z podcast with Martin Casado, Fraser called SAP's move "stupid" and predicted it would collapse the same way every previous open-API panic collapsed. He is probably right about the eventual outcome. Open APIs are a durable equilibrium. Closing them means declaring war on your own customers, and that war is unwinnable if you sell business software.

But eventually is doing a lot of work in that sentence.

Here is the specific SAP language, updated on April 27: no API use "for interaction or integration with (semi-) autonomous or generative AI systems that plan, select, or execute sequences of API calls" outside SAP-endorsed architectures. On the same investor call, SAP CEO Christian Klein said his company reserves the right to throttle "millions of calls coming towards an API" for stability reasons. Translation: even if the written policy softens, they now have a stability pretext to rate-limit any third-party AI you plug in.

Which means the moat is not the policy. The moat is the twelve to twenty-four months of ambiguity before it resolves back to open, during which you cannot cleanly move your operational data out and cannot cleanly plug your preferred agent in.

That is where the leverage is bleeding out.

Google is playing a different version of the same game. The TechCrunch piece points at a June email to customers announcing two new settings: Search Services History and Personalized Recommendations, both on by default. Google's own help documentation confirms the point in plain language: your saved history is used "such as training generative AI models" and reviewed with "the help of human reviewers." Every voice search, every Lens photo, every Translate practice session. All of it eligible for training and human review unless the user goes and unchecks a box most of them will never see.

You can call that quiet. I would call it strategic. Google is not banning anyone from anything. It is widening the funnel of proprietary training data faster than any regulator will move, and the regulatory clock Google actually cares about starts ticking in 2027.

Meanwhile, at the bottom of the stack, the source of the human judgment that used to fuel these models is drying up. AWS is closing Mechanical Turk to new customers because the AI labs have moved on to specialized annotation vendors like Scale AI and Surge AI, which employ vetted domain experts instead of anyone with a browser. A 2023 study found that Mechanical Turk workers were themselves quietly using LLMs to answer the tasks, which is the funniest kind of failure. The lesson the labs took from that is not "crowd data is bad." It is "generic human data is fungible; expert human data is a moat."

Set those three stories next to each other and the arc is clean. At the top of the stack, the platforms are widening their intake of your users' data. In the middle, the systems of record are choking off third-party access to your operational data. At the bottom, the labs are consolidating the human expertise that makes any of this work into a shrinking set of vendors. Every layer is being enclosed. Not by conspiracy. By identical incentives arriving at identical conclusions.

The founder's window on this is narrower than most people think. And this is where I want you to feel the sharp edge, because "we'll think about data governance at renewal" is going to age extremely badly.

Right now you can still choose. You can still pick a data platform, replicate a copy of your systems of record into it, and treat that copy as the canonical version an agent talks to. You can still evaluate two or three AI vendors against the same context and switch without pain. You can still pull your inputs out of a specialty vendor's control plane, because the market is fragmented enough that they have to compete on portability.

In eighteen months you probably cannot. Once your CRM's built-in AI is doing the work its API used to enable, once your users' voice data is baked into a foundation model, once your annotation vendor has calcified into an infrastructure dependency, the price of moving is not a migration project. It is a rebuild.

If you are building a company right now and you are not thinking about where your customer context lives, who gets to read it, who gets to train on it, and how you would move it out inside a quarter, you are about to find out.

Fraser's prediction is probably right on a five-year timeline. The vendors will discover they cannot supply every AI tool their customers want, they will fold, and the APIs will reopen. Great. But the companies that get through the intermediate period with their leverage intact are the ones building their own data foundation right now, running their own connectors, keeping a portable, complete copy of the customer context they depend on, and evaluating vendors on how easily they can be replaced instead of how deeply they can be integrated.

Photoroom just published the readable version of this playbook. On the same day the Google news broke, they released the fourth part of a series describing the actual data pipeline behind their open 7B image model on Hugging Face. Not the model announcement. The pipeline. Captioning strategy. Data formats. What they store, how they store it, why they store it that way. That is what a company looks like when it decides the moat is under its own roof instead of somewhere in a vendor's private roadmap.

The rest of us have about a year to make the same call.

Please tell me why I am wrong.