Portability is the concession

Context is the best lock-in software has ever had, because the customer builds it themselves for free. Two incumbents just published a standard for handing it over.

An open standard for moving users and their accumulated context between AI products, with Harvey and Thomson Reuters behind it. Read what is actually being given up.

The best lock-in ever invented

Context is switching cost the customer builds themselves. Every conversation, every correction, every document the system has read — none of it cost the vendor anything and all of it makes leaving worse. It compounds daily and it never appears on a balance sheet.

Incumbents do not publish standards for making that portable. That is the story.

So why do it

Because the alternative is worse for them. In professional services no firm runs one vendor for everything. The vendor that refuses to interoperate does not become indispensable — it becomes the one that gets ripped out wholesale at renewal, because it is the only one that made itself a single point of decision.

Handing over context is a bet that being one of four tools forever beats being the only tool briefly. For legal technology that is probably the right read.

The clause to read

A protocol that transfers user context across a vendor boundary is also a protocol for moving privileged material across a vendor boundary. Legal work is exactly the domain where that gets examined. Consent and scope handling is where this specification will be judged, not throughput.

The pattern underneath

The same month, the integration protocol crossed a billion SDK downloads and enforcement moved in front of the model rather than behind it.

Three announcements, one direction: the boundaries between AI products are being standardised by the people who would historically have defended them. Standards emerge where differentiation is impossible — not where cooperation is virtuous.

What this asks of you

If you are buying, ask every vendor what their export path is and make them answer in the contract. There is now a standard to point at, which means "we don't support that" is a choice rather than a limitation.

If you are building, assume your context moat evaporates. What is left is whether the product is better, which was always the uncomfortable version of the question.

SourceTrail — AI Agents in 2026: Protocol Updates, Enterprise Tools, and Autonomous Payments → · InfoQ — MCP Goes Stateless, and Developers Ask Whether That Just Makes it an API Again →