This deep dive expands on the Loop Asia conversation with Aki Ranin.

Aki Ranin raised a question in the episode that deserves more than the minute it got: "What if you stopped working with this company? Who keeps the agents?" It's aimed at managed agent platforms — the pitch where a vendor builds, hosts, and runs your agents for you. Good pitch. Also, Aki argues, a decision most companies are making without treating it as one.

The pitch and the trap are the same feature

The appeal is real: "This looks really attractive — you're going to build my agents for me, you're going to run them for me, so that's great." Aki isn't against managed services generally — the issue with agents is structural. "As we imagine a Venn diagram, there's a huge overlap between what we're talking about in terms of sovereignty and agentic AI, because agents are effectively how we're going to be running our businesses." An agent is an accumulation of decisions: how your processes work, your edge cases, your exceptions, what customers actually ask for. Build that inside a vendor's platform, and the knowledge lives there too.

Jon pushed on how hard that actually is to walk back: "Pulling the plug and switching to another provider is so much harder than even changing from one mobile phone company to another." Aki's answer was blunter than the question: "It could be impossible, either for technical or legal reasons."

Aki's 40-person-startup story illustrates the milder version of that risk. Claude API access was cut off by accident, and for 24 hours the business didn't function. The larger risk is what happens when the interruption isn't temporary — the relationship ends by choice, dispute, or the vendor's own decisions, and the agents running your operations aren't yours to take with you.

The questions a procurement conversation should actually include

Aki's framing suggests a due-diligence checklist most companies skip:

• Do we own the code, or a contract with somebody? Code you own, you can move; a contract is only as durable as the relationship. "Are you going to own the code, or do you just have a contract with somebody?"
• If we terminate, what do we take with us? Not the marketing answer — the contractual one. Configuration, prompts, fine-tuning data, integration logic: which are exportable, and in what format?
• What's the actual migration cost of moving elsewhere? Never zero — if nobody can estimate it, that's the answer.
• Is this use case sensitive enough that ownership should have blocked the vendor choice? Low-stakes and replaceable can live on a managed platform. Core to the business is different.

Scale changes the stakes

Aki's estimate — at least 100 agentic use cases per company, and "every company will have tons of agents by the end of the year" — turns this from hypothetical into operational risk. One bad decision on one agent is a minor annoyance. Repeated by default across 100 use cases, it compounds into a company that's discovered, as Aki put it, "we're building this outside of our company, and now we're cut out. We can't keep these, we can't move these, we're stuck with something." That discovery lands at the worst moment — mid-negotiation, mid-renewal, mid-dispute — when leverage has already shifted to the vendor.

Governance means putting a gradient behind procurement, not just IT policy

Aki's own practical starting point is an audit, not a policy document: know what's actually deployed today, most of it adopted bottom-up on individual credit cards, before you can set any rule about what happens next. From there, a gradient — free tools, paid subscriptions, private cloud, open models via API, open models on your own hardware — doubles as procurement policy. Where a use case sits on it should set how much ownership risk is acceptable in choosing its vendor. Public-information research can reasonably sit on someone else's managed platform; anything touching customer data or core to the business probably shouldn't, regardless of how good the demo looked.

Governing the tools is the easy part — approve a list, done. Governing the agents is harder, because agents encode process, and that's exactly why this isn't a decision to leave with "the AI guy." In Aki's words: "Don't just push it down the chain and say, 'You're the AI guy, you decide.' AI is so strategic that it sort of belongs maybe even at the board level."