2 min read

The AI Agent You Can't Leave

The AI Agent You Can't Leave

 

Hundreds of thousands of dollars. For one AI agent.

That's the quote the IT lead at a resources company described to me last week. A single AI "co-worker", with up to "two or 300" sub-agents underneath it.

But the price was only half the story. The second offer was the part that mattered: the vendor would build the agent, then run it for them, every month, as a service.

He turned it down, and his reason was blunt: "they tied into you for life."

He's right. I should know. My industry wrote that playbook.

 

This Time, the Lock-In Holds Your Process

Managed service providers have always known that a contract which is hard to leave is a good contract. From our side of the table, anyway. Agent-as-a-service is that model rebuilt, with one difference that matters. The old lock-in held your infrastructure. This one holds your process.

When a vendor builds and runs an agent for you, the valuable part isn't the model. It's the rules: how your purchase orders are checked, which exceptions go to a person, what "approved" means in your business. That's how your organisation actually runs, written down. If those rules live in the vendor's configuration screens, you haven't bought an agent. You've rented your own process back.

The industry's language makes this easy to miss. Microsoft's Frontier Firm glossary defines digital labour as "AI or agents that can be purchased on demand". Buying capacity on demand is fine. Renting the logic of how your business works is a different decision, and almost nobody is treating it as one.

 

An Agent You Can't Move Is a Bill You Can't Control

The same IT lead had already done the maths that worries him. Suppose his business spends $1 million a year on AI tokens. If prices rise sharply, he said, that could become "six or $7 million", and by then "all my business is now built around those coworkers or agents."

That's not a fantasy. BCG reports that average token costs for the top US language models rose from below $5 per million at the start of 2025 to $20 per million in 2026. The cheap end of the market is still getting cheaper. The frontier end is not. If you can't move your agent to a different model, you can't follow the price down.

Nor is this only a mid-market problem. Oliver Wyman surveyed large enterprises in Europe and North America this year and found only 10% hold a portability statement for their agents signed by the executive sponsor. Nine in ten don't.

 

A Tenfold Price Gap Is Telling You Something

Then look at what buyers are being asked to pay today. In the same fortnight, prospects told us about quotes for a single agent that ranged from around $17,000 to "hundreds of thousands". A more-than-tenfold spread for broadly similar work is not a market. It's a sign that nobody has a reference price, and that part of what buyers are paying for is the dependency, not the build.

 

Three Questions Before You Sign

  1. If we end this contract tomorrow, what do we keep? The prompts, the business rules, the test cases and the logs should be yours, in your tenant, in a format someone else can read.
  2. Can it run on a different model, and who decides when it does?
  3. What does it cost at twice the volume and twice the token price?

Then borrow one more question from the IT lead. Every buyer should be asking it: when the contract ends, "do they continue to use our data?" Get the answer in writing.

A Managed Intelligence Provider should be judged on how easily you could leave it.

That includes us.

We'd rather earn the renewal than engineer it.

Rent the compute. Own the process