When Your AI Vendor 10x’s the Price, This Is What Saves You

The repricing scenario is not paranoia; it is the observable history of enterprise software. Usage-based products reprice, acquirers harvest, favorable early terms expire, and the negotiating leverage in year three is set entirely by what it would cost the customer to leave. So the right way to hear the 10x question is as an architecture question wearing a procurement costume: on the day the price moves, what does this firm actually have to move? Inventory the stack honestly and every AI deployment decomposes into two layers. The model layer, the intelligence being rented, is a commodity market with several capable providers whose prices trend down under competition. The workflow layer is everything the firm taught the system: the prompts and skills encoding how the firm underwrites, the standards files, the normalized datasets, the templates, the review gates, the file structure the workflows read.

All the accumulated value sits in the second layer, and the entire lock-in question reduces to one fact about it: where does that layer live, and in what form? If the firm’s underwriting logic exists as portable artifacts, plain files the firm holds, in formats any capable model can execute, then the model underneath is a swappable part, and a repricing triggers a bake-off, not a crisis. If the same logic was built inside a vendor’s proprietary workflow builder, its agents configured in someone else’s interface, its data normalized into someone else’s schema, then the logic is a hostage, and the price will eventually be set accordingly. The vendor does not even need malice; the incentive gradient does the work.

The reason this discipline got dramatically cheaper is that the connective layer standardized. The Model Context Protocol, the open standard Anthropic introduced in November 2024 and published as an open specification, defines how AI applications connect to external tools and data sources, and its adoption is why Yardi, covered earlier in this series, could launch its connector on Claude first while stating that additional model support would follow: an MCP connector is built against the protocol, not against any one vendor’s product. That sentence is the strategic point. When the integration standard is open, the integrations the firm invests in stop being vendor commitments: the same connector serves whichever model sits behind it. The skills pattern points the same direction, workflow instructions expressed as plain, portable files rather than configurations trapped in a proprietary builder.

PSV’s own build standard follows from this, and it is the honest disclosure behind every recommendation in this series: workflows as plain instruction files the client holds, data in the client’s own systems in standard formats, integrations through open protocols where they exist, and the model reached through an interface a competitor could satisfy. Anthropic’s models are PSV’s default because they are currently excellent, and the architecture is designed so that sentence can change without the client rebuilding anything. That is what own your workflow means concretely. It has nothing to do with hosting your own model and everything to do with holding your own logic.

Treat portability like backup integrity: a thing verified, not assumed. Once a year, run the drill. Export every workflow artifact and confirm the export is complete and human-readable. Take one production workflow, the weekly variance memo is the usual candidate, and execute it against a different model provider in a sandbox, measuring what breaks and how long the re-point takes. Pull the data-exit answer from each vendor contract and check it against reality on a sample. Write the one-page result: what moved cleanly, what snagged, what the true switching cost is in days and dollars. The reviewer is whoever owns the AI budget. The gate: any new vendor whose honest answer to the export question is no, or whose exit clause is silence, prices its lock-in into the negotiation or does not enter the stack.

The drill’s value compounds in the negotiation nobody schedules. A firm that can truthfully say our workflows ran on a competitor’s model last quarter in testing negotiates renewals from a different chair than a firm that merely suspects it could leave. And the drill catches drift: the convenient feature adopted mid-year that quietly wrote firm logic into a vendor-only format, the dataset that migrated into a proprietary schema because the import was one click. Drift toward lock-in is the natural direction of every integrated product, not a scandal; the drill is simply the countervailing force, an afternoon a year spent keeping the exit real.