You Cannot Cancel the Stack: The License Stays, the Seats Go
Agents do not collapse the go-to-market stack. Every tool is two products, a data layer and a workflow layer above it, and agents dissolve only the second…
Every go-to-market tool sells two products under one contract. There is a data layer: the records, the history, the entitlements, the snapshot of what an account looked like on a given date. And there is a workflow layer built on top of it so that people can operate the data: the screens, the queues, the assignment rules. The workflow layer is what got demoed and what the purchase decision was made on. Agents dissolve it. The layer underneath does not get lighter. It gets heavier, because something now queries it constantly and at machine cadence. The consolidation pitch says the stack collapses and the software budget falls with it. What agents remove is one narrow class of tool. A tool whose only function was routing a request to the right person and telling the next person it arrived has nothing left once an agent does the routing. Systems of record stay, because the record is the thing being read. Systems of delivery stay for a harder reason. The thing that sends the message or moves the money holds an accumulated position with a counterparty: sender reputation with mail providers, a processor relationship underwritten by years of chargeback history. An agent needs an execution endpoint and cannot manufacture standing at one. That position transfers to nobody. Take a support desk. An agent takes the queue, the triage, the first response, and the routing of what it cannot resolve. Every ticket ever filed stays where it was. The invoice would arrive smaller for a system the business now depends on more heavily. No vendor gets cancelled. Spend per vendor falls while vendor count barely moves. Cancellations concentrate in that narrow class, which is small enough that the list shortens at the margin while the aggregate holds. The claim is falsifiable, so name what would break it. Twelve months after agents take over a set of workflows, pull the accounts payable export. If the vendor list is materially shorter and the software line has fallen by more than the removed seats explain, the consolidation case was right and this note was wrong. Finance already produces that file. A vendor watching its seat count empty does not sit still while the invoice shrinks. It reprices onto the unit that did not disappear, the volume moving through the data layer: resolutions, records, API calls, enriched rows. Where that repricing has already landed, spend holds or rises while seats collapse, and the change is visible only in the seat line. A vendor on a consumption denominator has already priced the agent in. Count seats before dollars. The consolidation case gets approved by a finance function shown a software number. If the prediction holds, that number sits where it was twelve months later and the program reads as underdelivering. It did not underdeliver. It moved a cost into a budget the case never touched, and the owner of that budget was not in the approval meeting. A program judged against the wrong line gets defunded in its second year while it is working. Name the line the saving will land on. Name it in the approval. A renewal needs a number, and there is a published one. A mid-market company's first production AI system runs $150K to $400K to build and $3K to $8K a month to operate, on a scope of one to three workflows. The operating half is $36K to $96K a year. Amortize the build across three years and the run rate for owning a workflow outright lands near $86K to $229K a year. That arithmetic is illustrative, not a quote, and it is not a cancellation device. It is a price on the alternative. A tool whose license sits under that number while its data layer is still doing work has been re-justified on arithmetic. A stack review run purely for cost is dangerous. The workflow layer is cheap to rebuild now: conditional logic, state, and a set of screens over a documented API, which is the exact class of work that agentic development tools compress. The data layer was never a cost problem. It is a time problem. History accrues at one day per day, and no budget accelerates it. Five years of resolved tickets, or the state of an account on the date a model was fit to it, is not recoverable by writing better code next quarter. Half the switching cost just got cheap. The other half did not. The failure mode does not show up at cancellation. It shows up at the next model refresh. A cost-driven consolidation cancels the enrichment vendor and keeps the contract data, on the reasoning that contracts are the asset. Then the territory model stops reconciling. The historical firmographic snapshots it depended on, the employee counts and industry codes as they stood when the model was fit, were the vendor's data under a standard enrichment license. They left when the contract did. Current data can be re-bought tomorrow. The version of the past the model was calibrated against cannot. A line item was saved. A model was lost. There is still a reason to run the review, and it is not cost. Every tool in the stack is a data contract somebody else wrote. What gets retained. In what form. Who owns the derived fields the vendor computed on your behalf. Those terms were agreed at signature by somebody optimizing for a discount rather than for an exit, and they have gone unread since. Ask, of every vendor on the list, what leaves the building with them. We run this review at the start of a Strategy engagement, before any workflow is scoped for an agent, because the map that says which workflows to automate is the same map that says which tools are still earning their license. The review is a reading exercise on contracts and a subtraction exercise on seats. The outcome it should produce is fewer seats and the same vendors.