Research March 2025 10 min GDA Research

01 Seat-based pricing prices access to a tool. It cannot price a system that removes the need for the seat.

02 Agent-as-a-Service prices completed work — per task, per outcome, or per unit of throughput.

03 Outcome pricing requires reliability the buyer can underwrite, which is an orchestration problem, not a model problem.

04 Orchestration and agent topology are where reliability is produced, and therefore where margin concentrates.

05 Model providers are becoming suppliers. The routing layer is becoming the business.

Per-seat pricing embeds an assumption: that value scales with the number of people using the tool. It held for two decades because software made individuals more productive and organisations bought one licence per individual.

An agent that completes a workflow end to end inverts this. Its value scales with volume of work, and its effect on headcount is neutral at best. A vendor charging per seat is charging on the metric its own product is designed to reduce, and buyers notice.

The transition is already visible in how enterprises procure. Requests are increasingly framed as volumes of work rather than numbers of users, and that framing is incompatible with the licence model on the other side of the table.

Agent-as-a-Service prices the work: per task completed, per outcome achieved, or per unit of throughput, frequently with a floor and a service standard attached.

This is closer to how enterprises buy business process outsourcing than how they buy software, and the comparison is instructive. Outsourcing contracts specify a standard, a volume, and a remedy for failure. Buyers understand the shape.

It also transfers risk. Under a licence, the buyer absorbs the cost of the tool not working. Under outcome pricing, the vendor does. That is a substantially better deal for the buyer, which is why it will win, and a substantially harder business to run, which is why most vendors are resisting it.

A vendor can only price outcomes if it can predict them. A single model called in a loop cannot deliver predictable outcomes on non-trivial work, because failure modes are correlated, silent, and difficult to bound.

This is where orchestration stops being infrastructure and becomes the product. Routing a task across specialised models, verifying intermediate output, escalating on low confidence, and holding an audit trail is what converts probabilistic components into a system a buyer can contract for.

The distinction matters commercially. A vendor selling model access competes on price against every other reseller of the same weights. A vendor selling underwritten completed work competes on reliability, which is defensible.

Early agent systems were chains: a fixed sequence of steps. Chains are simple to build and fail badly, because an error at any step propagates unchecked to the end.

Graph topologies — where tasks branch, run in parallel, verify each other, and converge — are harder to construct and materially more robust. Verification can be a node. Escalation to a human can be a node. Confidence thresholds can gate an edge.

The practical consequence is that reliability becomes a design property rather than a hope. That is the precondition for underwriting outcomes, and therefore the precondition for the pricing model.

Foundation models are converging in capability and falling in price. That is favourable for anyone building on them and difficult for anyone selling them.

The layer that decides which model handles which task, verifies the result, and stands behind it commercially is not converging, because it encodes accumulated operational knowledge about a specific domain's failure modes.

Investors should expect the orchestration and delivery layer to sustain margin longer than the model layer, and should be sceptical of any AI business whose only defensibility is access to weights it does not own.

Seat licence Prices access · vendor bears no delivery risk

Consumption Prices compute · buyer bears reliability risk

Agent-as-a-Service Prices completed work · vendor bears delivery risk