Author
- TL;DR
- How the Gateway Log Writes the Pricing Model Menu
- Three Measurements a Pricing Decision Needs
- The Wrong Subject: Limits per User, Prices per Customer
- When Your Own Agent and the Customer’s Agent Share One Usage Counter
- Reading Usage Telemetry for What It Does Say
- Decide the Value Metric, Then Build the Counter
- Making the Pricing Model Call
- FAQs
TL;DR
The pricing models a team considers are written by what its systems already log. Usage telemetry built to keep a service running becomes the option menu, and the value metric that would fit the buyer best is usually the one nobody can count yet. Decide the unit first. Then build the counter to it.
A team opens the question of how to price a new capability, often an agent-facing API or an MCP server, and the first move is to pull the logs. The analysis is sound. It is also the wrong starting point.
How the Gateway Log Writes the Pricing Model Menu
The sequence is the same regardless of product. Engineering exports what the gateway records: calls by tool, by user, by authentication method, by day. Someone builds a table, and every candidate on the pricing list is arithmetic on it. A price per call, a package of calls per month per user, a flat fee with a call ceiling above it.
Then a product leader proposes the unit that tracks what the customer receives: rows returned, records enriched, cases the agent closed. The gateway records none of those, so the proposal is marked “not measurable” and drops off the list. Nobody evaluated it against the buyer. It was eliminated on a logging question.
Instrumentation chosen to keep a service up becomes the pricing option menu by default. The team believes it chose a model. It chose among the models its logs permitted.
Three Measurements a Pricing Decision Needs
A pricing decision needs three kinds of measurement, and the gateway log holds them unevenly.
Consumption
How much of the product each customer draws, over what period, at what rhythm. Logs hold this completely; they were built for it.
Cost to serve
What each unit of consumption costs the vendor to deliver. A gateway sees the request and the response time. It usually does not see payload size, the downstream inference cost one call triggers, or the storage a result occupies afterward. One call that fans out into several model invocations and one that returns a cached answer count identically.
Value
What the customer received, in the customer’s own terms. Nothing a gateway writes says whether the record it returned was the one the customer needed, or whether the case the agent closed stayed closed. Value is never in a log.
A decision made from one of three inputs takes the form of that input. Work from the logs alone and you will produce a consumption decision whatever model you name, because consumption is the only quantity in front of you.
The Wrong Subject: Limits per User, Prices per Customer
Rate limits exist to protect the service. They are enforced per user, or per API key, over a short window: a burst ceiling against a runaway loop, a daily ceiling against a misconfigured job. The platform team set them long before anyone asked a pricing question.
When pricing arrives, those limits read as a ready-made structure.
A commercial allowance is a different object: a volume per customer over a billing period, with a price attached to the next unit above it. The subject differs, a user versus a customer account. The dimension differs, a rate versus a volume. And the running total per account that would enforce the commercial version is usually unbuilt.
A fair usage policy is the commercial cousin of the safety limit: boundedness of the grant, written at the policy layer because the licensing model declined to write it. Keyed to the user, an allowance also lets a customer buy headroom by adding logins, and the vendor ends up pricing the customer’s org chart.
Ask of any limit the platform team hands you: who is the subject, and over what period is it counted? If the answer is a user and a short window, you are holding a safety control. The commercial allowance has not been designed yet.
Naming the unit the business should count is the judgment call; the counter is often the first deliverable a pricing engagement produces. If your team is debating whether an allowance is “measurable,” talk to a pricing expert before the logging question closes the decision.
Talk to an Expert About the Pricing Problem in Front of You
Describe what you’re facing and a pricing expert will reply with a concrete read on your licensing, packaging, and pricing architecture.
When Your Own Agent and the Customer’s Agent Share One Usage Counter
Vendors now ship an agent inside the product. It reaches the same tools through the same gateway as the agent a customer connects from outside, and one per-user budget serves both.
The vendor’s own agent consumes the budget, and the customer’s integration calls are refused on a limit the customer never saw. And a meter on calls counts the same work twice, once for the in-product agent and once for the customer’s agent finishing the same job.
Neither stream can be priced until they are separable. That is instrumentation, and it comes first. Put the question to engineering: can the gateway tell which calls originated inside our product and which arrived from the customer’s agent, per account, per billing period? A no means the pricing conversation is early. MCP server pricing is the worked case.
Reading Usage Telemetry for What It Does Say
The telemetry answers two questions well, and both belong to the grant: what the unit entitles the buyer to do.
Who the user is
When an agent acts for a person at a desk, its volume follows the working day and drops at weekends; a scheduled integration runs flat. The log can tell you whether the user is a named person, a process, or an agent acting for either, and that decides what the grant covers.
How the use is distributed
In the patterns our library holds, a few integrations often account for most of the volume on an agent-facing path, and the long tail of human-driven use barely registers. That distribution says which accounts the grant has to bound and which it does not.
Decide the Value Metric, Then Build the Counter
The order that works: the metric’s owner names the unit, engineering builds the counter to it, and the price attaches to a quantity someone chose on purpose. Who should decide your value metric covers the ownership question. When nobody owns the value metric, the gateway chooses it by default.
The reverse order, counter first and unit second, is how a company ends up metering its own plumbing. Who owns your meter now follows what happens once that plumbing has an owner. The cheapest quantity to count is the one already counted, so the unit becomes whatever the gateway stores, and the gateway stores calls.
The counter built to a chosen unit does a second job. It lets the company re-read the unit when a person at a desk becomes a scheduled process. A unit chosen without a counter cannot be re-read, and the pattern we see most is the company learning at renewal, from the customer, that the unit stopped tracking value.
A software licensing model is two decisions: the value metric names the unit, and the grant defines what the unit entitles the buyer to do. Telemetry informs the grant today. It cannot choose the metric.
Making the Pricing Model Call
After the log analysis, the team holds the consumption half: how much, by whom, on what rhythm, with what concentration. It does not hold the value half: which unit the buyer recognizes as what they paid for, and what that unit should cost.
That half is the engagement. In Pricing Ground Truth, SPP’s entry engagement, a pricing expert reads your pricing architecture from the deals you have already closed, against the patterns in our library. LevelSetter, SPP’s pricing decision platform, holds the deal record and the analysis. The judgment about which unit the counter should be built to is made by people who have made it before.
If your pricing model was chosen from a list the gateway wrote, bring the log analysis to a pricing expert. It is the first half of the answer, and the second half decides whether the meter you build counts value or counts plumbing.
