Who Audits the Meter on Our AI Agents, and What Audit Rights Do We Need Before We Sign?

Nobody audits it unless you write the right in. On a consumption AI contract the vendor defines the billable unit, detects it with its own system, and holds the only full-fidelity record of what your agents did, so the audit clause and your own parallel count both have to exist before signature, not at the first true-up.

That is a different ask than a spend cap. A cap tells you when to panic. An audit right tells you whether the number you are being billed against is the number that happened.

Every term in the meter is a vendor-set lever

Look at what the buyer does not control in the Salesforce construct. A Flex Credit pack is priced at $500 per 100,000 credits, an agent action is decomposed as 20 credits, and an action is bounded at 10,000 tokens, which works out to $0.10 per action at list, per Software Pricing Partners' dated timeline (July 19, 2026, last verified Sep 03, 2026): softwarepricing.com/blog/agentforce-pricing.

Read that chain again. The vendor sets the pack rate, the decomposition, and the token boundary that decides when one action becomes two. None of those move the headline rate when they change.

The outcome SKU is the sharper version. Agentforce Help Agent bills $2 per autonomous resolution, announced June 25, 2026, and a resolution there is defined as the absence of human escalation or negative feedback, detected by the vendor's own system (same SPP source above). A buyer who quit halfway through can still register as resolved. SPP puts the buyer's real question in one line: at every true-up, who audits the meter.

There is already a market answer to that. Intercom prices Fin for Sales at $10 per qualified lead and lets each customer define what qualified means, which trades vendor control of the definition for a count both sides can check (SPP, same URL). That is the shape you are negotiating toward, whoever the vendor is.

The flat fee does not remove the meter, it defers it

If your answer to consumption risk is a flat-fee enterprise agreement, know what you bought. Salesforce introduced the Agentic Enterprise License Agreement on October 14, 2025 as a multi-year flat fee covering Agentforce, Data 360, MuleSoft, and Slack (SPP timeline, URL above). The meter keeps recording underneath it for the whole term, and everything it records becomes the baseline evidence for the next negotiation.

So the flat fee buys predictability for the term and hands the vendor a re-anchored baseline for the renewal. If your agents scale to real volume during those two or three years, you will sit down at that renewal against a consumption record only the vendor holds in full. That is the moment your own count either exists or does not.

What the advisory side is telling buyers to get in writing

UpperEdge's Adam Mansfield, writing September 1, 2026, tells Agentforce buyers to define what an action is (or what an outcome is for Help Agent) with examples before signing, to demand price certainty including removal of any conditions tied to maintaining volume or spend at renewal, and to push for the ability to exchange unused Flex Credits for other Salesforce products as well as rollover: upperedge.com.

He also names the Claudeforce version of the problem. Customers acquiring the Salesforce product in Claude get Headless Consumption Pricing on the Salesforce side and still have to negotiate the Anthropic deal separately (same URL). Two meters, two counters, two sets of definitions, and the reconciliation is yours.

On the legal side, Mayer Brown's February 2026 note on contracting for agentic AI describes the shift from a SaaS-style agreement to a service-style one, with a defined delegation of authority, real-time audit logs, and mandatory escalation triggers for human review: mayerbrown.com.

Those two documents cover the commercial and legal halves. Neither one tells a revenue operator what to count on their own side, which is the part that has to be running before the term starts.

The clause set

Four things belong in the order form, and none of them are exotic.

1. A definition of the billable unit that cannot change without your written consent, with worked examples of one unit, two units, and a case the vendor considers a single unit that you would have called two.
2. A right to the underlying event log behind each billed unit, in export form, at a stated cadence, not a dashboard summary. You are asking for the rows, not the chart.
3. A reconciliation and dispute window after each invoice, with a stated remedy when your count and their count diverge past a set threshold, plus the right to carry an unresolved dispute into the renewal rather than settling it under signature pressure.
4. Renewal terms that price from a jointly readable record. Strike anything that conditions your pricing on maintaining volume or spend, and get true-down rights, not only true-up.

Your own count

The security architecture actually helps here, if you use it commercially. Salesforce's Claudeforce page states that eligible customers will be able to keep activity data in cloud infrastructure they control, under their own encryption keys, access policies, and audit logging: salesforce.com/claudeforce.

That is your raw material. If the activity data sits under your logging, then a RevOps owner can maintain a parallel count of agent actions, escalations, and completed tasks from your own side of the boundary, and reconcile it monthly rather than annually. The vendor's number stops being the only number in the room.

Name the owner before the pilot starts. If nobody owns the parallel count, it will not exist at renewal, and the renewal is the only conversation where it matters.

Related: Claudeforce two contracts · Agentforce Flex Credits forecast · AI pricing seats to outcomes · Digital Wallet kill switch