Agentforce and the Per Conversation Question

How does Agentforce per conversation pricing work?

Agentforce is priced on a consumption meter measured per conversation, so the cost rises with how often the agent engages rather than with a headcount of seats. This is part of the broader meter shift, where Salesforce and other vendors move pricing away from seats toward usage, agent, and outcome models. The appeal to the vendor is obvious: a seat is capped at one human, but conversations have no ceiling, so revenue grows with adoption automatically. The risk to the buyer is the mirror image: a meter that scales with usage can scale faster than the value it delivers, especially when conversation volume is driven by customer behaviour you do not fully control. Understanding that the meter, not the seat, is the cost driver is the starting point for negotiating it.

This sits in the AI pricing defense cluster. For the agent layer mechanics more broadly, read Agentforce, negotiating the agent layer, and for the principle that governs any usage based premium, see the AI premium, paying for features you do not use.

Why does the definition of a conversation matter so much?

The definition of a conversation matters because it is the unit you are billed on, and an unclear unit is an open ended bill. If a single customer interaction that spans several messages, a follow up the next day, and an escalation to a human all count as separate conversations, your volume and your cost multiply against the same underlying work. This is the same problem that outcome pricing created when vendors billed per resolved case without first agreeing what resolved means. The counter is to write the definition into the contract before signing: specify what starts a conversation, what ends it, how follow ups and re engagements are treated, and whether a handoff to a human still counts. A precise definition is worth more than a headline rate, because the rate only has meaning once the unit is fixed.

How do you forecast and cap the meter?

Forecast and cap the meter by replacing open ended consumption with a realistic band and a hard ceiling. Start from real interaction data: your current contact volumes, the share the agent will plausibly handle, and the seasonal peaks that will spike it. Use that to commit to a band you will actually reach, because an inflated commitment chosen to win a lower unit rate simply converts to a different overpayment. Then cap two things in the contract: the unit rate, so a mid term repricing cannot raise it, and the total spend, so a surge in conversations cannot produce an uncontrolled invoice. Negotiate what happens above the band, whether overage bills at the same rate or a punitive one, and secure rollover so unused committed volume is not lost at the period end. The aim is a meter whose worst case is known before you sign.

How do you test whether Agentforce is worth the premium?

Test whether Agentforce is worth the premium by demanding ROI evidence before you accept it, not after. An AI agent that genuinely deflects volume and resolves cases has a measurable value, and the vendor should be able to show it against your own data, not a generic benchmark. In the current market, AI driven asks run 20 to 37 percent against a historical 3 to 9 percent annual uplift, by published market estimates, so the premium is large enough to demand proof. Run a scoped trial that measures deflection, resolution quality, and the human time actually saved, and price the agent against that result. Where the value is unproven or the volume is too uncertain to forecast, keep the option to run the process without the agent, so you are not committing to a meter before you know what it returns.

A worked example of capping the meter

Consider an indicative example. A retail business is offered Agentforce to handle customer service, with per conversation pricing and an opening commitment well above its current volumes. Rather than accept the headline rate, the buyer settles the definition first, grouping all messages in a session and excluding human handoffs, which alone cuts the projected count materially. It then forecasts from twelve months of contact data, commits to a band just above its realistic volume rather than the vendor's suggested figure, caps the unit rate for the term, and sets a total spend ceiling with overage at the committed rate. Finally it runs a scoped trial to confirm deflection before expanding. These figures are indicative, but the structure is what matters: the buyer converts an open ended meter into a bounded, forecastable cost, and lands the deal inside the 10 to 30 percent savings disciplined negotiation typically produces, by published market estimates.

What to do next

Before signing any Agentforce agreement, settle the definition of a conversation, forecast from real data, and cap the rate and the total. The full defense method for AI priced products lives in the AI Pricing Defense Guide, the Salesforce specific mechanics are covered on the Salesforce negotiation service, and the complete clause set is collected in the AI Price Defense Kit.

Last reviewed April 2026