Salesforce Agentforce at $2 a Conversation: Does the Math Actually Work?

Salesforce runs three Agentforce prices at once, and the one you get quoted decides your bill. At $2 a conversation, 20,000 support conversations a month costs $480,000 a year. Priced through Flex Credits instead, at $0.10 a standard action, the same volume can cost a fifth of that. Agentforce pricing is not a number. It is a choice of meter.

What $2 actually buys on the Agentforce meter

Salesforce publishes three Agentforce rates on one page. The first is $2 per conversation. The second is $500 per 100,000 Flex Credits, which works out at half a cent per credit. The third is an Agentforce user licence at $5 per user per month, and the page notes that it still requires Flex Credits underneath it.

The same page prices the work itself. A standard Agentforce action consumes 20 Flex Credits, and a voice action consumes 30. At half a cent a credit that is $0.10 and $0.15 respectively.

Put those two rates side by side and the conversation price stops looking like a price. $2 buys exactly 20 standard actions. A conversation that fires four actions, which is a normal shape for a knowledge lookup followed by a case update, costs $0.40 on Flex Credits and $2 on the conversation meter. That is a five times difference for identical work.

So the conversation rate is not a unit cost. It is a ceiling. You are paying a premium for the guarantee that one tool-heavy conversation cannot run away from you. That is a real product, and it is priced as though it were the default.

A conversation is a session, not a message

One conversation covers the whole exchange. A 40-message thread and a two-message ping bill the same, which is what makes the metric attractive to a finance team and dangerous to a support team.

Salesforce's pricing page states the rate. It does not state the session boundary. Third-party write-ups describe a session that ends on resolution, on session close, or after a fixed period of inactivity, and none of those is a Salesforce statement. Get the boundary rule in writing, because it silently sets your billable volume.

This matters more than it sounds. If your customers open a fresh session each time rather than continuing an old one, your conversation count is set by customer habit rather than by agent quality. You are then buying a variable you do not control.

Three meters, and the one you will be quoted

Every AI pricing debate collapses into the same question. What event triggers the charge? Salesforce currently answers that three different ways, and the answers are not close to each other.

Flex Credits price the action

Flex Credits meter what the agent does. Each retrieval, each record update, each tool call draws down a balance you have pre-purchased or agreed to pay for as you go. The unit is small, the visibility is good, and the exposure is unbounded on a bad month.

This is the meter a well-instrumented buyer should want. It also demands that somebody watch it, which is the part most organisations skip until the first surprise invoice arrives.

The seat licence prices the person

Service Cloud is still sold by the seat. Enterprise lists at $175 per user per month, Unlimited at $350, and Agentforce 1 Service at $550. A seat price is predictable, budgetable, and completely disconnected from how much work the agent does.

That disconnection is the whole appeal. It is also why seat pricing is under pressure everywhere, a shift covered in the piece on seat compression rewriting SaaS pricing.

Rates are Salesforce list prices read on 20 August 2026, except the pay-per-resolution row, which comes from June 2026 launch coverage. Nobody large pays list.

The three break-even points Salesforce does not print

None of this needs a model. All three crossovers fall out of published list prices and one line of division.

20 actions, 275 conversations, $10 an hour

The first is the one above. $2 divided by $0.10 is 20 standard actions. Below that count, Flex Credits win. Above it, the conversation cap wins.

The second compares the meter to a seat. Agentforce 1 Service is $550 per user per month and a conversation is $2, so 275 conversations per seat per month is the line. If every support seat you would otherwise staff handles fewer than 275 conversations a month, the meter is cheaper. Above that, the seat is.

The third is the comparison a vendor will make for you, and it is the one that closes deals. Assume a tier-1 conversation takes 12 minutes of human handling time. $2 for 12 minutes is a loaded labour rate of $10 an hour. Any support organisation paying more than $10 an hour, fully loaded, is buying that conversation cheaply.

That third number is why consumption pricing sells. It is also why the first two get skipped in the room. The unit price is not the risk. Volume and resolution rate are.

The model, run at three company sizes

Here is the same arithmetic applied to three support organisations. The volumes are mine and they are labelled as such. Everything they are multiplied by is a published rate.

The assumptions, stated

Three sizes: 2,000 agent conversations a month, 20,000, and 200,000. The agent engages every inbound conversation first, so the meter runs on 100% of volume. A typical conversation fires 4 standard actions. The resolution rate is 63%, which is Salesforce's own published internal figure and therefore a friendly assumption rather than a conservative one.

Sample size, window and exclusions, since the standard on this site requires them: this is a constructed model with no sample. It excludes implementation cost, Data 360 consumption, voice premiums at $0.15 an action, and the underlying Service Cloud licences that Agentforce sits on top of. Treat it as a shape, not a quote.

Volumes, action counts and seat productivity are the author's assumptions. All unit rates are Salesforce list prices. The seat row assumes 400 conversations per seat a month, above the 275 crossover, which is why seats undercut the meter at every size here.

Three things fall straight out of that table. The conversation meter is the most expensive option at every size. The gap is not marginal, it is 5 times the action-priced cost. And the seat licence, the model everyone assumes is being disrupted, undercuts the meter across the board.

That last result is uncomfortable and it is the honest reading of two published prices. It also depends entirely on the 400-conversations-per-seat assumption. Drop that to 250 and the seat row becomes the most expensive, which is exactly why the crossover number matters more than the model.

What none of this captures is the cost the vendor carries. Inference is not free, and the margin structure behind a $2 conversation is examined in the piece on inference costs and AI margins. A meter set below cost today gets repriced later.

Deflection is the number that moves the bill

Every scenario above assumed the agent gets a swing at every conversation. On the conversation meter, that means you pay whether it succeeds or not.

Salesforce publishes what its own deployment achieves. In April 2026 it reported that Agentforce had handled 2.6 million customer conversations with a 63% resolution rate and customer satisfaction scores matching human agents. By the Help Agent launch in June, coverage put the help site at 4.3 million inquiries and a 70% resolution rate.

Those are the vendor's own numbers on the vendor's own knowledge base, which is close to a best case. Your knowledge base is worse than Salesforce's knowledge base. Plan accordingly.

Cost per resolved conversation

Divide $2 by the resolution rate and you get what a resolution actually costs you. At 63%, $2 a conversation is $3.17 per resolved conversation. At 70% it is $2.86. At 50%, the number a poorly grounded first deployment can produce, it is $4.00.

That spread is the buyer's risk and nothing else. The vendor's revenue per conversation is $2 in every column.

In June 2026 Salesforce moved the risk back onto itself

On 25 June 2026, Salesforce announced Agentforce Help Agent with a different meter entirely. $2, charged only when the agent resolves an issue autonomously from start to finish.

Constellation Research recorded the terms plainly: no charge when the customer gives negative feedback or asks for a human. CIO reported general availability in July 2026 and a minimum pre-purchase of 1,000 resolution packets, with Agentforce Service EVP Kishnan Chetan describing it as customers paying when their customers' questions are resolved.

What counts as a resolution

Resolutions are questions answered without human escalation. Multiple questions inside a short window count as one. Escalations do not bill, and neither do cases where the customer states they did not get an answer.

Read that against the conversation meter and the concession is obvious. The old model charged for attempts. The new one charges for outcomes, and the difference in your annual bill at a 63% resolution rate is 37%.

My reading is that this is a pricing correction, not a pricing innovation, and it is the right one. Salesforce controls model quality, grounding, and the action library. It is also the party that can measure resolution honestly. Putting the failure rate on the buyer was always a defensible commercial choice and a weak product one.

Constellation's caveat holds anyway. Outcome pricing only works when both sides agree what an outcome is and who validates it. A resolution flag that the vendor sets, on a definition the vendor writes, is not an independently checkable event. That is the governance gap behind Gartner's forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027.

Where this argument is weakest

Three places, and the first is large enough to change the conclusion.

List prices are not the prices anyone pays. Salesforce closed more than 29,000 Agentforce deals in its first fifteen months, and enterprise contracts of that volume are negotiated bundles with credit allowances, ramp periods and floor commitments. Every crossover in this post moves when the rate moves. The method survives the discount. The specific numbers do not.

The seat comparison rests on an unverified inclusion. I have compared a $2 conversation to a $550 Agentforce 1 Service seat as though the seat covers unmetered agent usage. The pricing page separately lists a $5 Agentforce user licence that explicitly requires Flex Credits, which suggests seat pricing and consumption are stacked rather than substituted at some tiers. If they stack, the seat row above is understated and the crossover is wrong in Salesforce's favour.

The volumes are mine. There is no first-party dataset behind the 2,000, 20,000 and 200,000 conversation scenarios, the 4-actions-per-conversation figure, or the 400-conversations-per-seat productivity assumption. They are plausible, they are stated, and they are not evidence.

The case for the conversation meter

The strongest argument against everything above is that variance costs more than money. A conversation cap makes a bad month impossible, and a finance team that can forecast a line item will approve it faster than one that cannot.

There is also a real category of work where the cap wins on price. Complex, multi-system conversations that fire 30 or 40 actions are cheaper at $2 flat, and those are exactly the conversations an agent is worth deploying on. If your agent handles the hard queue rather than the password-reset queue, the conversation meter may be correctly priced for you.

Whether Salesforce holds this position at all is a separate question, argued in the piece on Salesforce as platform or roadkill. The pattern of platforms absorbing point tools is covered in the analysis of point SaaS absorption.

Frequently asked questions

Where to start before your next renewal

Three numbers, and you can pull all three this week.

Count your conversation volume for the last three months, taking the worst month rather than the average, and divide it by the number of tier-1 seats you run. If that figure is above 275, ask why you are being quoted a conversation rate at all. Then count actions per conversation across 50 transcripts. If it is under 20, the Flex meter is cheaper and you should say so in the negotiation.

Last, put one question in writing to your account team. Which events bill, which do not, and where does a session end? A vendor that answers all three precisely is selling you a meter you can manage. A vendor that answers with an ROI multiple is selling you something else, and the test for that is set out in the piece on where measurable AI return has actually shown up.

Related reading