What an Agentforce rollout costs beyond the licence

Salesforce publishes its Agentforce rates, which is more than can be said for a lot of enterprise AI. The published numbers are also the part of your budget least likely to surprise you, so this covers both: what the rate card says, and the four line items that usually dwarf it in year one.

#The three published models

Alongside those, Salesforce publishes add-ons giving unmetered employee usage at US$125 per user per month for Sales, Service and Field Service, and US$150 for Industries. Cloud and Industry user licences are published from US$550 per user per month with the Agentforce add-on included and 2.5 million Flex Credits per org per year.

Consumption is tracked through Digital Wallet, and the buying models are offered as pre-purchase, pre-commit and pay-as-you-go.

#The number that actually decides your model

A standard action consumes 20 Flex Credits. An Agentforce Voice action consumes 30.

At US$500 per 100,000 credits, a standard action is US$0.10 and a Voice action US$0.15. So twenty standard actions cost US$2.00, which is exactly one conversation at the published conversation rate.

That gives you the break-even directly: below roughly twenty actions per conversation, Flex Credits are cheaper. Above it, conversations are. The worked example Salesforce publishes runs a two-action use case at 40 credits, or US$0.20, and twenty requests a day across thirty days reaches 24,000 credits, or US$120 a month.

The practical consequence is that you cannot choose a model from a spreadsheet before you have a pilot. Actions per conversation is an empirical property of your agent and your grounding, and it is the single input the whole comparison turns on.

#The four costs that are not on the rate card

This is the model we use when scoping, and in year one the top row is routinely larger than everything below it.

None of these are Agentforce costs in the sense of being caused by Agentforce. They are pre-existing conditions that an agent project is unusually good at exposing, which is why they land unbudgeted. Where the answers depend on unified customer data rather than documents, that work sits in Data 360 and carries its own schedule.

We set out how to score readiness before any of this becomes a number in what separates an Agentforce rollout that is still running in six months.

#Quality and cost are the same problem

This is the part most cost models miss entirely. Because you pay per action, an agent that retrieves badly costs more and answers worse.

A well-grounded agent that finds the answer in two actions costs a fifth of one that takes ten attempts to assemble the same response from scattered sources. So the grounding work in the table above is not only a quality investment, it directly reduces the consumption line every month afterwards. Teams that treat cost control and answer quality as separate workstreams end up doing both badly.

#Reading the rate card properly

Three details on the published rates change estimates more than people expect.

The action, not the question, is the billable unit. One customer question can consume several actions if the agent retrieves, checks a record, calls out to another system and then composes. Estimating from expected question volume alone understates consumption by whatever your action multiplier turns out to be, and that multiplier is a property of your grounding rather than of your traffic.

Voice costs half again as much per action. At 30 Flex Credits against 20, a voice deployment with the same conversational shape as a chat one is materially more expensive. If both are in scope, model them separately rather than blending.

The user licence still requires credits. The Agentforce User License is published at US$5 per user per month and does not replace consumption; it sits alongside it. Estimates that treat it as an all-you-can-eat seat will be wrong.

#The cost curve over three years

Consumption and project cost move in opposite directions, which is why single-year business cases misrepresent this technology in both directions.

Year one is dominated by remediation: knowledge, permissions, integration, and standing up evaluation. Consumption is small because scope is narrow. A one-year case therefore makes the technology look expensive and the running cost look trivial.

Year two inverts. The remediation is done, scope widens, consumption rises, and the marginal cost of each new intent is low because the platform work is already paid for. This is where the return actually appears, and it appears fastest for organisations whose year-one remediation was genuinely thorough rather than minimal.

Year three is where operating discipline shows. Deployments with an owner, a regression set and a change gate keep their action counts flat as scope grows. Deployments without them see consumption climb while answer quality falls, because retrieval degrades and the agent compensates by doing more work per question.

#A worked twelve-month shape

Numbers here are illustrative arithmetic on the published rates, not a quote and not a benchmark. The point is the shape, which is consistent across the engagements we scope.

Take a service agent handling 2,000 conversations a month at an average of six standard actions each. That is 12,000 actions, or 240,000 Flex Credits, which at the published US$500 per 100,000 is US$1,200 a month. On the conversation model the same volume is 2,000 times US$2, or US$4,000. At six actions per conversation the credit model is clearly cheaper, and it stays cheaper until the average passes roughly twenty actions.

Now put the year-one project costs beside it. Knowledge remediation across the in-scope intents, a permission review and restructure, one integration to reach order data, and the evaluation set with the effort to maintain it. In every engagement we have scoped, that block exceeds twelve months of consumption, often by a multiple.

That is the shape worth internalising: consumption is the small, predictable, recurring number, and remediation is the large, uncertain, one-off number. Business cases that get this backwards are the ones that fail their first budget review.

#What drives the remediation number

Estimating remediation is where scoping conversations get vague, so these are the four variables that actually move it.

Number of in-scope intents. Cost scales with intents, not with volume. Ten intents at low volume cost more to ground than one intent at high volume, which is the opposite of how business cases are usually built.

Age and ownership of the knowledge. Content with a named owner and a review history needs tidying. Content nobody has owned for three years needs decisions, and decisions need people who are hard to get in a room.

Whether the sharing model is documented. If it is, permission work is verification. If it is not, it is discovery, and discovery has no reliable estimate until it is partly done.

Whether an integration is on the critical path. One system the agent must reach and cannot currently reach converts an agent project into an integration project with an agent attached.

A useful shortcut when scoping: ask how long it would take to answer "who can see this account, and why". If the answer is minutes, permission work is small. If it is days, it is the largest line in your estimate and nobody has costed it.

#If you are buying in Australia

Salesforce publishes local rates alongside the US ones: AU$2.80 per conversation, and AU$7 per user per month for the Agentforce User License. The Agentforce add-on giving unmetered employee usage is published at AU$175 per user per month against US$125.

Two things follow. Check the AUD line rather than converting the USD one, because the published local rate is not always the spot conversion. And if your org is on a Cloud or Industry licence that already includes the add-on and an annual Flex Credit allocation, model against that allocation first: orgs routinely buy consumption they were already entitled to because nobody checked what the existing licence carried.

#The question to ask your account executive

Two things are worth establishing before any commercial conversation, because both change the arithmetic and neither is obvious from the public page.

First, what your existing licences already include. Cloud and Industry licences are published from US$550 per user per month with the Agentforce add-on included and an annual Flex Credit allocation, and organisations regularly buy consumption they already hold. Ask for a written statement of what is included on your current agreement before discussing anything incremental.

Second, which buying model your commitment can move between. Pre-purchase, pre-commit and pay-as-you-go are published as options, and the model that suits a pilot is rarely the one that suits year two. The useful question is not which is cheapest today but what it costs to change your mind once you know your real action count.

#Modelling before you commit

Three things to do in that order.

Pilot narrow and measure actions per conversation, because that single number selects the buying model and no spreadsheet can predict it. Then price remediation as its own line, honestly, and present it as platform work rather than burying it inside an AI budget where it will look like overrun. Then choose the buying model, and revisit that choice once real volume exists, because the break-even moves as grounding improves and action counts fall.

#How to build the estimate

Score readiness first. Pilot one narrow intent and measure actions per conversation. Use that number to pick a buying model rather than guessing. Then size grounding and permission work honestly, because that is where the year-one budget mostly goes, and present it as what it is: platform remediation that happens to have been triggered by an agent project, and that pays back across everything else the org does afterwards.

A business case built on licence cost against deflected contacts will be wrong in both directions at once. It will understate the first year and overstate the ongoing.