Can you forecast your Agentforce bill three months out? For most finance teams, the honest answer is no. And that uncertainty, not the headline per-action price, is the real cost of agentic AI on the Salesforce platform today.
Salesforce has spent eighteen months reworking how Agentforce is priced. The original model launched in late 2024 at two dollars per conversation. By mid-2025 it had become Flex Credits, action-based metering, convertible licences and unlimited per-user add-ons, often all at once. The Summer ’26 release, generally available on 15 June, promises simplified pricing for Agentforce Self-Service. Simpler is welcome. Predictable is the harder problem.
The capability is no longer in question. Agentforce can resolve service cases, draft quotes and run multi-step back-office workflows, and the Summer ’26 multi-agent orchestration features push that further. The harder question, the one that decides whether a rollout survives its first budget review, is whether your organisation can predict and govern what those agents will cost.
What does Salesforce Agentforce actually cost in 2026?
Salesforce Agentforce in 2026 is priced through several models that often run in parallel. Flex Credits are the headline consumption option: roughly 500 dollars buys 100,000 credits, and a standard Agentforce action consumes about 20 credits while a voice action consumes around 30. Per-user licensing is the alternative, with unlimited Agentforce add-ons at 125 to 150 dollars per user per month and Agentforce 1 editions starting near 550 dollars per user per month. Salesforce Foundations, free on Enterprise Edition and above, includes 200,000 Flex Credits and Agent Builder so teams can pilot before committing. The list prices are straightforward. The bill is not, because it depends entirely on how much your agents actually do.
It is worth being precise about what is genuinely fixed and what is not. The per-user editions are predictable by design: you know the headcount, so you know the annual figure. Flex Credits are predictable only if consumption is predictable, and for most organisations in their first year of agentic AI it is not. The free Salesforce Foundations tier is useful precisely because it lets you generate real usage data before any money changes hands. Treat that pilot allocation as a measurement exercise, not a free trial to be spent quickly.
One nuance is often missed. The Agentforce 1 editions bundle Flex Credits and Data 360 credits into the per-user price, which blurs the line between the two models. A buyer can be paying per user and still hit a consumption ceiling. Reading the fine print on what is bundled, and what happens once the bundle runs out, matters as much as the headline figure.
What are Flex Credits, and why do buyers find them confusing?
Flex Credits are Salesforce’s consumption currency for Agentforce. Instead of paying per conversation, you pay per action: a record lookup, a generated summary, an email sent or a step in a workflow each draw down a set number of credits. Credits are bought in prepaid packs and consumed as agents work. The confusion is definitional. Buyers struggle to predict how many actions a given task will consume, because a single user request can trigger one action or fifteen depending on how the agent is built. Two functionally identical agents can carry very different running costs. Without instrumentation, most organisations cannot map a business outcome to a credit count, which makes the prepaid pack feel less like a budget and more like a meter.
This is why two organisations buying the identical Agentforce licence can report wildly different running costs. One has agents that answer in a single, well-scoped action. The other has agents that loop, re-query the same records and call external systems repeatedly. The pricing model did not change. The engineering did. Anyone forecasting Agentforce spend without auditing how the agents are actually built is forecasting in the dark.
There is a behavioural trap here too. Prepaid credit packs encourage a use-it mentality, and once an agent is live, internal teams tend to route more work through it than the original business case assumed. That is usually a healthy sign of adoption, but it quietly breaks the forecast. The fix is not to discourage usage. It is to make consumption visible in close to real time, so growth becomes a decision rather than a surprise.
Why is consumption-based AI pricing so hard to forecast?
Consumption-based AI pricing is hard to forecast because usage is both new and volatile. Three forces compound the problem. First, adoption curves are unknown: nobody can reliably say how many requests employees or customers will route through an agent in month three. Second, action counts are opaque, because the same task can consume very different credit volumes depending on agent design, retries and the data it touches. Third, the pricing surface itself keeps moving, so a forecast built on this quarter’s model may not survive the next release. The result is a budget that behaves like a variable-rate loan. The technology can deliver real value, but finance cannot sign off on a number it cannot bound.
Compare this to how the same finance team treats other software. A SaaS subscription is a known annual number. Cloud infrastructure is variable, but most organisations have years of internal benchmarks to draw on. Agentic AI has neither a fixed price nor an established baseline, which is why it unsettles budget owners far more than its absolute cost would suggest. Predictability, not price, is the real adoption barrier.
What did the Summer ’26 release change?
The Summer ’26 release, generally available on 15 June 2026, brings Agentforce Self-Service with a new Help Agent, a redesigned portal experience and what Salesforce describes as simplified pricing. The release also introduces multi-agent orchestration, letting several specialised agents work as one team across channels with shared context. For pricing, simplification is real progress: fewer SKUs and clearer entry points reduce the upfront choice paralysis that pushed many buyers to delay. But multi-agent orchestration cuts the other way. When one customer request fans out across several cooperating agents, each taking its own actions, the action count behind a single interaction rises. Simpler pricing on the page does not automatically mean a more predictable bill.
The strategic point is that orchestration changes the unit of cost. In a single-agent world, you could reason about cost per agent. In a multi-agent world, the meaningful unit becomes cost per resolved outcome, because one outcome may involve a routing agent, a service agent and a back-office agent each doing chargeable work. Buyers who keep measuring at the agent level will consistently under-forecast once orchestration is switched on.
Per-user or consumption: which pricing model fits you?
Choosing between per-user and consumption pricing for Agentforce comes down to how predictable you need the bill to be. Consumption pricing through Flex Credits is efficient when usage is genuinely low or highly variable, because you pay only for what agents do. Per-user unlimited licensing costs more on paper, but it converts a volatile bill into a fixed line item, which is why many enterprises accept the premium. A practical pattern is to pilot on consumption to learn real usage, then move heavy, high-volume agents to per-user pricing once the pattern is clear. The wrong move is committing to either model before you have measured a single agent in production.
There is no universally correct answer, and that is the point. The decision depends on volume, volatility and how much your finance function values certainty over efficiency. What does not work is treating the choice as permanent. Usage patterns in year two rarely look like year one, and the Flex Agreement option, which lets you convert between licences and credits, exists precisely because the right model shifts as you learn.
How should you budget for an Agentforce rollout?
Budgeting for an Agentforce rollout starts with instrumentation, not a spreadsheet. Before scaling, run a small set of agents in production for several weeks and measure the actual credit cost per resolved case, per quote or per workflow. That unit cost, multiplied by realistic volume, gives a defensible forecast. Then set governance: per-agent consumption limits, alerting when credit burn exceeds plan, and a named owner for every agent so usage is attributable. Treat agent design as a cost lever, because a poorly built agent that retries and over-fetches data can cost several times what a lean one does. A rollout governed this way stays inside its budget. A rollout sized by list price alone rarely does.
This is also where independence matters. The instinct, encouraged by every vendor, is to scale fast and licence generously. A disciplined rollout sometimes means telling a stakeholder that a particular agent should not go live yet, because its cost per outcome is worse than the human process it would replace. That judgement is hard to make from inside a licence-driven incentive, and it is exactly the judgement that protects the budget.
The Sirocco perspective
In our work across Salesforce, HubSpot and Dynamics 365, we see the same pattern: the technical case for agentic AI is treated as the hard part, while the financial case is left as an afterthought. As an independent CRM consultancy, we are not paid on Agentforce licences, so our advice is straightforward. Decide what you want the agents to do, instrument a pilot, and let measured unit costs rather than vendor list prices drive the rollout plan. Simplified pricing in Summer ’26 is genuinely good news, but a clear price tag is not the same as a predictable bill. The organisations that win with Agentforce are the ones that govern consumption from day one rather than discovering it on an invoice.
If you are weighing an Agentforce rollout and want a cost model you can defend to finance, it is worth a conversation before the credits start burning. You can schedule a consultation with our team to pressure-test the numbers.
Get in Touch
If you are sizing an Agentforce rollout and need a consumption forecast finance will actually sign off on, tell us where you are and we will help you build one.
