The demo is free. The agent is not. Every Agentforce conversation eventually arrives at “and what does it cost to run?” — usually two weeks after the pilot got approved, when someone in procurement asks the question the keynote skipped. This guide explains where the bill actually comes from, why nobody can honestly quote you a universal number, and how to model your number before you commit to anything.
Key takeaways
- Agent usage is metered through consumption credits (Flex Credits) on top of your platform licence — the licence gets you the capability, consumption pays for the usage.
- The bill is driven by volume × action complexity: how often agents act, and how much work each action does. Content generation, campaign drafting and segment queries meter differently.
- Editions include allowances that cover early experimentation for many teams — which is precisely why costs feel invisible in month one and visible in month six.
- Anyone quoting you a universal €-per-month figure is guessing. The honest method is: read your order form, count your expected actions, model the consumption — then pilot with the meter visible.
- The same modelling logic applies to Data Cloud credits underneath — our free credit estimator lets you play with the drivers yourself.
Where the money actually goes
Salesforce’s AI pricing has converged on a consumption model. Conceptually there are three layers, and confusion between them is where most budget surprises are born:
- The platform licence. Marketing Cloud Next (Growth or Advanced) or Account Engagement with the AE+ path. This is the fixed part — it entitles you to the agentic capabilities at all.
- Consumption credits. Flex Credits are the metering currency for agent actions: a campaign drafted, content generated, a conversation handled. Different action types consume at different rates, and the rates are set in your contract’s rate card — not on a public price list you can Google reliably.
- The data layer underneath. Agents grounded in unified customer data ride on Data Cloud, which meters its own credits for ingestion, identity resolution, segmentation and activation. If your agent use case needs real-time segments, this layer quietly becomes part of the “AI cost”.
Why this article contains no price table
Because I would be inventing it. Credit rates vary by contract, edition, region and negotiation date, and Salesforce adjusts published rates as the products mature. A consultant who quotes you “Agentforce costs €X per month” without reading your order form is marketing, not modelling. What follows instead is the method that produces a defensible number for your org.
The honest modelling method
Four steps, one afternoon, no vendor required:
- Read your actual entitlements. Your order form (or your account team, asked in writing) tells you: which credits are included in your edition, at what volume, and what overage costs. This single step eliminates most surprises.
- Count the actions, not the users. Estimate per month: campaigns drafted, content pieces generated, segment queries, agent conversations. Multiply by working frequency honestly — the marketer who “will use it daily” uses most tools twice a week.
- Model the data layer separately. If the use case needs Data Cloud, model ingestion volume, identity-resolution runs and segment refreshes on their own meter. Our Data Cloud credit estimator exists exactly for this — the rates are editable fields, so you can plug in the numbers from your own quote.
- Pilot with the meter visible. Run the first use case four weeks with consumption monitoring on from day one. Real consumption data from a scoped pilot beats every forecast — and it is the number your CFO will actually believe.
The patterns that inflate bills
Watch for these in your own plan
- Unbounded conversational use cases. A website agent answering unlimited public traffic meters very differently from an internal campaign assistant used by five marketers. Volume caps and escalation rules are cost controls, not just quality guardrails.
- Segment refresh frequency. “Real-time” is a consumption multiplier. Most B2B use cases are fine with daily refreshes at a fraction of the metering.
- Regeneration loops. Ungrounded agents produce output teams reject and regenerate — paying per attempt. This is the unglamorous financial argument for proper grounding and brief templates: fewer retries, lower consumption, better output.
- Pilot sprawl. Three parallel “experiments” without owners burn allowance without producing a decision. One scoped pilot with a go/no-go gate is cheaper and more conclusive.
Budgeting rules of thumb that survive contact with reality
Without inventing your numbers, three planning behaviors consistently hold up:
- Treat included allowances as your experimentation budget, not as “free forever”. Plan the paid-consumption decision for the moment the pilot proves value — not as a surprise when the allowance runs out.
- Put consumption review into the monthly marketing ops routine from day one. Five minutes a month; prevents the six-month surprise.
- Negotiate at renewal with usage data. Consumption pricing cuts both ways — real usage numbers from your own org are leverage nobody can argue with.
Where this fits in the bigger decision
Cost modelling is one input into the readiness picture, not the whole of it. If you are earlier in the journey: score your org’s foundations with the free AI Readiness Score, check whether your specific use case should be an agent’s first job with the use-case readiness check, and read grounding & guardrails before anything ships. If you are on Pardot and wondering how you even get access to agents without migrating — that is exactly what Account Engagement Plus is for.
And if you want the modelling done for you: consumption cost modelling is part of every Agentforce engagement at MWCS — because a pilot whose running costs are a surprise is a pilot that dies at the CFO’s desk, regardless of how well it worked.
Seeing this in your own org?
The first call is free: thirty minutes, no slides, and an honest read on what a fix would take. If it is not worth doing, you will hear that too.