Agentforce: What Salesforce AI Agents Actually Cost

Agentforce is priced by consumption, and that single fact should change how you budget for it. Most Salesforce buyers arrive with a seat-licence mental model, ask what an agent costs per user per month, and get a number that does not describe their bill. The bill is a function of how many actions the agent takes, and that is a function of how well the agent was designed and how good your knowledge base is.

This makes it unusual among enterprise software purchases. Design decisions inside the agent are line items. An agent that answers a question in one step costs a fifth of one that grinds through five, and nothing in the licence tells you which of those you have built. You find that out from the invoice, usually in month three.

What does Agentforce actually cost? Salesforce prices it two ways. Flex Credits are USD 500 or GBP 400 per 100,000 credits, and one standard action consumes 20 credits. Conversations are a flat USD 2 or GBP 1.60 each. A standard action therefore costs GBP 0.08, so the two models break even at exactly 20 actions per conversation. Below that, credits are cheaper, and a well-designed agent sits at three to five.

What Agentforce Is, and What It Is Not

An Agentforce agent is not a chatbot with a bigger script, and it is not a workflow rule with natural language on the front. The distinguishing property is that it selects what to do rather than following a path you drew in advance.

Salesforce’s own description of the mechanism is specific. Its documentation on how agents work describes the reasoning engine as a series of orchestrated prompts sent to a large language model, where each step of classification and reasoning is one or more model calls. The agent first classifies the user’s message against the names and descriptions of the subagents available to it, using roughly the last six turns of conversation as context.

It then enters a reasoning loop. It can run an action, ask the user for missing information, ask a clarifying question, or answer directly. After each action it checks the result, adds the output to what it already knows, and decides whether to go round again. Salesforce documents this as continuing for up to seven reasoning loops.

That is the whole design brief in one paragraph. A scripted bot has a fixed cost per conversation because it has a fixed path. An agent does not, because the path is chosen at runtime.

The Pricing Model, Precisely

There are two consumption models and they cannot coexist in the same org.

Flex Credits are the model Salesforce leads with. A pack of 100,000 credits is USD 500, and Salesforce publishes the same pack at GBP 400 on its own Agentforce pricing page, so there is no conversion involved in quoting the sterling figure. One credit is GBP 0.004.

Conversations are the older model, and they came first. Salesforce added Flex Credits alongside them on 15 May 2025, and its announcement of the change describes the existing conversational pricing at USD 2 per conversation as the thing it was extending rather than replacing. Conversations are a flat USD 2 or GBP 1.60 regardless of how much work happens inside them, and the model has not been retired. It sits on the pricing page today as a live option, and Salesforce’s own FAQ explains how to switch an existing org from Conversation SKUs to Flex Credits.

Licences and editions sit on top

Alongside consumption there is licensing. The Agentforce User License is USD 5 or GBP 4 per user per month and still requires Flex Credits behind it. The Agentforce add-ons for Sales, Service and Field Service are USD 125 or GBP 100 per user per month and cover unmetered employee agent usage. Agentforce 1 Editions start at USD 550 per user per month and include 2.5 million Flex Credits per org per year.

What a credit is actually spent on

The published Flex Credits Rate Card, updated 21 April 2026, is the document to read before signing anything. It sets a standard action and a custom action both at 20 credits in production and 16 in a sandbox. Voice actions are 30 and 24. Prompt usage types are billed separately at 2, 4 and 16 credits for basic, standard and advanced prompts, and the card’s footnote says these apply in addition to the action usage types.

The card also prices Data 360, and this is where budgets go wrong. Data 360 Unification consumes 75,000 credits per million rows processed at the base tier multiplier. Four million rows exhausts the 300,000 credit base tier, which is GBP 1,200 at the published rate before the agent has answered a single question. Tier multipliers reset on the first day of each calendar month.

Speech is metered separately again. Speech to text is 150 credits per hour of transcription, text to speech is 6,000 credits per million characters, and translation is 4,000 credits per million characters. Those are small numbers next to Data 360, but they are additional to the 30 credit voice action, so a voice deployment carries at least three meters running at once and needs modelling as such.

A Worked Agentforce Cost Example

Everything below is a worked example with stated assumptions, not a quote. Change the assumptions and the answer changes, which is the point.

Take a UK retailer running a customer-facing service agent for order status and returns. Assume 1,200 conversations a month reaching the agent, three standard actions per conversation (identify the customer, retrieve the order, answer from knowledge), production environment, no voice.

Three actions is 60 credits, so one conversation costs GBP 0.24. Twelve hundred conversations is 72,000 credits a month, which is GBP 288, or USD 360. Under the Conversations model the same 1,200 conversations cost GBP 1,920. That is not a rounding difference. It is 6.7 times more expensive for identical work.

Now scale it. At 12,000 conversations a month the same agent consumes 720,000 credits, GBP 2,880 a month, GBP 34,560 a year. Still modest against the salary cost it is offsetting.

Then degrade the design. Suppose grounding is poor, so the agent averages six actions per conversation instead of three because it retries, asks clarifying questions and re-queries knowledge. The credit consumption doubles to GBP 69,120 a year. Nothing about the licence changed. Nothing about the volume changed. The agent simply thinks harder because it was given worse material to think with.

The break-even nobody calculates

One standard action is GBP 0.08. One conversation is GBP 1.60. Divide and the answer is 20. Conversations only become the cheaper model above 20 actions per conversation, and that number is currency independent because the USD figures give the same ratio.

Twenty actions in a single conversation is a broken agent, not an ambitious one. So for practically every deployment, Flex Credits is the correct model, and the Conversations model is worth keeping only where you want a fixed, forecastable unit cost and are willing to pay a large premium for it.

Why Consumption Pricing Changes the Design

On a seat licence, grounding quality is a quality problem. On a consumption model it is also a cost problem, and that reframing is the most useful thing in this article.

Every extra action is another 20 credits. An agent that cannot find the answer in your knowledge base does not fail quietly. It runs another retrieval, asks another clarifying question, and burns another 20 credits doing it. Salesforce documents the loop as running up to seven times, so the worst case for a single user turn is roughly seven times the cost of the best case.

Action design has the same property. Three narrow actions that each fetch one field cost 60 credits. One well-shaped action that returns the whole record costs 20. Nothing about the user experience differs. The bill differs by a factor of three.

This is why we treat agent design and retrieval augmented generation as the same conversation. The retrieval layer is not a nice-to-have that improves answers. On this pricing model, it is the mechanism that determines your unit economics.

The Data Prerequisite That Is Not an AI Cost

Here is the part vendors skip. An agent is only as good as the knowledge it is grounded in, and in most UK mid-market businesses that knowledge is stale, contradictory, or scattered across a wiki, a shared drive, three inboxes and one person’s head.

You cannot ground an agent in that. What actually happens is a content project: finding every place an answer lives, deciding which version is authoritative, deleting the ones that are not, and rewriting the survivors so a retrieval system can find them. In our experience this is where the majority of the project cost lands, and it is a documentation cost, not an AI cost.

It is worth saying plainly, because it changes the business case. If you have never audited your knowledge base, the honest estimate for an Agentforce programme is mostly a number for cleaning up your own content. The agent build is the cheap half.

The ICO’s guidance is a useful forcing function here, because its data minimisation guidance for AI requires you to identify the minimum personal data needed for your purpose and process no more. Doing that properly means knowing what is in your knowledge sources, which is the same audit.

Guardrails, Testing and Observability

Deciding an agent is safe to expose to customers is an engineering judgement, and it needs evidence.

Topics, or subagents in the current builder, are the primary boundary. A subagent has a scope, a set of instructions and a fixed list of actions it may call. An agent that has no matching subagent falls to an off-topic handler rather than improvising. Action boundaries are the second layer: an action that cannot issue a refund cannot be talked into issuing one, no matter how the prompt is phrased.

For customer-facing service agents, Salesforce documents a final response validation step that checks the proposed answer is grounded in the data sources, adheres to the subagent’s scope and instructions, and does not contain hallucinations, unverified claims or prompt injection artefacts. A failed response is deleted and regenerated, even mid stream.

Testing costs credits, so budget for it

Agentforce Testing Center runs batch tests across generated or uploaded scenarios and scores response accuracy, subagent and action validation, completeness, coherence, conciseness, latency and instruction adherence. Salesforce warns explicitly that running tests consumes requests and credits, and that tests can modify CRM data, so they should run in a sandbox.

The sandbox multiplier is 16 credits per action rather than 20. A thousand test scenarios averaging four actions is 64,000 credits, GBP 256 per full regression run. Run that weekly through a build and testing is a real line in the budget.

What an escalation path should look like

Escalation is a design artefact, not a fallback. It needs three properties. It has to trigger on evidence rather than on sentiment, which in practice means a rule on repeated failed retrievals, a request the subagent has no action for, or an explicit user request for a person. It has to carry the full transcript and the records the agent touched, so the human does not restart the conversation. And it has to be measurable, because the escalation rate is the only honest counter to a deflection number.

Set the threshold deliberately low at launch and raise it with evidence. An agent that escalates 40% of conversations in week one and 15% by month three is a project going well. An agent that escalates 2% from day one is usually one that has been told not to escalate rather than one that is answering correctly.

Where Agentforce Deployments Fail

Deflection measured badly. A conversation that ended is not a conversation that was resolved. If the customer abandoned the chat and rang the contact centre the next morning, the agent scored a deflection and the business paid twice. Measure resolution by whether a human ticket followed within a defined window, not by session termination.

An agent that commits the business. The failure that costs real money is an agent that offers a refund, quotes a price or promises a delivery date it had no authority to give. This is an action design problem, not a model problem. If an action can write a commitment, it needs an approval step or it needs to not exist.

Hallucination against thin knowledge. An agent grounded in six outdated articles will answer confidently and wrongly, because answering is what it does. Response validation helps and does not eliminate this.

The month three cost surprise. Pilots run inside the 100,000 free Flex Credits that Salesforce Foundations gives Enterprise Edition and above. That is 5,000 standard actions, around 1,660 conversations at three actions each, so a pilot at low volume never sees a bill. Real volume arrives and the invoice is a shock, which the Digital Wallet usage alerts exist to prevent. Note also that unused credits do not roll over into a subsequent subscription term.

Agentforce Versus a Custom Agent

The comparison a UK business should actually run is Agentforce against a custom agent built on a model API and integrated to the CRM. Both are legitimate and the decision is not close once you frame it correctly.

Agentforce buys you three things that are expensive to build. The platform integration is real: actions read and write Salesforce objects with the running user’s permissions, which is a security model you would otherwise have to design. The audit trail and the governance surface exist and an enterprise buyer’s procurement process will ask for them. And the commercial terms around the model calls are already negotiated on your behalf.

A custom build buys you model choice and per-token economics. This matters more than it sounds. Agentforce charges 20 credits for an action whether the underlying model call was 500 tokens or 50,000. That flat rate is excellent value on a heavy retrieval call and poor value on a trivial one, and a bespoke agent delivered as ordinary software development lets you price each call on what it actually consumed.

The decision rule. If the majority of your actions are reads and writes against Salesforce records, and the audit record has to live in Salesforce, buy Agentforce. If most of the work happens outside Salesforce, or your volume is high and your actions are individually cheap, a custom build will undercut GBP 0.08 per action. We walk clients through this the same way we handle any other build versus buy decision, and the answer is usually determined by where the data lives rather than by the software.

Compliance: UK GDPR and the EU AI Act

Two obligations bite on any customer-facing agent, and both need documenting before launch rather than after.

Disclosure. Article 50(1) of the EU AI Act requires providers to ensure that AI systems intended to interact directly with natural persons are designed so that those persons are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person. Article 113 of Regulation (EU) 2024/1689 states that the Regulation applies from 2 August 2026, so this is live law, not a future deadline. It reaches UK businesses whose agents serve customers in the EU.

Automated decisions. The ICO’s guidance on rights related to automated decision making sets out that solely automated decisions with legal or similarly significant effects are only lawful where necessary for a contract, authorised by domestic law, or based on explicit consent, and require a simple route to human intervention and to contest the decision. A DPIA is expected.

What goes into the context window

One Agentforce-specific detail belongs in your DPIA. Salesforce documents that in the Einstein Trust Layer, pattern-based and field-based data masking is disabled for agents, because masking removes the context an agent needs to answer accurately. The zero data retention contract with the model provider still applies, and Salesforce-hosted models keep data inside the trust boundary, but masking is off. Record that, rather than assuming the platform strips personal data for you.

An enterprise buyer will also ask for the retention position on conversation transcripts, the escalation path to a human, and evidence that the agent’s scope was tested. Have those written down before the first customer touches it. Our wider AI integration services work usually starts with this documentation set, because it is the part that blocks a launch.

What Implementation Actually Costs

These are house estimates for UK delivery, excluding Salesforce licences and consumption, and they assume one well-scoped customer-facing agent rather than a platform rollout.

Discovery and use-case scoping runs GBP 4,000 to GBP 8,000 for one to two weeks, and its job is to kill the use cases that will not pay back. Knowledge base remediation is GBP 8,000 to GBP 30,000 and is the line most likely to blow out, because nobody knows how bad their content is until someone reads all of it.

The agent build itself, meaning subagents, instructions, actions, grounding configuration and the escalation path, is GBP 12,000 to GBP 35,000. Integration to systems outside Salesforce adds GBP 6,000 to GBP 20,000. Testing, guardrail design and observability is GBP 5,000 to GBP 12,000.

A realistic first-year total is GBP 35,000 to GBP 90,000 of professional services, plus consumption. Ongoing tuning and monitoring runs GBP 1,500 to GBP 4,000 a month, and it is not optional, because instruction quality drifts as your products and policies change. For comparison against other AI work, our AI integration cost guide covers the same bands for non-Salesforce builds.

Deciding Whether to Proceed

Agentforce is a reasonable purchase when your customer data already lives in Salesforce, your knowledge base is in decent shape or you are willing to fund fixing it, and you have a use case where three actions genuinely resolve a customer’s problem. It is a poor purchase when it is being bought to avoid a documentation project, which is the most common reason it disappoints.

Run the arithmetic before you commit. Multiply your realistic monthly conversation volume by your realistic action count by GBP 0.08, then double it for the design being worse than you hoped. If that number still works, the project works. The same discipline applies to any agent programme, which is why we wrote separately about what AI agents cost and where they fail.

Mecanik builds and integrates agents on Salesforce and elsewhere, and the work usually starts with the knowledge audit rather than the agent. If you want the cost model run against your actual volumes before a contract is signed, our AI integration services and software development teams do exactly that, and the answer is sometimes that you should not build it.

Frequently Asked Questions

How much does Agentforce cost per action? One standard Agentforce action consumes 20 Flex Credits. Salesforce sells credits at USD 500 or GBP 400 per 100,000, so a standard action is USD 0.10 or GBP 0.08 in production and 16 credits in a sandbox. Voice actions consume 30 credits, or USD 0.15 and GBP 0.12.

Is Agentforce per conversation pricing still available? Yes. Conversation pricing at USD 2 or GBP 1.60 per conversation remains a live option on Salesforce’s pricing page and has not been retired. It cannot run in the same org as Flex Credits, but an existing org can switch by swapping its Conversation SKUs for Flex Credit ones.

Which Agentforce pricing model is cheaper? Divide the conversation price by the action price and you get 20, so Conversations only become cheaper above 20 actions per conversation. Most well-designed agents resolve a request in three to five actions, so Flex Credits are substantially cheaper for almost every real deployment.

What is the difference between Agentforce and a chatbot? A chatbot follows a path you scripted in advance, so its cost per conversation is fixed. An Agentforce agent selects which actions to invoke at runtime, looping up to seven times as it checks results and decides what to do next, so both its behaviour and its cost vary with how well it was designed.

Do I need to tell customers they are talking to an AI agent? Article 50 of the EU AI Act requires providers to inform people that they are interacting with an AI system unless that is obvious, and it has applied since 2 August 2026. UK businesses serving EU customers are in scope, and the ICO separately expects safeguards where automated processing has legal or similarly significant effects.