What Agentforce really costs - and when to weigh a headless alternative

The cost of Agentforce is not the cost of the licences. And before calculating it, it is worth asking whether the architecture you need is really an agent inside Salesforce.

What follows is the structure of the cost, not a price list. Salesforce commercial models change over time and a figure reported by a third party ages quickly: terms, discounts, included credits and thresholds should always be checked against your own order form.

Before the cost: which architecture do you need

The first question is not "what does Agentforce cost", but what experience you want to give users, which data it has to use and where those people already work today.

There are two roads, and they lead to two different cost structures.

The headless road now has a name: with Salesforce Headless 360 the platform's data, logic and governance become reachable from other surfaces - Slack, Teams, WhatsApp, ChatGPT - through skills, MCP, APIs and a dedicated experience layer. The point that matters to a buyer: according to Salesforce's documentation, agents inherit the permissions, compliance rules and approval chains already configured, rather than requiring a new security model.

One useful clarification, because it changes how you should put the question to your vendor: Salesforce does not present Headless 360 as an alternative to Agentforce. It presents it as an extension of the surfaces the platform can appear on - the module section is literally titled "An 'And,' Not an 'Or'". For you, though, a real choice remains, and it is about sequence: if the value of the first use case is querying and updating Salesforce from where people already work, that value can be obtained without immediately introducing a second interface to drive adoption for.

The six cost lines

1. Licences and included capacity

This is the most visible part of the offer. The useful questions are not only about year one:

• which features are included in the proposed edition?
• which ones require additional products or tiers?
• what changes at renewal?
• which credits or capacity are included, and for how long?
• which commercial terms are guaranteed in writing?

A demo can show features that require components not covered by the configuration you have been quoted. Ask for every capability shown to be mapped to the contractual line that enables it.

2. Variable consumption

Agentforce also offers consumption-based models. On the Salesforce pricing page the options are Flex Credits consumption, per-conversation consumption, or a per-user licence; one agent action is worth 20 Flex Credits and a voice action 30, and consumption is monitored through Digital Wallet. The per-credit amounts sit in the official rate card, which is dated: always check the current version.

A unit price is not enough to build a budget, though, because the cost depends on the real volume of the process. Ask the vendor to make explicit:

• which monthly volume they used in the estimate;
• which actions or interactions consume credits;
• which alert thresholds are available;
• what happens when consumption exceeds the forecast;
• whether a spending cap or approval procedure exists.

Variable consumption is not a problem in itself. It becomes one when it is not tied to a use case, an estimated volume and a periodic review.

3. Data and the data platform

If the agent only uses data already present and well structured in Salesforce, the data scope stays limited. If it has to reach ERP, ticketing, documents, knowledge bases, SharePoint or e-commerce, how that data is made available, refreshed, authorised and audited has to be designed.

Data 360 (the product Salesforce used to call Data Cloud) has its own commercial models, based on credits and on configurations that vary by usage. There is one question to ask:

Is this data component needed for the first use case, or is it deferred to a later phase?

If it is deferred, work out whether the first project is still useful without that data, or whether you are simply postponing an essential part of the cost.

4. Integrations

AI creates value when it can read reliable data and, where appropriate, take actions in the systems that run the process: ERP, warehouse, ticketing, document management, e-commerce, proprietary applications.

Integration is architectural and development work, and it should be estimated separately from licences and AI consumption:

• which systems have to be involved?
• which APIs already exist?
• which data has to be read or written?
• which security and authorisation controls are needed?
• who maintains the integration after release?

In a headless architecture this line does not disappear. What changes is the interface through which users and agents reach Salesforce; the work of connecting systems, data and processes remains.

5. Data quality and governance

Duplicates, unclear ownership, fields used differently by different teams, conflicting sources and stale data are not secondary problems: if the agent has to answer or act on that information, the quality of the result depends on the quality of the data. That is the subject of the Salesforce whitepaper on data governance for Agentforce.

Before signing, run a concrete check on the chosen process:

• sample real records;
• verify the completeness and consistency of the essential data;
• identify the system of record for every critical piece of information;
• measure duplicates, missing data and exceptions;
• define what has to be fixed before go-live.

Remediation does not necessarily have to cover the whole company: it can be bounded to the first use case. But it has to appear in the plan, if it is a precondition for a reliable result.

6. Internal time and accountability

An AI project involves business, IT, security, legal, data governance and process owners. After release somebody has to review results, feedback, unhandled cases, errors, sources and metrics - and Salesforce provides dedicated tooling to build, test and diagnose agents, which is itself a signal of how much continuous monitoring is part of the lifecycle.

The budget has to include:

• the process owner's time;
• the time of users involved in testing;
• IT, security and compliance support;
• management of the integrations;
• monitoring and improvement after release;
• training and adoption.

If these responsibilities do not already exist, the project has to set out how they will be assigned. It is not a line you can leave implicit.

How to build a business case

Ask for an estimate over at least 24 months. Year one contains the project, the release and often introductory commercial terms; year two makes renewals, actual consumption, maintenance and the costs that do not surface during implementation visible.

1. What does the process cost today? People's time, delays, errors, tickets, rework, service impact.
2. How much value can the first use case create? With a concrete metric: time saved, tickets avoided, response times, activities automated.
3. Which costs are one-off and which are recurring? Separate the project, remediation, integrations, licences, consumption and ongoing operational cover.
4. What has to be true to continue? The minimum metrics for extending, correcting or stopping, agreed before the pilot.

The useful comparison is not between the cost of Agentforce and zero. It is between the full cost of the solution and the current cost of the process you want to improve.

When to weigh a headless architecture

Before approving the investment, put this road on the table too if one of these applies:

• people already work predominantly in Slack, Teams, ChatGPT or Claude;
• the main value is querying or updating Salesforce from an interface already adopted;
• you want to test a bounded conversational use case quickly;
• you already have AI skills or platforms outside Salesforce;
• the use case needs several data sources and a tailored user experience;
• you want to keep the AI interaction layer separate from the system that holds data, rules and permissions.

This road does not remove the requirements for security, permissions, audit, testing and maintenance, and it does not reduce the integration work. What it does allow is choosing the interface based on how people actually work, rather than adding one and then having to drive its adoption.

Frequently asked

Why does the quote not include every cost?

Because the quote covers the vendor's scope: licences, configuration, development or agreed integration. Your budget has to add what stays with the company: internal people, data availability, security, remediation, testing, training and ongoing management. That is not a flaw in the quote, it is the client's planning responsibility.

How do you estimate Agentforce variable consumption?

Start from the target process: how many requests a month, which seasonal peaks, how many actions each request needs, how many users and what adoption you expect at 6 and 12 months. Then ask the vendor to translate those assumptions into the consumption model and to document what happens if volume exceeds the forecast.

Is a pilot worth running?

Yes, if it exists to reduce a real uncertainty. A good pilot uses a bounded use case, representative data, real users and metrics defined before it starts, and it has to establish whether the process is suitable, whether the data is sufficient, whether users adopt the solution and whether the value justifies extending it. Showing that the technology works in the abstract is already a demo's job.

Agentforce, or ChatGPT, Claude or Slack?

They are not necessarily competing products. Salesforce can remain the system that governs data, processes, permissions and traceability, while ChatGPT, Claude, Slack or Teams become the interface through which people and agents reach those capabilities. The choice depends on the use case, the level of integration required and the skills available to govern the architecture.

Related reading

• Is Agentforce worth it?
• Data 360 or Snowflake? - which architectural layer you need, and when you need neither.
• What to ask a partner before signing
• Readiness assessment - use case, data, integrations, governance and architecture assessed before you approve the investment.