AI Agents vs Agentforce: What It Actually Costs

AI agents vs Agentforce is really a build-or-buy question wearing a product comparison's clothes. Agentforce is Salesforce's own agent platform, sold as a metered add-on inside Salesforce, and it can read and act on your CRM data the moment it is turned on. A custom-built agent starts with nothing and has to be wired to every system by hand, but it only does the one job you actually need and nothing you are paying for and not using. The real comparison is not features. It is what your own setup actually costs once Flex Credits, add-on tiers, and integration work are counted, and how well an agent performs on your specific job once you move past the demo. Getting the speed-to-lead setup right matters more than either vendor's pitch, because a slow handoff from a fast system is still a slow handoff.

What is Agentforce, and how does it differ from a custom AI sales agent?

Agentforce is Salesforce's agent-building layer, sold as a metered capability on top of the CRM you already pay for. It ships with pre-built actions for Salesforce records, so an agent can read an opportunity, update a field, or draft a follow-up without you writing the integration that connects an agent to your pipeline data. That native access is the entire pitch: if your sales process already lives in Salesforce, Agentforce starts closer to useful than almost anything built from outside it.

A custom-built sales agent starts from the opposite place. Nothing is wired up on day one, not even to Salesforce if that is where your deals live, and someone has to build and test every connection it needs. What you get in exchange is an agent scoped to exactly one job, lead qualification, renewal outreach, or quote follow-up, rather than a general-purpose platform licensed per seat whether every seat uses it or not. Neither description tells you which one is cheaper or which one actually gets the job done. Both of those are answerable, and the numbers below are Salesforce's own.

How much does Agentforce actually cost once Flex Credits and add-ons are counted?

More than the headline number most people quote. Salesforce's own Agentforce pricing page lists two separate consumption models, and a company can only run one of them per org. "Conversations" bills $2 for each 24-hour customer session, no matter how much the agent does inside it. "Flex Credits" bill $500 per 100,000 credits, with a standard agent action costing 20 credits and a voice action costing 30. The two cannot be mixed in the same org, and a company has to commit to one before it knows for certain which its actual usage pattern favors.

Conversations versus Flex Credits: the break-even math

Twenty credits per action at $500 per 100,000 credits works out to $0.10 per action, which means $2 buys exactly 20 actions under Flex Credits. That is the break-even point: a conversation that takes 20 or more agent actions costs less under the flat $2 Conversations rate, and one that takes fewer than 20 costs less under Flex Credits. A simple question-and-answer exchange that resolves in three or four actions is clearly a Flex Credits conversation. A multi-step qualification flow that checks inventory, pulls account history, and drafts a personalized follow-up can blow past 20 actions before the lead ever talks to a person, and that is a Conversations-priced interaction even if the headline rate looked like the cheap option on the pricing page.

A worked example: pricing a 25-action qualification flow

Say a sales agent needs 25 actions to qualify an inbound lead: look up the account, check for an open opportunity, read the last three activities, score the lead, and draft a follow-up email. Under Flex Credits, that is 25 actions times 20 credits times $0.005 per credit, or $2.50. Under Conversations, the entire session is a flat $2, because the meter counts the session, not the action count inside it. For that specific flow, Conversations wins by $0.50 per lead, which is real money once you multiply it by a few hundred leads a month, but it only wins because the flow crossed the 20-action line. The same company running a simpler flow on the same plan would want the other meter, and Salesforce will not let them run both at once.

The per-user add-on tier that pushes the real price past the per-action rate

The consumption meters are not the whole bill. The Agentforce add-on for Sales, Service, Field Service, or Operations lists at $125 per user per month with unmetered employee usage, and the Max Edition, which bundles the Agentforce add-on with 2.75 million Flex Credits per org per year, starts at $550 per user per month. A 12-person sales team on the Max Edition tier is looking at $6,600 a month before a single customer conversation runs, a number that rarely shows up in the demo but shows up on the renewal invoice every time.

Does Agentforce perform well on real CRM sales tasks, even by Salesforce's own research?

Unevenly, and Salesforce's own research team is the one that measured it. CRMArena-Pro, a 2025 benchmark built by Salesforce AI Research, tested leading LLM agents against 19 expert-validated tasks spanning sales, service, and configure-price-quote work, the exact categories a sales-focused Agentforce deployment is meant to handle. The agents completed about 58% of single-turn tasks and held above 83% on single-turn workflow execution specifically. Multi-turn tasks, the kind that require a back-and-forth instead of one clean request, dropped to about 35% success.

What CRMArena-Pro found about handling sensitive account data

The same study measured something that matters more for a CRM agent than a general chatbot: whether the agent respects the confidentiality boundaries baked into real account data, deal size, internal notes, another rep's pipeline. CRMArena-Pro found the agents it tested showed near-zero inherent confidentiality awareness, and that prompting the model to try to fix that awareness often made its task performance worse, not better. That is not a reason to rule out either an Agentforce deployment or a custom agent. It is a reason to test the specific multi-turn, sensitive-data scenarios your own team actually runs before trusting either one to touch a live opportunity unsupervised.

What does Agentforce do well that a custom-built agent would take months to match?

Day-one access to data a custom agent has to be taught to see. An Agentforce action already knows what an opportunity object looks like, what fields exist on a lead, and how your validation rules are supposed to behave, because Salesforce built the platform around its own schema. A custom agent starts with none of that, and someone has to map every field, test every write, and handle every edge case Salesforce's own engineers already handled once for every customer running their CRM. For a company whose entire sales process already lives inside Salesforce, with clean data and validation rules that actually get followed, that head start is real and it is hard to replicate quickly from outside the platform.

Agentforce's own growth numbers back up that the companies using it are finding real use for it, not just buying it and letting it sit idle. Salesforce's Agentic Enterprise Index, now in its second edition, reports that the average number of activated agents per organization grew from 5 in February 2025 to 13 by April 2026 among customers who kept agents in production every month of that window. That is a real adoption signal inside an existing, paying cohort. It is not evidence about the broader market, and the report itself states the figures "are not indicative of Salesforce performance," a caveat worth reading before quoting the 5-to-13 jump as proof that Agentforce works for every company, including yours.

When does a scoped, custom-built agent cost less and do more than Agentforce?

When your sales process does not live entirely inside Salesforce, or when the one job you actually need is narrower than the platform you would be renting to do it. A company running its pipeline out of a lighter CRM, a spreadsheet-and-email process, or a mix of tools that nothing inside Salesforce was built to read gets none of Agentforce's native-data advantage and still pays the per-user and per-action rates above. A custom agent built for one specific job, answering after-hours inbound and booking the first call, say, costs what that one job costs to build and run, with no $125-to-$550-per-user floor sitting underneath it whether the team uses the full add-on's capability or not.

The sales team Agentforce was not priced for

A 12-person team that does not run Salesforce, or runs a scaled-down edition without the add-on tiers available, is paying for a platform fit built for an enterprise CRM deployment, not for the one lead-qualification job it actually needs automated. What a B2B sales agent actually needs to do on a smaller team usually looks nothing like what the add-on tiers are priced around, and sizing the build to the job instead of the platform is the difference between a system that pays for itself in a quarter and one that sits on the renewal list getting questioned every year.

Who should own this decision, and what should they ask before the renewal call?

One named person with both the budget authority and enough technical fluency to read a pricing page past the headline number. The US Bureau of Labor Statistics counted 670,570 computer and information systems managers employed nationally as of May 2025, earning a median hourly wage of $84.20, the role that most often owns a platform-versus-build call like this one at a company large enough to have a dedicated IT or systems lead. At a smaller business without that title, the job usually falls to the owner or the sales lead who signs the renewal, and the questions are the same either way: how many actions does our actual qualification flow take, which consumption meter does that favor, and what would the equivalent job cost to build once instead of rent every month.

Before that renewal call, pull a real transcript of what your sales process needs an agent to do, count the actions it takes end to end, and run both the Flex Credits and Conversations math against your own volume rather than Salesforce's example numbers. Comparing a platform's own figures against the job it has to do is the same exercise regardless of which vendor is on the other side of the table.

Is there a middle path between buying Agentforce and building from scratch?

Often, yes. Plenty of companies run Agentforce for the service work that genuinely benefits from native CRM access, while building a narrower custom agent for the one sales job that Agentforce was never priced to do cheaply, after-hours lead response, renewal outreach, or quote follow-up, outside the platform entirely. That split avoids paying the full per-user tier for a job a scoped build can do for less, without giving up the native data access where it actually earns its price. How long that custom build actually takes is the next real question once the cost math points away from the all-in-one platform for your specific job.

If you want a second set of eyes on which side of that split your own sales process falls on, your free Systems Plan walks through what your actual lead flow needs before you commit to either path.

Methodology

This post on AI agents vs Agentforce draws on four sources. Salesforce's own Agentforce pricing page, fetched directly on 2026-11-04, supplies the Conversations rate, the Flex Credits rate and per-action cost, and the per-user add-on tiers; those are the vendor's published figures, not an independent audit, and worth confirming against a sales quote before budgeting. CRMArena-Pro (arXiv:2505.18878), a 2025 benchmark built by Salesforce AI Research itself, supplies the single-turn and multi-turn task success rates and the confidentiality-awareness finding, which makes it the one source in this post testing the underlying technology rather than describing the pricing around it. Salesforce's Agentic Enterprise Index, second edition, supplies the activated-agents-per-organization growth figure, disclosed here as the vendor's own usage data drawn from an existing paying cohort, not a market-wide survey. The US Bureau of Labor Statistics' Occupational Employment and Wage Statistics for computer and information systems managers (May 2025) supplies the headcount and wage baseline for the role that most often owns this kind of decision.