Agentforce Help Agent and pay-per-resolution: what changes when you only pay for outcomes

Every conversation about deploying a service agent used to snag on the same fear: what if it doesn’t work and I’ve paid for it anyway? Under the original Agentforce economics (roughly $2 per conversation, or a pile of Flex Credits burned per action) you paid for the agent to try, whether or not it succeeded. A confused customer who got bounced to a human still cost you. That single fact made every CFO cautious and every pilot smaller than it should have been.

In June 2026, Salesforce shipped Agentforce Help Agent and, with it, the answer to that fear: pay-per-resolution. You pay a flat $2 only when the agent autonomously resolves an issue from start to finish. If the customer asks for a human, leaves negative feedback, or abandons the chat, there’s no charge. It’s the largest shift in Agentforce pricing since launch, and it lines Salesforce up directly against the outcome-priced upstarts (Intercom’s Fin, Sierra) that made “pay for results, not attempts” the new default. It’s also more complicated than the headline, because the definition of “resolved” is doing all the work, and the party that defines it is the party that gets paid. This post is what Help Agent is, how the pricing functions, and the questions to settle before you sign.

What Help Agent is

Strip the pricing for a second, because the product underneath matters. Agentforce Help Agent is a packaged, autonomous service agent built on the Agentforce 360 platform, designed to be turned on fast and deployed everywhere a customer might reach you. Two things distinguish it from rolling your own service agent from the template:

Guided setup and omnichannel from one screen. It ships with prepackaged workflow actions and deploys across voice, web, portal, and messaging, configured from a single setup surface. Salesforce’s pitch is minutes, not a project. The prebuilt action library covers the bread-and-butter of service: manage cases, schedule appointments, update orders. If you’ve fought the Messaging for Web snippet or wired channels by hand, the packaging is the point.
A reimagined self-service portal. The customer-facing experience is rebuilt around a single conversation bar instead of a nav tree of articles. As the customer describes what they need, the portal adapts in real time, surfacing personalized responses and dynamic cards that let them complete the task inside the conversation rather than being handed a link to a form. It’s the assistive-to-autonomous shift made concrete: the portal stops being a place to search and becomes a place to get the thing done.

That’s a good service product. But it’s not why the whole industry is talking about it. The pricing is.

How pay-per-resolution works

Here’s the mechanic, precisely. A resolution, when it bills, costs $2, equivalently 400 Flex Credits. It bills only when the agent handles an issue autonomously end to end. It does not bill when:

the customer requests a human, or the conversation escalates;
the customer gives negative feedback (“this didn’t answer my question”);
the customer abandons the conversation.

And the part that’s easy to miss but changes the math: there’s no variable cost inside a resolved session. However many actions the agent takes, however many Data 360 queries it runs, however long the reasoning chain, a resolution is $2 flat. Under Flex Credits, a chatty session that hit six actions and grounded on three retrievers cost more than a one-shot answer. Under pay-per-resolution, they cost the same. The meter no longer punishes a hard-won resolution, which is exactly the behavior you want from a service agent. Dig in and solve it, don’t bail to save credits.

The comparison to the old model is the story:

	Old Agentforce Service Agent	Help Agent (pay-per-resolution)
Billing event	Per conversation (~$2)	Per resolution ($2)
Failed / escalated session	Billed	Free
Cost of a compute-heavy session	Scales with actions/credits	Flat, no variable cost
Who carries the risk of a bad agent	You	Salesforce (in theory)

On paper, this moves the risk of an underperforming agent from you to Salesforce. That’s real, and it’s why the model is attractive. But “in theory” is carrying weight, and it’s worth understanding why.

The accuracy gap nobody puts on the slide

Read the definition of “resolution” one more time, adversarially: a session bills when the issue is closed without escalation and without explicit negative feedback. Notice what that definition does not contain: any check that the answer was correct.

By default, pay-per-resolution charges for a conversation that ended without the customer objecting, not for one that was verifiably right. A customer who accepts a confidently wrong answer and hangs up, satisfied in the moment, is a billable resolution. A customer who gets the right answer but grumbles is not. The metric is a proxy for success, and like every proxy, it can drift from the thing it stands for. Salesforce Ben put the caution bluntly, and it’s the right caution: “you only pay when it works” is a headline, not a contract.

This isn’t a reason to avoid the model. It’s a reason to instrument it. Three questions decide whether pay-per-resolution is a good deal for you specifically:

What exactly counts as a resolution, in the contract? Get the billable event defined in writing, including the treatment of edge cases: a customer who resolves and then reopens the same issue an hour later; a session that “resolves” but generates a follow-up case; a multi-turn conversation that spans two channels. The headline is clean; the edges are where the invoice lives.
Who adjudicates a disputed resolution? There’s a structural incentive worth naming plainly: the party that decides whether a session was “resolved” is the party that gets paid for it. That doesn’t make the model bad, outcome pricing works across the industry, but it makes an audit path non-optional. You want to be able to sample resolutions and disagree.
How will you measure accuracy, separately from billing? Because billing won’t do it for you. This is the operational shift pay-per-resolution forces: you now have to run your own quality loop (sample transcripts, score correctness, watch reopen rates and downstream case creation) to know whether a “resolution” was real. The same testing and observability discipline you’d apply to any agent becomes a financial control here, not just a quality one.

Outcome pricing aligns the vendor’s incentive with your success, but only as tightly as the definition of “outcome” is honest. When the billable event is “customer didn’t complain,” your job is to measure the gap between didn’t complain and was helped, and to hold the contract to the difference.

The competitive context: Salesforce is following, not leading, here

Pay-per-resolution didn’t originate with Salesforce, and knowing the field sharpens the evaluation. Intercom’s Fin publishes a flat $0.99 per resolution, with no platform fees and no per-seat charge for the AI agent, plus spend caps. Sierra sells enterprise outcome-based pricing with no published rate. Against that, Salesforce’s $2 per resolution is the premium option on price, roughly double Fin’s published number.

The obvious question, “why pay double?”, has a real answer, and it’s the same answer that decides most Agentforce-vs-standalone calls. Fin and Sierra are bolt-on agents that have to integrate into your systems. Help Agent runs inside the platform where your cases, orders, entitlements, and customer history already live, grounded natively on Data 360 and acting through governed Salesforce actions under the Einstein Trust Layer. If you’re already a Service Cloud shop, the integration you’d pay for elsewhere in engineering time is the thing you’re not paying for here, and the resolution happens against live, governed CRM data rather than a synced copy. If you’re not on Salesforce, that premium is much harder to justify: the per-resolution delta is real money at volume, and Fin’s transparency is a genuine advantage. This is the build-vs-buy calculus wearing a pricing hat: the platform tax buys you grounding and governance you’d otherwise assemble.

Modeling the cost honestly

Run the numbers before the model seduces you, because “free when it fails” can become expensive when it succeeds a lot. Suppose you deflect 50,000 tier-1 contacts a month and the agent autonomously resolves 60% of them:

Monthly resolutions   = 50,000 × 0.60 = 30,000
Monthly cost          = 30,000 × $2   = $60,000
Effective cost per contact handled    = $60,000 / 50,000 = $1.20

Two things fall out of that arithmetic. First, a higher resolution rate costs you more, not less. You’re buying outcomes, and more outcomes is a bigger bill. That inverts the Flex Credit instinct where a smarter agent that resolved in fewer actions saved money. Here, success is the cost, and the savings show up in the human cases you didn’t staff, not in the agent line item. Model it as deflection value against agent spend, exactly as we lay out in the honest ROI method, or the finance conversation will go sideways when the bill scales with your win rate.

Second, the crossover with Flex Credits depends on your escalation rate. If your agent escalates or fails half its sessions, pay-per-resolution is a bargain. You were paying for all those failed conversations under the old model and now you’re not. If your agent resolves almost everything cheaply in one action, a well-optimized Flex Credit setup might come in lower. The model rewards agents with a middling success rate on expensive sessions, and penalizes agents that resolve trivially at high volume. Know which one you have before you pick the meter.

What to do

If you’re on Service Cloud and evaluating Help Agent, treat the pricing as the easy part and the definition as the work. Get “resolution” defined in the contract, edges included. Stand up your own accuracy loop (transcript sampling, reopen-rate tracking, downstream-case monitoring) so you can tell a real resolution from a customer who gave up politely, and so you can dispute the ones that weren’t. Model the cost against your actual deflection volume and resolution rate, not the demo’s, and compare it honestly to a tuned Flex Credit setup and to Fin’s published rate if you’re platform-agnostic. Deploy the packaged agent and the conversation-bar portal fast, because that part is quick, but keep a clean human handoff as the release valve, because every session that escalates cleanly is a session you don’t pay for and a customer you don’t lose.

Pay-per-resolution is a better model than paying for attempts, and for most Service Cloud teams Help Agent will be the right way to run an autonomous service agent. Just remember that outcome pricing is only as good as the outcome’s definition, and measuring the distance between “closed” and “actually solved” is now your job, not the invoice’s.

Understanding the basics

How much does Agentforce Help Agent cost?

Agentforce Help Agent uses pay-per-resolution pricing: a flat $2 per resolution (equivalent to 400 Flex Credits), charged only when the agent autonomously resolves a customer issue from start to finish. There’s no charge when the customer requests a human, escalates, gives negative feedback, or abandons the conversation, and no variable cost for the number of actions or Data 360 queries within a resolved session. It launched in June 2026 on the Agentforce 360 platform and deploys across voice, web, portal, and messaging.

What counts as a “resolution” in pay-per-resolution pricing?

By default, a resolution is a session that closes without human escalation and without explicit negative feedback from the customer. Crucially, that definition does not, by itself, verify that the answer was accurate: a customer who accepts a wrong answer and leaves can still register as a billable resolution. That’s why the practical advice is to get the billable event defined precisely in your contract, establish who adjudicates disputes, and run your own accuracy measurement (transcript sampling, reopen rates, downstream case creation) separate from billing.

How does Agentforce Help Agent pricing compare to Intercom Fin and Sierra?

Help Agent charges $2 per resolution. Intercom’s Fin publishes a flat $0.99 per resolution with no platform or per-seat fees; Sierra uses custom enterprise outcome-based pricing that isn’t publicly disclosed. Salesforce is the price premium of the three. The justification is platform-native grounding: Help Agent runs inside Salesforce where your cases, orders, and customer data already live, grounded on Data 360 and governed by the Einstein Trust Layer, so you avoid the integration work a bolt-on agent requires. If you’re not already on Salesforce, that premium is much harder to justify.

Weighing Help Agent’s pay-per-resolution against a tuned Flex Credit setup, and want the resolution definition pinned down before you commit? Talk to us. Modeling agent economics against real transcripts and real deflection volume is exactly the work we do.