On June 25, Salesforce announced Agentforce Help Agent. Most of the coverage led with the same thing: it deploys in minutes. Fair enough, that is genuinely useful. But it buries the part that actually reshapes how you design a service org.

You pay only when the agent resolves an issue, start to finish. If the customer asks for a human or walks away unhappy, there is no charge. Salesforce calls it pay-per-resolution, and it ties your bill to outcomes instead of activity.

Sit with that for a second. For two years, the cost conversation around Agentforce has been about consumption. How many credits per conversation, how many actions, how do we forecast it, what happens at the overage. Architects spent real hours building consumption models nobody fully trusted. This pricing takes that whole anxiety and moves it onto Salesforce’s side of the table.

Here is why that matters to you specifically. When the meter only runs on success, the incentive to over-engineer a chatbot to “handle” volume disappears. A conversation that ends in escalation is free. So the design question changes. You stop optimizing for how many tickets you can keep away from humans, and you start optimizing for how many issues the agent can genuinely close.

Those are not the same goal. Deflection rewards an agent that ends conversations. Resolution rewards an agent that solves problems. The pricing now points everyone at the second one.

In an outcome model, the definition of the outcome is the contract. So before anything else, pin down what counts as a billable resolution and what does not.

From the announcement, the boundaries are clear in their shape. A resolution is the agent autonomously handling an issue from start to finish. Two things explicitly fall outside it: a customer who gives negative feedback, and a customer who asks for a human. In both of those cases there is no charge, and the agent hands the full context to your service team.

There is one more detail worth flagging, because it removes a real source of cost surprise. During the agent interaction, both Data 360 and Agentforce are unmetered. You are not watching credits tick down on every grounding query or every action call inside a conversation. The conversation runs, and the only commercial event that matters is whether it ended in a resolution.

That said, a clean definition still leaves architect-level questions. What happens when an agent partially solves something and the customer leaves satisfied without explicit feedback? How is “negative feedback” captured, and is it the customer’s signal or a model’s guess? These are the things to confirm in writing, which is exactly what section 5 is about.

Salesforce frames a good service agent as three things: it knows your business, it can act, and it works everywhere your customers are. Help Agent ships all three configured, which is the reason it sets up fast.

The portal piece is more interesting than it sounds. The Agentforce Customer Service Portal has been rebuilt around a single conversation bar. As a customer describes what they need, the page adapts live, surfacing personalized responses and dynamic cards so the task gets completed inside the conversation. Because it runs on real-time data, it can also trigger workflows proactively, reaching the customer before an issue becomes a ticket.

One line in the announcement deserves an architect’s attention. Salesforce calls messy data “the single biggest reason agents fail,” and says grounding on Knowledge takes that off your plate. Read that carefully. It does not mean your data problems vanish. It means the agent leans heavily on Knowledge being good. If your Knowledge base is thin or stale, a fast setup just gets you to a confident wrong answer faster.

If you only get paid-for outcomes, you design for outcomes. Here is the sequence I would run, and notice how different it looks from a classic chatbot rollout.

The mindset shift is the whole game. A deflection-era rollout asks “how do we keep people away from agents.” A resolution-era rollout asks “how do we make sure the agent can finish the job, and gracefully hand off when it cannot.” Same product, very different setup.

Pay-per-resolution genuinely lowers the risk of paying for nothing. It does not remove every sharp edge. Here is what a careful architect checks before committing a contact center to it.

The Knowledge Dependency Is Now a Cost Lever

In a consumption model, weak Knowledge wasted credits. In a resolution model, weak Knowledge means fewer billable resolutions, which sounds cheaper until you remember those unresolved issues still flow to your human team. You did not save money. You moved the cost to payroll and slowed the customer down. The lesson holds from my last write-up: data and content readiness decide whether any of this works.

None of this is a reason to wait. It is a reason to go in with eyes open. The model is fairer than what came before. It still rewards the orgs that did the unglamorous work on Knowledge, actions, and escalation design.

Pay-per-resolution is not a one-off pricing experiment. It is a statement about where Salesforce thinks agent value should be measured. When a vendor agrees to only get paid for resolutions, they are betting heavily on their own resolution rate. The 4.3 million inquiries and 70% resolution on their own help site are the proof point they built that bet on.

The Fin acquisition tells you the same thing from another angle. Salesforce signed a definitive agreement to acquire Fin, an autonomous customer agent platform built for small and medium businesses and trusted by more than 30,000 companies. Pair a prepackaged, fast-setup Help Agent with an SMB-focused agent platform, and the direction is obvious: get autonomous service into the hands of companies that never had the resources to build it.

For architects, the role does not shrink here, it sharpens. When setup is fast and pricing is fair, the differentiator is no longer who can stand up an agent. It is who can design Knowledge, actions, and escalation so the agent resolves the right issues and gracefully hands off the rest. The plumbing got easier. The judgment got more valuable.

So the question to take into your next planning session is not “can we deploy a service agent.” You can, quickly. The real question is sharper: when the bill only arrives on a genuine resolution, is your Knowledge good enough to earn it?