Salesforce spent this summer removing the two largest obstacles that sat in front of customer service AI, and the result is worth attention even if you never intend to buy a Salesforce licence. Standing up a support agent used to be a project with a scoping phase attached to it. In the Summer ’26 release it takes six clicks or fewer. Paying for one used to mean committing to a consumption forecast that nobody in finance could produce with confidence. The new Agentforce Help Agent charges a flat two dollars when it autonomously resolves a case, and nothing at all when it does not.
That is a strong position for a buyer to be in. The cost of trying agentic service has fallen close to zero, and the vendor has taken on a share of the outcome risk that used to sit entirely on the customer side of the contract. What that leaves is the genuinely interesting part of the story. With deployment and pricing largely solved, the difference between a service agent that resolves three quarters of contacts and one that resolves a third is no longer the model or the platform. It is the quality of the knowledge the agent reads, and the care taken over what happens the moment it hands a conversation to a person.
Both of those are things a service organisation can improve inside a single quarter. Better still, both keep their value whichever platform you end up running, which makes this an unusually safe piece of groundwork to start now.
Setting up a support agent now takes six clicks
The Summer ’26 release reworked Agentforce Self-Service around three pieces: a new Help Agent, a new Portal experience, and a simplified pricing model. The Help Agent can be configured in six clicks or fewer, and it runs either on the new Portal or on your own website. The Portal itself is an agent-first experience with a conversational, personalised interface, rather than a search box bolted onto a knowledge library and a case form.
Alongside it, Multi-Agent Orchestration moved from beta to general availability on 15 June 2026. That capability lets several specialist agents work as a team on one request, sharing context between them, so the customer keeps a single point of contact and never has to repeat themselves or work out which agent handles which question. It is a quiet but meaningful change in architecture. The older pattern was one broad chatbot trying to cover everything. The new pattern is a set of narrow, well-scoped agents that hand work between themselves, which is both easier to govern and easier to improve one piece at a time.
The practical effect of all this is that the pilot conversation changes shape. A service leader no longer has to argue for a budget line and a delivery window in order to find out whether agentic service works on their content. They can find out in an afternoon, on a narrow slice of their contact volume, and bring evidence to the budget conversation instead of a business case built on vendor benchmarks.
What is Salesforce Agentforce, and what does the Help Agent do?
Salesforce Agentforce is the platform layer Salesforce uses to build, deploy and govern AI agents that act inside CRM data rather than only generating text about it. An Agentforce agent is given a role, a set of permitted actions, guardrails for when it must stop, and grounding in your own records and knowledge articles. The Agentforce Help Agent is a packaged version of that aimed squarely at customer support. It answers product and account questions on a self-service channel, carries out permitted actions such as looking up an order or updating a case, and escalates to a human with the full conversation context attached when it reaches the edge of what it is allowed or able to do.
Why pay-per-resolution pricing favours the buyer
The Help Agent is billed at a flat two dollars per autonomous resolution. If the customer asks for a person, leaves negative feedback, or abandons the conversation, there is no charge, and the agent passes the full context across to the service team regardless. That is a meaningful departure from how most AI pricing has worked so far, and it is worth being precise about why.
Consumption pricing charges for effort. Every action the agent takes draws down credits whether or not the customer left satisfied, which puts the forecasting risk on the buyer and quietly rewards a chatty agent. Outcome pricing charges for results, which lines the vendor’s incentive up with yours: Salesforce only earns when the agent genuinely closes the loop. For a service leader that turns an awkward budget request into a straightforward one, because the cost per resolution sits next to the fully loaded cost of a human-handled contact and the comparison speaks for itself.
It also makes a phased rollout much easier to defend. You can point the agent at one product line or one language, watch the resolution count, and expand on evidence. The unit economics scale with success rather than with ambition, which is a far more comfortable way to grow a programme than committing to a credit pool twelve months ahead of knowing how well the thing performs on your content.
What is CRM data quality, and why does it decide agent performance?
CRM data quality is the degree to which the records and content in your CRM are accurate, complete, current, consistent and non-duplicated, judged against the decisions people and systems actually make with them. For a service agent it covers three things in particular: whether your knowledge articles describe current product behaviour, whether case records carry clean resolution notes the agent can learn the shape of, and whether entitlements and account data are reliable enough for the agent to act on rather than merely read. An agent grounded in accurate, well-structured content resolves a high share of contacts. The same agent on stale content produces confident answers to the wrong version of the question.
The published numbers make the point better than any argument does. Agentforce has handled 4.3 million enquiries on the Salesforce Help site and resolved roughly 70 per cent of them without a person, and Salesforce reports an average end-to-end resolution rate of about 76 per cent for the Help Agent. 1-800Accountant has reported that Agentforce now helps resolve up to half of incoming requests. Those are not results that came from a better prompt. They came from support content that was already maintained as a product in its own right.
This is the good news buried in the technology becoming easy. Knowledge curation was always worth doing. It shortens handle time for human agents, it improves search on your own site, and it is the single most reliable predictor of how a service agent will perform. A team that spends a few weeks retiring outdated articles and rewriting the twenty most-read ones to answer one question each has improved its human service operation and prepared its agentic one at the same time. Very little technology work offers that kind of double return.
How do you measure the return on a service agent?
Measure a service agent on resolution rate, cost per resolved contact, escalation quality and customer satisfaction after escalation, and track all four from the first week rather than at the end of a pilot. Resolution rate tells you how much of the contact volume the agent genuinely closes. Cost per resolved contact compares directly against the fully loaded cost of a human-handled case. Escalation quality measures whether the handover arrives with enough context that the person picking it up does not restart the conversation. Satisfaction after escalation is the one most teams skip, and it is the number that tells you whether the agent is protecting the customer relationship or quietly eroding it.
Outcome-based pricing makes three of those four much easier to calculate, because the billing record is itself a resolution record. The thing worth adding is a view of what the agent chose not to attempt. Every escalation is a free piece of research telling you which topics your knowledge base does not yet cover well, and teams that review that list monthly tend to see their resolution rate climb steadily rather than plateau after launch.
The escalation path is where trust is won
The most common improvement available to service teams running agents today is not a smarter agent. It is a better handover. Customers are markedly more forgiving of an agent that says it cannot help and moves them cleanly to a person than of one that keeps trying. Salesforce has built for that: when the Help Agent escalates, the full conversation goes with it, and there is no charge for the attempt. Designing the escalation deliberately, deciding which topics should route straight to a human, which queue receives them, and what the agent says as it hands over, is a short piece of work with an outsized effect on satisfaction.
Adoption follows the same logic on the internal side. Service agents adopt AI readily once they trust that it improves the cases arriving in their queue rather than skimming the easy ones and leaving them a harder mix with the same handle-time target. The teams getting this right involve their most experienced agents in reviewing what the AI escalates, give them a direct route to fix the knowledge article behind a bad answer, and adjust handle-time expectations to reflect the changed case mix. Do that and your best people become the improvement engine for the agent, which is exactly where you want that expertise pointed.
The Sirocco perspective
We work across Salesforce, HubSpot and Microsoft Dynamics 365, and we like seeing capability of this kind ship as standard. When a vendor removes the setup barrier and takes on part of the outcome risk, it raises the baseline for every buyer in the market, including those who never move to that platform. Our advice to service leaders right now is to take the invitation. Run a narrow pilot on one product area, measure the four numbers above, and let the results rather than the roadmap decide how far you go.
An independent CRM partner is a consultancy that implements and advises across multiple CRM platforms without reselling any single vendor’s licences, so its recommendation is shaped by your requirements rather than by a quota. In a moment like this one, where three major platforms are shipping strong agentic service capability within months of each other, that independence is practically useful rather than merely principled. It means the question we help clients answer is which capability fits your contact mix, your content estate and your team, not which badge is on the box.
The underlying shift here is a good one. The work that decides whether service AI succeeds has moved from procurement and configuration to knowledge and service design, and those are things your own team already understands better than any vendor does. That is a far better place for the differentiator to sit, and the organisations that start curating now will be the ones quoting resolution rates in twelve months while everyone else is still scoping a pilot. If you would like a second opinion on where to point that effort first, we are happy to help.
Get in Touch
If you are weighing up an agentic service pilot, whether on Agentforce or on the platform you already run, tell us about your contact mix and knowledge estate and we will tell you where the quickest wins sit.
