On September 11, 2026 — four days before the Dreamforce keynote — Salesforce stopped asking customers to build their own AI agents and started shipping finished ones with names. Casey handles service. Piper qualifies inbound leads. Hunter works outbound pipeline for weeks at a time. Marshall runs back-office processes. Six of the seven are generally available today. For CRM buyers, the interesting question is no longer “can we build an agent?” It’s “what does it cost per action, who owns its mistakes, and does it need data we don’t have yet?”

Key Takeaways

• Salesforce announced seven pre-built, job-scoped Agentforce agents on September 11, 2026: Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin. All are GA except Hunter, which is in pilot with GA targeted for November 2026.
• The underlying technical shift is the long-horizon runtime — memory across sessions, durable execution, and dynamic steering — which lets an agent pursue a goal over days and weeks rather than one conversation.
• Pricing is consumption-based: Flex Credits at $500 per 100,000 credits, where a standard action costs 20 credits ($0.10) and a voice action 30 credits ($0.15). One action covers up to 10,000 tokens; past that, you are billed for multiple actions.
• Fin arrives via acquisition. Salesforce closed its $3.6B purchase of Fin (formerly Intercom) on September 10, 2026, bringing 30,000+ customers and a claimed 76% average resolution rate.
• Governance is a separate, later story: the Trusted Enterprise AI Harness and its AI Control Plane were previewed September 10 but roll out in early fiscal FY28 (early 2027).
• The ROI gap is real. McKinsey’s 2026 survey found 80% of workers say AI improved their personal productivity, but only 37% say it contributed to organizational EBIT — essentially flat year over year.

What Salesforce Actually Announced

Salesforce is calling these agents “job-ready,” and the phrase is doing real work. Each one ships with pre-built skills, actions, and a data model scoped to a business function, connected to Customer 360 and operating inside your existing business rules and permissions. You configure and constrain them rather than assembling them from scratch in Agent Builder.

Alongside the agents, three platform pieces matter more to implementers than the names do. Agent Script is Salesforce’s open-source language for agent behavior, letting you blend LLM reasoning with deterministic rules — the practical answer to “how do we stop it from improvising on a refund policy.” Multi-Agent Orchestration routes work between agents so Piper can hand a qualified lead to a seller or to Hunter. And Agent Optimizer (GA October 2026) handles building, testing, and reading session traces — the closest thing yet to a QA harness for agent behavior. AI Skills in Agentforce Coworker, which let an employee teach a task once and reuse it, is in pilot with GA in October 2026.

Salesforce published customer numbers with the launch, and they’re narrower than they first look: Engine reports 50% of chat inquiries fully resolved, Autism Queensland 70% of administrative requests resolved, Hibbett 90% of core shopper journeys handled, and Perk says 60% of its sales pipeline is now built by Hunter. Anthropic reports 79% of conversations resolved autonomously. Every one of those is a containment or coverage metric — not a margin metric. Keep that distinction in mind when you build your own business case.

The Part Nobody Puts in the Keynote: What It Costs to Run

Consumption pricing is where agent projects quietly go over budget, so here is the current rate card as Salesforce publishes it.

Flex Credits cost $500 per 100,000 credits — $0.005 per credit. A standard production action consumes 20 credits ($0.10); the same action in a sandbox costs 16 credits ($0.08). Voice actions are 30 credits ($0.15). Prompt-only work is cheaper: basic prompts 2 credits, standard prompts 4, advanced prompts 16. Help agent resolutions are billed at 400 credits ($2.00). Speech and translation meter separately — transcription around $0.75 per hour, translation around $20 per million characters.

Two caveats deserve to be in your budget model, not your footnotes. First, one action covers up to 10,000 tokens of processing. An action that consumes 15,000 tokens bills as two actions ($0.20); past 20,000 tokens, three. That means a verbose system prompt, a fat knowledge article, or an over-stuffed context window doesn’t just slow the agent down — it multiplies your bill. Second, unused Flex Credits do not roll over into the next subscription term. Over-committing is a real cost, not a safety margin.

On the licensing side: an Agentforce User License is $5 per user/month but still requires you to buy Flex Credits. Agentforce add-ons run $125 per user/month for Sales, Service, Field Service, and Operations, or $150 for Industries Clouds, and include unmetered employee agent usage. Agentforce 1 Editions start at $550 per user/month and bundle 2.5M Flex Credits annually per org. The legacy $2-per-conversation model is still available as a pre-purchase option — but Flex Credits and Conversations cannot coexist in the same org, so that’s an architectural decision, not a billing preference.

Run the math on a mid-market service desk. Say 3,000 customer conversations a month, and each one averages eight metered actions end to end (classify, retrieve, respond, look up an order, post an update, and so on). That’s 24,000 actions × 20 credits = 480,000 credits × $0.005 = $2,400 per month, or roughly $0.80 per conversation. Now assume your knowledge articles are long and half those actions blow past the 10,000-token line: you’re at $3,600. The difference between those two numbers is content hygiene and prompt design — which is to say, implementation quality is now a line item on your invoice.

How This Compares With HubSpot and Zoho

Salesforce is not alone in shipping pre-built agents, but the three major CRM vendors have landed on meaningfully different meters — and the meter shapes your risk.

If you only take one thing from that table: outcome pricing and action pricing fail differently. HubSpot’s $0.50-per-resolved-conversation model caps your downside when an agent flails. Salesforce’s per-action model means a badly scoped agent that loops, retries, and re-retrieves is billable the whole time. Neither is wrong — but the Salesforce model demands the observability that Agent Optimizer provides, and it demands that someone actually watches it.

The Governance Gap You’ll Feel First

The day before the agent launch, Salesforce previewed the Trusted Enterprise AI Harness — six capabilities (Trusted Context, Agency, Action, Governance, Security, and Models) plus a new AI Control Plane to discover and register agents, assign identity and policy, manage lifecycle, evaluate performance, observe behavior, and control cost across both Salesforce and third-party AI.

This exists because the multi-vendor reality is already here. VentureBeat Intelligence’s July 2026 Agentic Orchestration Pulse Survey (107 qualified respondents) found enterprises running an average of 3.1 agent orchestration platforms, with 85% running two or more, and 53% expecting their primary control plane by end of 2026 to combine provider-native and external orchestration. As Salesforce president and chief platform officer Rohan Kumar framed it, model intelligence is becoming ubiquitous; the differentiator is your proprietary context.

The timing problem is straightforward: the agents are GA now, and the unified governance layer begins rolling out in early fiscal FY28 — early 2027. If you deploy Casey and Piper this quarter, your audit trail, cost controls, and agent inventory are your responsibility in the interim, built from Agent Optimizer traces, platform event logs, and permission-set discipline.

CRM Experts Online’s Perspective

We implement these systems for small and mid-market companies, and we’ll say plainly what we tell clients on discovery calls: pre-built agents remove the build cost, not the readiness cost. Casey resolving 70% of requests at one organization tells you what’s possible when the knowledge base is clean, the case taxonomy is coherent, and entitlements are modeled. If your help content lives in a shared drive, your service console has 40 required fields nobody fills in, and your order data sits in an ERP nobody has integrated, a named agent will simply fail faster and more expensively than a custom one would have.

Three positions we’d defend in a room full of stakeholders:

1. Pick one agent, one journey, one meter. Not a platform rollout. Take the single highest-volume, lowest-variance workflow you have — order status, password resets, inbound demo requests — and instrument it end to end before you touch the second one. Coverage metrics mean nothing without a per-transaction cost next to them.
2. Treat Agent Script as an obligation, not an option. Anywhere there’s a policy — refund windows, discount thresholds, eligibility rules, regulated disclosures — encode it deterministically. Let the model handle language and routing; never let it handle entitlement.
3. Budget the token diet. The 10,000-token action boundary means knowledge-base cleanup, retrieval scoping, and prompt compression are cost-reduction projects with a measurable payback, not hygiene tasks to do later. We’ve seen this single line item move projected run-rate by 40–50%.

And if you’re running HubSpot or Zoho rather than Salesforce, the lesson transfers intact. Zoho’s 25+ marketplace agents and 700+ Zia Agent Studio actions will get a mid-market company to a working agent quickly and cheaply; HubSpot’s outcome pricing gives you the most predictable pilot budget of the three. The decision that matters isn’t which vendor named their agent better — it’s whether you have the clean data, clear policy, and named human owner that any of them require.

Common Mistakes We Expect to See in the Next Two Quarters

• Deploying without a cost ceiling. Set a Flex Credit budget alert and a hard escalation rule before go-live, not after the first invoice.
• Mixing pricing constructs. Flex Credits and Conversations can’t coexist in one org. Choose deliberately, with a three-year volume projection in hand.
• Over-committing credits. Unused credits don’t roll over. Start with PayGo or a modest pre-purchase and true up once you have real consumption data.
• Skipping the sandbox. Sandbox actions cost 16 credits instead of 20 — testing is 20% cheaper than production. There is no excuse for learning in front of customers.
• Measuring containment instead of margin. A 60% resolution rate that costs more per ticket than a human agent is a demo, not a deployment.
• No agent inventory. With the AI Control Plane a 2027 story, keep your own register of every agent, what data it touches, whose permissions it inherits, and who approves changes.

FAQ

Do I need Agentforce 1 Edition to use these agents? No. Salesforce offers several paths: a $5/user/month Agentforce User License paired with Flex Credits, per-cloud add-ons at $125/user/month ($150 for Industries), or Agentforce 1 Editions starting at $550/user/month with 2.5M Flex Credits included annually per org. Most mid-market buyers should model the add-on plus metered credits first.

What does one “action” actually mean? A discrete step the agent takes — a retrieval, a record update, a reply — covering up to 10,000 tokens of processing. Exceed that and it bills as multiple actions. Standard actions are 20 credits ($0.10) in production, voice actions 30 credits ($0.15).

Is Hunter available now? Hunter, the outbound sales agent, is in pilot with general availability planned for November 2026. The other six — Casey, Fin, Piper, Carter, Paige, and Marshall — are generally available.

What is the long-horizon runtime, in plain terms? It’s what lets an agent work a goal across days and weeks instead of a single session: memory that preserves context and progress, durable execution that keeps plans running, and dynamic steering that adapts behavior from feedback. It’s also why multi-week agents need cost monitoring — a long-running plan is a long-running meter.

Does the Fin acquisition change our Intercom contract? Salesforce completed the $3.6B acquisition of Fin (formerly Intercom) on September 10, 2026, months ahead of its original Q4 FY2027 schedule, bringing 30,000+ customers into the fold. Existing Fin customers should expect roadmap and packaging convergence with Agentforce; get renewal terms and data-residency commitments in writing before your next cycle.

When can I govern Salesforce and non-Salesforce agents in one place? The Trusted Enterprise AI Harness and AI Control Plane were previewed September 10, 2026, with the unified experience rolling out beginning early fiscal FY28 — early 2027. Until then, plan on your own inventory and controls.

How does this compare with HubSpot’s pricing if we’re on HubSpot? HubSpot moved to outcome-based pricing on April 14, 2026: roughly $0.50 per resolved conversation, $1.00 per recommended lead, and $0.10 per data answer, with Customer Agent requiring Service Hub Professional or Enterprise and Prospecting Agent requiring Sales Hub Professional or Enterprise. Included monthly credits generally expire rather than carrying forward.

Can a small business realistically use any of this? Yes, but start where the meter is forgiving. Zoho’s Agent Marketplace (25+ deployable agents) and prompt-based Zia Agent Studio are the lowest-friction entry point for a 10–50 person company. On Salesforce, Foundations gives you Agentforce Builder access to experiment before committing to credits.

What’s the honest ROI picture? Mixed, and worth being candid about. McKinsey’s 2026 survey found 80% of workers reporting individual productivity gains from AI but only 37% reporting a contribution to organizational EBIT — flat year over year. Agents can absolutely pay off; they pay off when someone owns the measurement.

Conclusion

The seven named agents are a genuine step forward: they shorten time-to-first-value from months of building to weeks of configuring, and Agent Script plus Agent Optimizer finally give implementers the control and observability that first-generation agent projects lacked. But the pricing model has moved the risk. It now sits in your data quality, your prompt design, your knowledge base, and your willingness to watch a per-action meter. Salesforce’s own governance answer doesn’t fully land until 2027, which means the next two quarters belong to teams that instrument carefully rather than deploy broadly.

If you’re evaluating a pre-built agent on Salesforce, HubSpot, or Zoho this quarter, we’ll build the cost model with you before you sign — real action-volume estimates, a data-readiness assessment, and a scoped pilot with a measurable exit criterion. Schedule a consultation with CRM Experts Online and let’s find out whether your CRM is ready to hire an agent, or needs a month of cleanup first.

Further Reading

• Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work
• Salesforce Introduces the Trusted Enterprise AI Harness
• Salesforce Completes Acquisition of Fin (September 10, 2026)
• Agentforce Pricing — Flex Credits, Conversations, and Editions
• The New Agentforce Pricing Model: From $2 Conversations to $0.10 Actions
• SiliconANGLE: Salesforce introduces new AI agents to automate sales, support tasks
• Unite.AI: Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime
• VentureBeat: Companies already run 3 agent platforms — Salesforce’s Enterprise AI Harness wants to govern all of them
• HubSpot: Customer Agent and Prospecting Agent — now you pay when the task is complete
• Zoho: Zia LLM, Prebuilt Agents, Agent Builder, MCP, and Marketplace

John M Perez

CRM & ERP Enterprise Technology Expert and Entrepreneurial Executive with 20+ years of leading CRM, ERP, Customer Experience, and Block-chain initiatives and projects across internal and customer facing technologies.  Proven success in closing large deals in Pre Sales customer facing engagements and deploying enterprise wide CRM & Customer Experience solutions internationally and domestically.