Dreamforce 2026 closed at Moscone Center on 17 September after three days and, by Salesforce's own count, more than 1,600 breakout sessions and 50-plus keynotes. The two announcements that took the main stage — the AIforce interface layer and Koa, Salesforce's first CRM reasoning model — are the ones that will appear in next quarter's board decks.
Neither is generally available.
That is the whole closeout for anyone holding a Salesforce renewal. The headline reveals are a pilot and an open beta with no published price. The things you can actually put into production this quarter got far less stage time: a router, an observability console, and an optimizer. That inversion is worth understanding before you build a 2027 roadmap on a keynote slide.
The headline reveals are not products yet
Koa is Salesforce's first CRM-specific reasoning model. It is built on NVIDIA Nemotron 3 Super and post-trained on synthetic datasets derived from what Salesforce describes as 27 years of CRM knowledge, spanning more than 14 industries including financial services, healthcare and manufacturing. Salesforce says it "matches or does better than leading model performance on CRM actions with 3x fewer errors".
Two things about that claim matter more than the claim itself. The benchmark is Salesforce's own CRM Benchmark, not an independent evaluation — so "3x fewer errors" is a vendor-run result against a vendor-defined task set, and it has no published methodology you can replicate against your own org. And Koa is available to select pilot customers only, with general availability expected in winter 2026 and only in U.S. regions. If your data residency sits in the EU or APAC, Koa is not a 2026 line item at all.
AIforce is the other headline: an interface layer that carries Data 360 context, Customer 360 semantics and Agentforce actions into any AI surface. Salesforce Ben's write-up summarised the pitch as "AI replaces the UI" — you build your own front end instead of navigating Salesforce's. Its launch partners are Anthropic, AWS, Google, Lovable, Vercel, DocuSign, Gamma, Jasper, Ramp and Rippling, and requests run on the existing permission model, so an agent sees only what the person behind it can see.
The useful clarification from that same session: AIforce does not replace Agentforce. Agentforce remains the digital workforce layer; AIforce is an interface layer above it; Claudeforce is a specific surface. Three brands, one stack, and the licensing for each is negotiated separately.
Claudeforce moved to open beta, which is not the same as shipped
The Salesforce–Anthropic partnership was announced on 26 August, and at Dreamforce it moved to open beta. Applications President Patrick Stokes told the keynote audience the plugin is available from the AppExchange, and the Salesforce in Claude plugin ships with 37 prebuilt sales skills.
Open beta is a real milestone and a real risk. There is still no published price for the Salesforce side, and the consumption runs on two separate meters — Salesforce metering API calls against your org's entitlement, Anthropic billing tokens against a contract with no cap. We covered that structure in detail when the partnership was announced; nothing at Dreamforce changed it.
One more data point from the keynote worth noting for anyone modelling vendor concentration: Salesforce Ben reported that 45 Salesforce employees have joined Anthropic since the start of 2026. That is a partnership with unusually deep personnel overlap.
The operations layer is what you can actually buy
The Agentforce keynote, not the main stage, is where the shippable material was. Mark Wakelin, the GM of Agentforce, put the installed base at 30,000 customers live on the platform. What got demonstrated was the machinery for running agents that already exist:
- Agent Router — one controller that hands parts of a process to dedicated subagents. Southwest Airlines demonstrated a customer name change moving through separate subagents for intake, document validation, reimbursements and ticket refunds.
- Agentforce Observability — session logs, response quality, action traces and latency. Vivint showed it running against live voice calls, which is the hardest case to instrument.
- Agent Optimizer — the most decision-relevant demo of the week. Salesforce showed a production issue where refund requests over $200 escalated at a much higher rate than expected, traced it to a gap in the subagent and router configuration, proposed a change, and generated smoke and regression tests for the fix.
- Long-Horizon Runtime — goal planning, human-in-the-loop checkpoints, memory and durable execution for agents that run longer than a single session.
Agent Optimizer is the one to ask your account team about, because it is the only announcement that addresses the failure mode most buyers actually hit. A demo of an agent succeeding tells you nothing. A tool that detects a $200 threshold quietly driving escalations, and then writes the regression test, is aimed at the post-deployment reality where agent projects go wrong. It is the closest Salesforce came to answering the question the ecosystem had been asking going into the week.
Fulton Bank is the first proof point with real numbers attached
For two years Agentforce customer stories have been adoption counts and interaction volumes. This year one number set was specific enough to test.
Fulton Bank reported 80,000 hours saved, $389 million in attributed loans and deposits, and a four-day reduction in complaint resolution time, running an internal agent integrated with Gmail.
Treat the $389 million with the caution the word "attributed" deserves — attribution in a lending pipeline is a modelling choice, not a measurement, and no methodology was published. The 80,000 hours and the four-day resolution improvement are the more usable figures, because both are operational metrics a bank measures anyway and both are the kind of number your own team could reproduce as a before-and-after. If you are building a business case, those two are the template; the loan attribution is not.
What it costs once the agents are actually running
None of the Dreamforce announcements changed the meters, and the meters are where agent programmes overrun. Salesforce's published rate card now runs several concurrent models at once:
| Model | Published rate | What triggers a charge |
|---|---|---|
| Flex Credits | $500 per 100,000 credits ($0.005/credit) | A standard agent action costs 20 credits ($0.10) |
| Agentforce Voice action | 30 credits ($0.15) | Each voice action |
| Conversations | $2 per conversation | Each customer conversation |
| Help Agent resolution | $2 per resolution | Only a successful autonomous resolution |
| Per-user add-ons | From $125–$150 per user/month | Seat-based, on top of platform |
| Agentforce 1 Editions | From $550 per user/month | Bundled seats plus credits |
The pay-per-resolution model on the Help Agent SKU went generally available in July 2026 and is the most buyer-friendly of the set: a resolution only bills when the agent handles the issue autonomously, the customer does not ask for escalation, and the customer shows no dissatisfaction signal. Failed attempts and escalations are free.
The trap is the interaction between meters. A standard action includes processing up to 10,000 tokens; exceed that and a single action counts as multiple actions — 20,001 tokens bills as three actions, not one. An Agent Router that fans one request across four subagents multiplies that count before anything resolves. The same architecture that got stage time as an availability win is also a billing multiplier, and no session at Dreamforce reconciled the two.
What to do before the winter release
Four things, in order, for a buyer leaving this week with a roadmap to write:
- Do not budget Koa for 2026 outside the U.S. It is a pilot with a winter GA target and a U.S.-region constraint. Build your plan on the models you can already contract for, and treat Koa as an upgrade path.
- Ask for the CRM Benchmark methodology in writing. "3x fewer errors" against a vendor-defined task set is a starting point for a conversation, not a procurement input. If it cannot be reproduced against a sample of your own org's actions, it does not belong in your business case.
- Instrument before you scale. Observability and Agent Optimizer are available now; Koa is not. The ordering is convenient — the highest-value work this quarter is measuring the agents you already run, and that work does not depend on any 2027 release.
- Model the router against the meter. Before adopting multi-agent routing, price a representative transaction end to end at 20 credits per action, with the 10,000-token action boundary applied. Southwest's name-change flow touched at least four subagents.
The honest summary of Dreamforce 2026 is that Salesforce showed a coherent architecture and shipped the least glamorous third of it. For a buyer, that is better news than the reverse: an operations layer you can run against agents you already have is worth more this quarter than a reasoning model you cannot contract for until winter.
Continue Reading
- Claudeforce Has No Published Price and Two Separate Meters
- Why 88% of AI Agent Pilots Never Reach Production
- AI Agents Have Admin Access. 57% of Firms Aren't Ready.
- MCP Goes Stateless: Enterprise AI Governance Just Got Real
- AI Control Gap: Why Only 11% of CIOs Are Ready
- OpenAI Won't Sell You AI Agents — They're Sending Engineers