This year's Dreamforce in San Francisco produced the usual wave of product names, and most professional services leaders might forget half of them by the time the news cycle turns over. What's worth carrying forward isn't the name of any single feature, it's the shape of the argument Salesforce spent three days making about how AI belongs inside a business that runs on live client work, and that argument has a lot to say about how PS teams should be thinking about delivery in 2027.

Trust Became the Headline, Not Intelligence

The loudest new product was AIforce, a layer Salesforce is positioning above Agentforce, Data 360, and the rest of the platform, meant to let people work with the system through whatever interface they're already in. Underneath it sat something more telling. Salesforce paired the announcement with a governance framework it's calling the Enterprise AI Harness, covering identity, permissions, observability, and traceability for agents that don't just answer questions but modify records and take action on their own. That distinction, an AI that answers versus an AI that acts, is exactly the one delivery leaders have been living with quietly for the past year. A chatbot that gets something wrong is an annoyance. An agent that updates a project record or reallocates a resource based on a stale read of the data is a different kind of risk, and Salesforce spending a full keynote block on governance could signal an admission that the market has caught up to that concern.

For a PS org evaluating any AI layer sitting on top of delivery data, that's the question worth asking before the demo:”Where does it run, and can you trace every action back to the record it touched?” An AI that lives natively inside the same Salesforce org as the project, the resource, and the client record can answer that question directly. One that routes through a separate model or a bolted-on integration usually can't, and that gap is becoming harder to paper over now that the buyer's own vendor is naming it out loud.

No One Is Betting on a Single Model Anymore

Salesforce also introduced Koa, its own CRM reasoning model built with NVIDIA, alongside news that Claude is now available inside Agentforce through an open beta. Add that to the Gemini and OpenAI integrations Salesforce has been rolling out over recent months, and the pattern becomes clear: Salesforce isn't picking a favorite AI brand, it's building a platform several models can plug into. The durable asset, in Salesforce's own framing, is the platform and the data model underneath it, not whichever LLM happens to be fashionable this year.

That's a useful correction for PS leaders who've been evaluating vendors on which AI they've bolted on. The model behind an agent will keep changing, probably faster than any PSA roadmap can track. What doesn't change is whether that agent is grounded in your actual project data or working from a stale export, and that's an architecture question, not a model question.

Agents Need an Operating Rhythm, Not a One-Time Rollout

The Agentforce keynote spent less time on new features than on how customers were running the agents they already have, built around getting an agent started, extending what it covers, and improving it through ongoing observation. Salesforce previewed Agent Optimizer, aimed at turning a recurring review into a reusable, traceable improvement, with general availability targeted for next month. The company also said about 30,000 customers are now live on Agentforce, a number that says less about adoption and more about how much production experience is now available to learn from.

That operating rhythm is the same discipline closed-loop delivery intelligence has been arguing for all along. A project health signal, a resourcing gap, a client sentiment shift, none of it is useful as a one-time report. It has to be watched, acted on, and checked against what actually happened, continuously, which is a harder thing to build than a single dashboard and the reason so many PSA tools stop at reporting instead of driving delivery.

What to Take Into 2027 Planning

None of this changes what a PS delivery leader should be doing differently by January, in some ways it actually confirms it. The AI layer worth trusting is the one that never leaves the system where the project data already lives, running on a platform built to plug in whatever model comes next, watched continuously rather than deployed once and left alone. That's the same architecture case Cloud Coach has been making before it was Salesforce's headline story, and last week's keynotes are as good a signal as any that the rest of the market is arriving at the same conclusion.

Sources Cited

Salesforce, "Salesforce Launches AIforce," Dreamforce 2026 keynote coverage, September 2026. https://www.salesforceben.com/complete-roundup-of-the-agentforce-keynote-at-dreamforce-26/

Moor Insights & Strategy, "At Dreamforce 2026, Salesforce Goes All In on Agentic AI," September 2026. https://moorinsightsstrategy.com/field-notes/at-dreamforce-2026-salesforce-goes-all-in-on-agentic-ai/

CRM Hacker, "Dreamforce 2026 Agentforce Recap," September 2026. https://www.crmhacker.com/content/dreamforce-2026-agentforce-keynote-recap