Dreamforce 2026 Fixed the Model Side. The Data Side Is Still Yours

The three companies on stage were each fixing a different problem

Jensen Huang described AI as a layer that sits on top of software rather than replacing it. Dario Amodei showed Claude working with Salesforce data through the Claudeforce plugin. Marc Benioff introduced Koa, which is Salesforce's own reasoning model for CRM tasks, trained on synthetic data modelled on 27 years of deployments and, as Salesforce stresses, not on a single byte of customer data.

Put those together and the picture is this. NVIDIA supplies the compute and the Nemotron model Koa is built on. Anthropic supplies the interface people will use to ask questions. Salesforce supplies a model that understands the structure of a deal and makes three times fewer errors on CRM tasks than general models do, by its own benchmark.

All three are improvements to the model side, and the records in your instance are the same as they were before the keynote.

Why Siemens got the most useful line of the keynote

Siemens was on stage as the customer example, and the most useful line of the keynote came from their segment.

"Hallucination does not really work on the shop floor."

They can run Agentforce agents across sales and service because their Salesforce instance is already the single source of truth for products, parts and manufacturing instructions. When an agent answers a question, the answer is made from records that are correct, because a decade of work went into making them correct.

That's the condition the demos don't mention. A CRM model with fewer errors still builds its answer from what's in the fields. If the close date on an opportunity has been pushed every month this year, Koa will reason about it correctly and Claude will report it fluently, and the answer will still be wrong, because the record is.

The four questions Claudeforce will ask of your data first

When a company connects Claude to Salesforce, the first questions people ask are predictable, because they're the questions they've wanted answered for years. Each one depends on a field that tends to be in poor shape in an instance that's a few years old.

• Which deals are we most likely to lose this quarter? Depends on close dates, stage and next step being current. A pipeline where half the opportunities have no next step gives you a confident list built on nothing.
• What do our biggest customers have open with us? Depends on accounts not being duplicated. Three records for the same customer means the answer is a third of the truth.
• Where are our service hotspots? Depends on cases being categorized consistently. Ten spellings of the same product give you ten small problems instead of one big one.
• What's the health of this account before my meeting? Depends on activity being logged. Reps who keep the real notes in their phone leave Claude preparing you from the last email that got synced.

Each of these is a process problem, and Koa can't fix it. The fix is a decision about who owns each field and what happens when it's empty.

What I'd do before the pilot, in this order

• Pick the one question your sales leader would ask first. Pull the report that answers it today and look at how many of the underlying records are complete. That percentage is your readiness number.
• Fix ownership before you fix data. A cleanup without a rule for who keeps the field current doesn't stay clean.
• Merge duplicate accounts before connecting anything, because every question Claude answers about a customer assumes there's one of them.
• Run the permissions review. Claude sees what each person can see, and in an older instance that's more than the job requires.
• Turn it on for one team, read-only, and watch what they ask. The questions they ask in week one are your cleanup list for month two.

Amodei put it plainly at Dreamforce 2026: the value being captured from this technology today is five or ten percent of what's possible. I'd agree, and after this week I'd say the models are no longer what's holding the number down. The models got better in San Francisco. Your data is the remaining variable, and that's the work my team at Apexify Consulting does for corporate Salesforce customers. If you'd like to know how far your instance is from what was on stage, you can find us at apexifyconsulting.com.