At Dreamforce this month, Marc Benioff unveiled AIforce — a new live interface layer that sits above Agentforce, Data 360 and Customer 360. The message was blunt: customers won’t need to log into Salesforce anymore. AI agents will act on the platform for them.

Salesforce also announced Koa, its first CRM reasoning model for Agentforce, built with NVIDIA. The company says Koa already matches or exceeds leading models on CRM actions with three times fewer errors.

That’s impressive. But read it carefully, because it hides the real story.

The model is not the bottleneck. Your data is.

An agent that reasons brilliantly over opportunity stages your reps each define differently will simply get the wrong answer faster. An AI that logs every call, email and meeting — like HubSpot’s new self-updating Smart CRM — still can’t tell you whether that call actually moved the deal forward.

Gartner’s prediction puts a number on it: through 2026, organizations will abandon 60% of AI projects that lack AI-ready data. McKinsey’s State of AI survey tells the same story from the other side — 62% of organizations are experimenting with AI agents, yet fewer than 10% have scaled them in any single business function.

The blocker isn’t the model. It’s the data underneath it.

What “AI-ready CRM data” actually means

Before any agent — Koa, Breeze, or a custom build — can act reliably on your CRM, four things have to be true:

  • One shared definition of your pipeline. What “qualified” and “committed” mean must be agreed, configured, and enforced — not left to each rep’s interpretation.
  • Clean master data. Duplicate accounts, stale contacts and half-migrated fields poison every downstream AI answer.
  • Governed integrations. Your CRM, ERP and marketing tools each claim to be the source of truth. Only one can be. The rest must sync on clear rules.
  • Scoped permissions and audit logs. Agents that can write to your CRM need guardrails: approval steps for money-moving actions, least-privilege access, and a record of every change.

None of this is model work. All of it is implementation work — and no vendor ships it in a box.

This is where most companies get stuck

The platforms are racing to sell you the agent layer. What’s missing is the unglamorous layer beneath it: data cleanup, ETL pipelines, field governance, integration architecture and RevOps analytics that tell you whether the AI is actually producing revenue.

That’s exactly the work we do at Falafel Technologies — HubSpot and Salesforce implementation, data engineering and ETL, custom AI agents with proper guardrails, and analytics that connect AI activity to collected cash, not vanity metrics.

Planning to pilot agentic AI on your CRM in 2026? Start with a data-readiness audit, not a model demo. See how we can help →