Giving away the interface is not the same as giving away control

On 15 September 2026, Salesforce opened Dreamforce by unveiling AIforce, an interface layer built so work happens inside Claude, Slack or Lightning rather than Salesforce's own screens, according to Salesforce Ben's coverage of the keynote. The interface is the part Salesforce is willing to give away. What it announced keeping, on the same stage, is the part worth reading closely.

We read the full announcement set, not just the headline, because a vendor conceding the interface is a bigger claim than most of this week's commentary is treating it as. On Salesforce's own account, it is also a more selective concession than "AI replaces the UI" makes it sound.

What AIforce actually bundles

AIforce is not one new product. It is an interface layer placed over three components Salesforce announced separately across 2026, now sold as a single layer.

AIforce brings together Agentforce Coworker, announced in May 2026, Slackforce, announced during Salesforce's second quarter earnings call in August 2026, and Claudeforce, the Salesforce and Anthropic partnership that moved into open beta on the day of the keynote, 15 September 2026 (Salesforce Ben, 15 September 2026; Salesforce Break, 17 September 2026).

The pitch, in Salesforce Ben's framing of the keynote, is that people "don't have to come to Salesforce to get work done" because Salesforce now meets them inside Claude, Slack or Lightning instead. Every one of those three surfaces already existed before 15 September. What changed that day is that Salesforce started marketing them as one layer rather than three separate integrations.

What Salesforce kept: Koa

The same keynote introduced Koa, a CRM reasoning model Salesforce trained itself and runs only inside its own infrastructure, never inside Claude or Slack.

Salesforce's own announcement describes Koa as its first CRM reasoning model, built by post-training Nvidia's Nemotron 3 Super on a proprietary synthetic dataset modelled on nearly three decades of CRM deployments (Salesforce, 15 September 2026). Salesforce controls the model weights and says it handles both post-training and inference inside its own trust boundary, and that no customer data was used to train it.

Salesforce also says, in the same release, that Koa matches or exceeds leading models on its own CRM benchmark, with three times fewer errors on tasks such as case routing and opportunity updates. That is Salesforce's benchmark, run on Salesforce's tasks, and we have no independent measurement of it, so we are treating it here as a vendor claim rather than a verified result.

Koa is available now to select Agentforce pilot customers, with general availability expected during the winter of 2026 in the US (Salesforce, 15 September 2026).

Why "who owns the interface" is the wrong question for a buyer

Most reaction to the launch asked what Salesforce becomes if Claude is the interface. That question skips over where the judgement actually runs.

An interface is where a person clicks. A reasoning model is what decides which record to update, which lead to flag or which value to write, before the click ever happens. Salesforce gave up ground on the first and, on its own account, tightened its hold on the second, in the same announcement.

That split matters more than the headline suggests, because the reasoning layer is what actually produces the write. We have written before about why a permission check on a write is not the same as a record of why the write happened, and about what breaks in a CRM once more than one agent can write to the same field. Both problems get harder, not easier, when the reasoning that produces the write sits inside one vendor's trust boundary while the interface that surfaces it sits inside three others.

None of this makes Salesforce's approach wrong. Keeping training data and inference inside a controlled boundary is a defensible answer to the governance question our own control room argument raises, and it is a more concrete answer than most vendors in this category have given. What is worth watching is what it costs a buyer once the interface and the judgement behind it live inside different companies' infrastructure.

What to check before buying an "AI replaces the interface" pitch

Ask where the reasoning model runs and who can see its training data and outputs, not how many surfaces the interface now reaches.

Three questions we put to any vendor pitching this kind of launch on client engagements:

• Where does the reasoning model run, inside the vendor's own infrastructure or inside a third party's, and does that change under a different contract or tier.
• What happens to the audit trail behind a write when the interface that triggered it changes from one surface to another.
• Whose benchmark is being quoted for accuracy or error rate, and has anyone outside that vendor measured it independently.

We set out the general version of this test in what you should never hand an agent: judge a task by its reversibility and its blast radius, not by which surface it happens to run on. An interface layer changes where a request enters a system. It does not change what happens if the judgement behind the answer turns out to be wrong.

AIforce is a genuinely well executed piece of interface engineering, and Salesforce deserves credit for shipping Claudeforce, Slackforce and Agentforce Coworker as one coherent layer rather than three disconnected betas. Our point is narrower. Judge the launch by what moved, the interface, and by what stayed exactly where Salesforce wants it to stay: the reasoning behind every write.