Salesforce Says 'AI Replaces the UI'—Your Customers Still Have One

On September 15, 2026, Marc Benioff opened the Dreamforce ‘26 keynote with a slide the enterprise software world will replay for years: “AI replaces the UI.” Behind it stood AIforce, Salesforce’s new agentic interface layer, along with Claudeforce (Anthropic’s CEO Dario Amodei on stage), Slackforce, Agentforce Coworker, and Koa—Salesforce’s first CRM reasoning model, built on NVIDIA Nemotron. The pitch is elegant: stop asking your salespeople and support agents to open Salesforce. Let Salesforce come to them, inside Claude, Slack, Microsoft Teams (soon), or wherever they already work.

For internal employees, it’s a genuinely powerful idea. For the customer who just clicked your Google ad, arrived from an AI Overview, or typed your URL into their browser? It changes nothing. They still see a UI—your website—and they still expect an answer in seconds. That gap is what this article is about.

What AIforce actually does

AIforce is a composable interface layer that exposes Salesforce data, workflows, business logic, permissions, and governance to any AI client. It sits on top of the Headless 360 architecture that Salesforce launched in April 2026 under the tagline “No Browser Required”—APIs, Model Context Protocol (MCP) tools, and CLI commands that let external agents read and act inside Salesforce without opening a Salesforce screen.

At Dreamforce ‘26, three pre-configured agentic surfaces launched on top of it:

• Claudeforce: a prebuilt Salesforce MCP server inside Claude, shipping with 37 sales skills. Reps ask Claude to summarize an account, draft an outreach email, or update an opportunity—Claude reasons over Salesforce data and writes back to it, all without switching apps.
• Slackforce: the same idea inside Slack—turn a thread into a case, escalate to an agent, close an opportunity.
• Agentforce Coworker: Salesforce’s own agentic desktop, tuned for cross-cloud work.

Salesforce also announced Koa, its first CRM reasoning model (NVIDIA Nemotron), an expanded Google Cloud partnership (Agentforce plus Gemini Enterprise over MCP), and a broader AWS deal covering Amazon Quick, Bedrock model choice, and Agentforce Voice.

Two populations, two problems

Every business running on AI in 2026 serves two very different populations, and AIforce only touches one.

Population A — Internal employees. Sales reps, support agents, HR partners, ops coordinators. They live inside Salesforce, Slack, Outlook, and now Claude. They have accounts, roles, permissions, and multi-factor authentication. AIforce is built exactly for them: it takes the tool they already use daily and injects Salesforce data into it, so they stop switching apps forty times a day.

Population B — External customers. Prospects who read your blog post from an AI Overview, buyers comparing your pricing page, existing users who need to check refund policy at 11pm. They don’t have a Salesforce account. They don’t have Claude. They don’t have Slack. They arrive on your website—via Google organic, direct traffic, paid ads, email campaigns, or, increasingly, AI-driven referrals. They see a UI, they ask a question, and they judge you by how fast and how accurately you answer.

AIforce does nothing for Population B. That was never its job.

Where trying to conflate them breaks

There’s a natural temptation once you have AIforce running for your reps: “if Claude can read our Salesforce data for internal use, can it also answer customer questions on our website?” That path looks efficient. It’s structurally dangerous once you hit European regulation.

• CLOUD Act exposure. Anthropic is a US company. Every customer question routed through Claudeforce becomes a US-hosted conversation, subject to CLOUD Act warrants regardless of where Salesforce hosts its data. For B2B chat, maybe tolerable. For any conversation that touches account details, health, finance, legal, HR, or public-sector data—not tolerable.
• EU AI Act Article 50, active since 2 August 2026. Any AI system talking to end users in the EU must disclose it’s an AI at the first interaction, and fines run up to €15M or 3% of worldwide turnover. CNIL, Germany’s BfDI, and Spain’s AESIA opened the first inspection wave in September 2026. A cross-vendor chain (Salesforce → AIforce → Claudeforce → Anthropic → customer) makes it very hard to prove who’s responsible for the disclosure and the audit trail.
• DSA and VLOSE. ChatGPT was classified as a Very Large Online Search Engine under the Digital Services Act on August 31, 2026. Claude is on the same trajectory. Routing customer conversations through a designated VLOSE adds a third regulatory layer on top of GDPR and the AI Act.
• Cost, for anyone smaller than the Fortune 500. The AIforce stack assumes Salesforce Data Cloud (roughly $108K/year to start), Service Cloud as a foundation, Claude Enterprise seats (~$60/user/month), and often Slack Enterprise Grid. Total cost of ownership starts in the tens of thousands of euros per year, before the implementation work.
• Attack surface. Multi-agent chains are the frontier of prompt-injection research in 2026. An attacker who plants a poisoned document in one system can pivot through connectors to exfiltrate deals, contacts, or opportunities. Bounded RAG—retrieval limited to a small, moderated corpus—is a much smaller target than a headless CRM plumbed into three LLM vendors.

None of this means AIforce is a bad product for its intended job. It just isn’t a customer-facing chatbot, and it doesn’t want to be.

What a customer-facing chatbot has to do

A chatbot on your public site has a completely different job description than an internal agent:

1. Answer only from your published knowledge. Product pages, PDF spec sheets, FAQs, terms and conditions, blog articles. Not internal opportunities, not deal notes, not employee data. Bounded scope is a feature, not a limitation.
2. Load in under a second on any device. No login, no plugin, no browser extension. A JavaScript widget on the page—one line of code.
3. Comply natively with Article 50. The AI disclosure must be visible in the first message, not buried in a footer.
4. Capture leads with GDPR consent. When a visitor is qualified, the chatbot should collect an email or phone with explicit consent, not silently forward the conversation into a CRM they didn’t ask about.
5. Be auditable end to end. One vendor, one jurisdiction, one log.
6. Cost predictably. No per-resolution fee, no per-seat surprise. Growth in traffic shouldn’t become a P&L crisis.

Salesforce’s own customer-facing product for this job is Einstein Service Agent, and it lives inside the Salesforce ecosystem—assuming Service Cloud, Data Cloud, and the CRM footprint that goes with them. That fits large enterprises that already run on Salesforce. It doesn’t fit the roughly 82% of French SMBs now investing in AI who need something bounded, sovereign, and priced for a monthly budget, not an annual procurement cycle.

DoxyChat: the orthogonal tool

DoxyChat is deliberately built for Population B. It doesn’t compete with AIforce and it doesn’t try to replace your CRM—it does what AIforce doesn’t.

• Bounded RAG on your published documents. Ingest PDFs, DOCX, Excel, your website, an RSS feed. The chatbot answers only from what you loaded, with citations. Nothing else.
• Deployed in two minutes. One JavaScript snippet on your site—less than a hundred and fifty kilobytes—and the widget is live. No IT project, no MCP server, no browser extension.
• Hosted in France on Scaleway, Mistral as the primary LLM (Apache 2.0 open weights), Gemini fallback. No CLOUD Act exposure. No US-mandated warrants over EU customer conversations.
• Article 50 disclosure built in. The AI declaration appears in the opening message, screen-reader friendly, in the customer’s language.
• Row-level security in PostgreSQL. Every chatbot is isolated per tenant. Every message is logged for audit.
• Lead capture with native GDPR consent. When a visitor is ready, DoxyChat asks for an email with explicit consent, cleanly, in the conversation—no dark patterns.
• Discovery plan at zero euros. One chatbot, ten documents, two hundred queries per month. Enough to prove the point before your first invoice, and enough for many small sites to run indefinitely.

If you already run Salesforce and you’re rolling out AIforce, keep going—it’s excellent for your internal workflows. Bolt DoxyChat onto your public website in parallel, and treat it as a separate concern for a separate audience.

If you don’t run Salesforce and never plan to, AIforce isn’t the tool your customers ever needed. DoxyChat is.

Conclusion

The Dreamforce ‘26 keynote framed AIforce as the end of the browser-and-click enterprise. That framing is accurate—for the people inside your company. Your customers, on the other hand, still open a browser, still land on a page, and still expect a helpful answer immediately. That interaction is the front door of your business, and it deserves its own AI, hosted in your jurisdiction, bounded to your own knowledge, and audited under a single vendor.

Try DoxyChat free—one chatbot, ten documents, no credit card—at doxychat.com.