Salesforce Named Its Chatbot Casey: Should Yours Have a Name Too?

On September 11, 2026, days before Dreamforce 2026 opened in San Francisco, Salesforce unveiled seven named AI agents with job titles and long-term memory: Casey for customer service, Paige for IT and HR support, Carter for shopping assistance, Hunter for outbound sales, Marshall for supply chain, Piper for inbound lead qualification, and Fin as an additional customer-service agent. Five went generally available immediately. Hunter, powered by a new long-horizon runtime designed to pursue goals over weeks instead of a single chat session, entered pilot with general availability planned for November 2026.

The messaging is deliberate: Salesforce is no longer selling “chatbots.” It is selling teammates with names, faces, and memory that sit on Customer 360, follow your business rules, and act autonomously. That is a significant enterprise story. It is also a story that has been oversimplified into “chatbots are dead, name your AI agent” — and that oversimplification is dangerous for most businesses.

Here is what actually happened, what it costs, and why for the vast majority of small and mid-sized businesses, a bounded RAG chatbot — not a named autonomous agent — is still the right first purchase in 2026.

What Salesforce Actually Launched

Casey, Paige, Carter, Marshall and Piper are generally available now. Hunter enters pilot with general availability planned for November 2026 and is the first to use the long-horizon runtime, which lets an agent pursue an outbound sales goal over multi-week timelines rather than a single conversation. Fin, a customer-service agent, was folded into the announcement alongside the six new hires. Salesforce also released the Trusted Enterprise AI Harness, a governance layer for companies already running multiple AI agent platforms.

Under the hood, these agents sit on Salesforce’s Customer 360 data platform. They operate within your existing business rules, permissions, and security setup. They call tools, chain actions, and hand off to humans through Service Cloud. RAG makes them context-aware by retrieving from trusted internal sources, so they are grounded — but they are not bounded to a single document set. They can access accounts, opportunities, cases, orders. That is the whole point.

This is not a bad product. For a Fortune 500 running Salesforce top to bottom, it is a coherent next step. For a 20-person real-estate agency in Lyon or an accounting firm in Nantes, it is a category error.

The Sticker Price Is Not the Real Price

Salesforce Agentforce today has three pricing paths:

• $2 per conversation (introduced October 2024), where a conversation runs from the agent’s first response until the case closes or 24 hours of inactivity elapse.
• Flex Credits at $500 per 100,000 credits — one standard action costs 20 credits ($0.10), one voice action costs 30 credits ($0.15).
• Per-user licenses starting at $125/user/month.

Independent 2026 analysis by pricing consultancies has flagged what Salesforce does not put on the pricing page. Agentforce requires an existing Service Cloud foundation. It requires a mandatory Data Cloud subscription starting at $108,000/year. It requires knowledge-base setup, custom development, and — per multiple 2026 field reports — a typical implementation timeline of five to eleven months. Before Casey answers a single ticket, the total cost of ownership for a mid-market rollout is often past six figures per year and a full quarter of engineering runway.

That model works for enterprises with a Salesforce admin team, a data-cloud budget line, and a multi-year AI transformation program. It does not work for a business that just wants a chatbot on its website that answers questions from a PDF product guide, captures leads at 2 a.m., and does not hallucinate.

Named Autonomy Is Not the Only Way to Do AI

The dominant 2026 narrative — pushed by Salesforce, OpenAI (GPT-6 Astra, September 3), and xAI (Grok Bot for Enterprise, September 3) — is that chatbots are dead and the future is autonomous agents with memory, tools, and multi-step goals.

Half of that narrative is correct. Half of it is misleading.

The correct half: for internal productivity — a rep drafting a pitch deck, an IT tech provisioning access, a finance analyst reconciling ledgers — an autonomous agent that can traverse ten systems and coordinate weeks-long workflows is a genuine step-change. Cisco rolled out MyAgent to 90,000 employees this summer for exactly this reason. Grok Bot for Enterprise, Le Chat with 20 connectors, ChatGPT Work and Claude Cowork all serve this Job A: internal productivity.

The misleading half: the AI that talks to your customers on your website does not have the same job. It is not asked to reconcile ledgers or schedule multi-week outbound campaigns. It is asked to answer questions truthfully from your documented knowledge, capture qualified leads, and defer to a human when it is unsure. That is Job B: customer-facing RAG. Job A and Job B look similar from a distance and are governed by completely different economics, risks, and regulations.

McKinsey’s State of AI 2026, published August 25, made this explicit. Chatbots are the single most-scaled AI tool inside enterprises (47% of respondents), but only 37% of organizations report EBIT impact from AI — flat versus 2025. Deployment is not the problem. Fit-for-purpose deployment is. Named autonomous agents win productivity headlines; bounded customer-facing RAG wins the P&L.

Where Named Agents Break for a Non-Enterprise

Three specific gaps make named Salesforce-tied agents the wrong first purchase for a small or mid-sized business.

1. Data residency and the CLOUD Act. Salesforce is a US-headquartered company. EU hyperscaler regions exist, but the parent-company governance is intact: the US CLOUD Act (2018) can compel US providers to disclose data held anywhere in the world. For an EU business under GDPR — and for any regulated sector, from financial services to legal to healthcare — a chatbot that ingests customer conversations under US governance is a structural risk, not a configuration one. DoxyChat runs on Scaleway in France, on Mistral open-weight models under Apache 2.0, with no CLOUD Act exposure by architecture.

2. Attack surface and EU AI Act Article 50. Since August 2, 2026, EU AI Act Article 50 has been enforced across all 27 member states. Any chatbot must identify itself as AI at first interaction, and operators must be able to produce an audit trail on demand. Fines reach €15M or 3% of global revenue. A bounded RAG chatbot that only reads from a controlled document set is straightforward to audit: every answer traces to a citation, every citation to an ingestion event. A named autonomous agent that reads from Customer 360, calls tools, chains actions, and pursues goals over weeks has an audit surface that is orders of magnitude larger — and a hallucination surface that is proportionally larger too. Bounded is auditable.

3. Cost of failure and lock-in. A €19-per-month chatbot that misfires costs you a bad answer and a lesson learned. A $125-per-user-per-month agent bundle that misfires — after nine months of implementation, $108K of Data Cloud, and Customer 360 dependencies you cannot unwind — costs you a strategy. In September 2026, OpenAI announced it would stop supplying models to Cursor on November 12, 2026, a reminder that the LLM supply chain can break at any time. A chatbot architecture that abstracts the model (DoxyChat uses Mistral in production with a Gemini fallback) survives supplier upheaval. A named agent stack fused to a single CRM does not.

What DoxyChat Actually Is

DoxyChat is not an autonomous agent with a name. It is a bounded RAG chatbot for customer-facing conversations. You upload your product PDF, your terms and conditions, your FAQ, your policy document, your knowledge base, or point it at your website. DoxyChat ingests, chunks, embeds, indexes. When a visitor asks a question, DoxyChat retrieves the relevant passages, generates a grounded answer, cites the source, captures the lead when appropriate, and hands off to a human when it does not know.

It does not access your CRM. It does not call arbitrary tools. It does not pursue multi-week goals autonomously. It does exactly what a customer-facing chatbot should do in 2026: be present, be accurate, be auditable, and be contained.

Because it is contained:

• The answer set is bounded to what you gave it — no wild hallucinations, no unpredictable actions.
• The audit trail is trivial to produce for GDPR and EU AI Act inspections.
• Deployment is minutes, not months. The widget is one line of JavaScript.
• Pricing is predictable: a free Discovery plan for a first chatbot on ten documents, then transparent plans up to standard business tiers. No per-conversation surprises, no mandatory Data Cloud floor, no six-month implementation.
• Hosting is Scaleway France. The LLM is Mistral (open-weight, Apache 2.0). No CLOUD Act exposure. Article 50 native.

The 2026 Decision Framework

If your business is a Fortune 500 running Salesforce as its system of record, staffing an AI center of excellence, and buying a workforce of AI agents to sit next to your reps — Casey, Paige, Hunter and their colleagues are the coherent choice. Salesforce built them for you.

If your business is a small or mid-sized company that needs the chatbot on its website to answer customer questions accurately, capture leads, and stay out of trouble with the CNIL — you do not need a named agent. You need a bounded RAG chatbot hosted in your jurisdiction, on models you can inspect, with a price you can predict, and a deploy time that fits between two coffees.

The industry is naming its agents this week. That is fine. It does not mean your chatbot needs a name. It means you should be clearer than ever about which job you are hiring for — customer-facing containment, or internal-productivity autonomy — and buy the tool that matches, not the one with the loudest press release.

Try the Contained Alternative

DoxyChat gives you a bounded RAG chatbot on your documents, hosted in France, GDPR- and Article 50-compliant, deployed in two minutes with one line of JavaScript. The Discovery plan is free forever — one chatbot, ten documents, 200 queries per month — enough to prove the model on your own content before you commit to anything. Try DoxyChat free at www.doxychat.com and see what a chatbot looks like when it is designed for the job it actually has to do.