From copilots to operators
For most of CRM's history, "AI" meant a sidebar that summarised an email or drafted a reply. The human still clicked, typed, confirmed. The AI was a copilot.
That ended in 2025. Anthropic's Model Context Protocol (MCP) made it standard for any AI agent to call external tools through a typed contract. Vercel v0 ships with native MCP support. Cursor reads mcp.json. Claude Desktop has Connectors. ChatGPT shipped Apps SDK.
The implication for CRM is direct: agents can now operate your customer data, not just describe it.
What does "agentic" mean concretely?
An agentic CRM exposes its functionality as typed tools that any MCP-compatible agent can call. Every tool carries:
- A name (
create_booking,attach_tags,void_invoice) - An input schema (Zod / JSON Schema)
- An output schema (the shape of what comes back)
- Annotations (
readOnlyHint,destructiveHint,openWorldHint,idempotentHint)
The annotations are the critical bit. They tell the agent which tools mutate data, which hit external services, which are safe to retry. A well-annotated catalog lets the agent self-gate destructive operations — it knows to ask for confirmation before calling cancel_booking, but not before list_members.
Three things change
1. Onboarding. With an agentic CRM, "sign me up" is a tool the agent can call. FavCRM's register_organisation_request → register_organisation_verify pair is OTP-gated and rate-limited. The user never leaves their chat. No portal click. No copy-paste API key.
2. The shape of work. Pre-agent CRM workflows were "open the screen, find the customer, click the action, fill the form." Agentic workflows are "tell the agent what you want." The agent picks the right tool sequence — search_members → get_loyalty_balance → earn_loyalty_points → confirm — and runs it.
3. Skill packaging. MCP has a separate primitive called prompts (named, multi-step playbooks). FavCRM ships 37 of these covering customer onboarding, booking lifecycle, billing, subscriptions, AI training, reporting. Each playbook is a slash command the agent already knows how to run.
What does this mean for service businesses?
Service businesses — beauty, fitness, tutoring, retail, hospitality — have been underserved by enterprise CRM for decades. Salesforce and HubSpot price per seat from $50+ and target B2B sales pipelines. Mindbody and Calendly own the booking layer but are weak on customer relationship and comms.
Agentic CRM rewires this. The cost of "having a CRM that does X" drops from "subscribe to a vendor" to "ask the agent." Five examples that work today:
| Ask | Tools the agent chains |
|---|---|
| "Find customers who haven't booked in 30 days, tag them 'win-back', send a WhatsApp." | search_members → attach_tags → send_whatsapp_message |
| "Bobby wants a yoga class Friday 10am. Book it." | list_services → get_available_slots → create_booking → confirm_booking |
| "Bill Acme Co for the November engagement, $2,400, due in 14 days." | create_invoice → mark_invoice_sent |
| "Show me past-due subscriptions and reach out to each." | list_subscriptions (status=past_due) → send_whatsapp_message per row |
| "Train the agent on our new cancellation policy." | add_knowledge_text |
None of these involves opening a screen.
Why MCP is the standard to bet on
Three reasons.
1. Cross-vendor adoption. Anthropic, OpenAI, and Vercel all ship MCP. Cursor reads mcp.json. Smithery, Continue.dev, Windsurf, Zed, Cline, Roo Code, Replit Agent all speak the same protocol. One server reaches every client.
2. Typed contracts. Function-calling without schemas turns into spaghetti — the agent guesses at parameter names, the server returns ad-hoc shapes. MCP's input + output schemas let the agent reason about results without parsing prose.
3. Annotation discipline. The destructive/readOnly/openWorld hints stop the worst class of agent failures (catastrophic mutations on confused intent). Pre-MCP function-calling had no equivalent.
What to evaluate
If you're picking an agentic CRM today, three questions matter more than feature lists:
How many tools, and what's the annotation coverage? A 154-tool catalog with 100% annotation coverage beats a 30-tool catalog without. Tools that lack
destructiveHintare landmines for agent users.Are there named playbooks? Multi-step workflows hidden behind a slash command beat raw tool exposure. Counting tools without counting playbooks is like counting bricks without counting buildings.
What's the signup story? If the user has to leave the chat to create an account, the agentic story breaks. OTP-gated agentic signup is the cheapest way to validate this.
What's next
The protocol is settling. The clients are shipping. The vendors who don't expose typed agentic surfaces in 2026 will look the way mobile-only-after-2010 vendors looked.
For service businesses specifically: the question isn't "should we adopt an agentic CRM" — it's "which one, and how soon."
Browse FavCRM's MCP catalog to see what 154 typed tools and 37 named playbooks look like in practice. The same workspace runs across Cursor, Claude, ChatGPT, and Vercel v0.
