Weekly signal

This week (covering September 21–29, 2026) saw three concrete moves that materially affect customer-service teams building or buying agentic AI: Salesforce tightened how agentic clients are identified and signalled that agent-driven API/MCP interactions will be metered (billing + identity); Zendesk pushed voice into production-grade agent workflows and is expanding specialized industry/custom agents for end-to-end resolution; Intercom (Fin) shipped small but important connector and phone improvements that make agentic resolution more data-connected. These are practical, near-term platform shifts — not abstract research — and they change how you design agentic customer workflows, security boundaries, and cost models.

What changed

1. Salesforce: Agentic Identity + Flex‑Credit metering. Salesforce published guidance and third‑party reporting that registered agent identities (Agentic Identity) will be the canonical identity for AI clients talking to Salesforce, and that successful agent calls over Hosted MCP servers or direct agentic API usage will consume Flex Credits (they will be billed as Headless Platform Interactions). Salesforce says customers will get 30 days' notice before metering begins; admins must register agents and move to Agentic Identity/OAuth flows. This is part of their hosted MCP + Agentforce strategy.

2. Salesforce: enterprise reasoning model (Koa). Salesforce announced Koa, a CRM reasoning model built with NVIDIA Nemotron for Agentforce pilots; it targets multi‑step CRM tasks so agents can act (not just reply). That model rollout signals vendors and integrators will push more autonomous, action‑oriented agents into service workflows.

3. Zendesk: voice AI agents EAP -> broader availability and Specialized Agents. Zendesk opened voice AI agents (EAP details updated late Sept) and continues to expand Specialized (Industry + Custom) Agents that can run end‑to‑end workflows, call external integrations, and create tickets/transcripts on every call. Voice agents include escalation paths and transcription/summary handoffs into Agent Workspace.

4. Intercom (Fin): product updates that matter to agents. Intercom’s changelog shows Sep 21–23 updates: more robust phone diagnostics (outbound call failure reasons), richer reporting, and continued data‑connector improvements (Snowflake MCP connector), which reduce friction for agents that need canonical order/billing/context during automated resolution. These are incremental but operationally important for agent reliability.

What to do with it

• Immediate operational guardrails: audit any agentic workflows that call your CRM or run on Hosted MCP. Require agent registration, test Agentic Identity flows in a sandbox, and set credit‑use alerts.
• Cost modelling: run a small production pilot and measure headless platform interactions per resolved case now — Salesforce will meter those calls. Treat Flex Credits as a second vendor bill (model both LLM provider and Salesforce credit consumption).
• Security & governance: map agent permissions per Agentic Identity; design least‑privilege scopes for agents that act (writes, refunds, tickets). Keep sandboxes for development so you aren’t unexpectedly billed.
• Voice pilot checklist: for contact centers, enroll in Zendesk voice EAP (or early pilots), validate transcript/summarization quality, escalation latency, and PII handling across handoffs. Use Specialized Agents for high‑value industry flows where available.
• Data connectivity: use Intercom/Fin connectors (Snowflake MCP) or Hosted MCP patterns so agents query canonical data instead of copy‑pasting. That reduces incorrect agent actions and rework.

Sources: numbered in order below.