Support's AI agents shift from runtime to authoring — copilots draft and test them, billing meters them.

The week in customer support

Nineteen products shipped this week, producing nine sparks against fifty-nine improvements — a ratio that tells the sector's story. The dominant move is not a new feature but a change in who, or what, does the authoring. Support vendors are turning the AI agent from a runtime you configure into something a copilot drafts, tests, and meters. Respond.io put Copilot into every workspace as an agent builder that writes an AI Agent from a description and runs test cases before publishing, then paired it with AI Credit On-Demand, which bills overage instead of halting automation mid-conversation. The two ship together deliberately: one lowers the cost of creating agents, the other turns the agents into metered consumption revenue.

The second pattern is capture and retrieval leaving the product entirely. Canny replaced its Slack slash commands with an @Canny mention that reads plain language from any thread, and Frill shipped an MCP server — ten tools, one key — on every plan, so its feedback board answers to Claude, ChatGPT, or Cursor. Both bet that the value is the corpus, not the interface. Against that, Twilio is doing the unglamorous platform work the layer requires: retiring Functions Classic with an automatic September-to-October migration while turning consent and deliverability into API surface.

Leaders

Respond.io shipped the week's clearest strategic pairing. Copilot lands free and credit-free as an agent-authoring assistant, while AI Credit On-Demand quietly reclassifies AI from a bundled plan feature into metered overage. Widen the funnel to build agents, charge for the agents that run — the commercial and product halves of one move, alongside Microsoft 365 arriving as a full email channel.

Canny turned its Slack integration from a command surface into a conversational one. @Canny captures feedback from any thread, routes it through Autopilot for deduping, and creates ideas, users, and companies inline — the first agent Canny operates itself rather than exposing over MCP. It sits on a reporting rebuild that ties feedback to renewal dates and pipeline revenue, the other half of a capture-everything strategy.

Twilio, the sector's highest-velocity product, is collapsing parallel generations of the same product onto one platform. Functions Classic is being retired on Twilio's schedule with no customer action, and the same consolidation shows in dedicated IPs for Twilio Email and a Consent Management API that can now clear network-level toll-free opt-outs. Compliance is becoming something you configure programmatically rather than file a ticket about.

Frill moved its MCP server from June's public beta to a full release on every plan — a pricing choice that optimizes for the feedback corpus being queried rather than for upsell. With a Chrome extension capturing ideas from anywhere and AI summaries collapsing long threads, Frill is systematically removing the requirement to be in Frill.

Hatz AI shipped Activate, a rollout console built not for end users but for the MSPs who resell it: a shadow-AI scan, regenerated enablement material, and a sales-user role. Against a window otherwise made of model-selector maintenance, it is the one release aimed at the partner rather than the operator, and it fits a phone agent being hardened for repeated multi-tenant deployment.

Wildcards

Richpanel added Telebroad as a third phone provider, but the notable part is how voice arrives: not as audio to replay but as a recording, a labeled transcript, and an AI summary an agent reads during a handoff. Voice is being normalized into the text timeline rather than treated as a separate medium.

Live Helper Chat is the sector's counter-current. While everyone else adds AI agents, this self-hosted product spent the window on operator oversight — configurable offline reasons backed by a new permission, a forms-module audit trail, and per-operator performance widgets. Nothing in its releases points toward AI or automation; it is being fitted out for teams large enough to need supervision.

Themes that compounded

• Agent authoring is the new battleground: Respond.io's Copilot and Hatz AI's phone-agent tooling both invest in building and validating agents, not running them.
• MCP is becoming table stakes: Frill put it on every plan while Formbricks spent its 5.3 line correcting MCP scopes rather than extending them.
• Capture is migrating out of the product, from Canny's Slack mention to Frill's browser extension.
• In-house models and governance are consolidating trust: Deskpro resolves datasource access server-side for regulated buyers, and ManageEngine ServiceDesk Plus moved its AI onto Zoho's self-hosted Zia.
• Metered AI is the emerging commercial model, most explicitly in Respond.io's consumption billing.

Watch this week

Expect the authoring-and-metering pattern to spread: where Respond.io leads with test-case validation before publish, others building agents will need the same pre-flight tooling. MCP write paths are the next frontier — Frill and Canny can already be read by assistants, and filing, merging, and status changes are the obvious follow-on. Watch HelpCenter.io, whose FAQs.me free tool and Zendesk-migration content form a funnel aimed at incumbent switchers, and Twilio's September compliance cutovers, where new TFV and A2P error codes should land as API-readable signals.