← All articles

A coaching tool that never reads your customers' chats — how that works

Support Coach AI · September 2026

An AI tool can coach a support agent's replies without ever reading the customer's side of the conversation — by analyzing only the reply the agent is drafting. That sounds like a limitation. For the person responsible for what leaves the company's browsers, it's the whole point: there is no customer conversation data to protect, because none is ever collected.

The design choice: draft-only analysis

Most of what makes a support reply risky is visible in the reply itself. Blame language ("that's your mistake"), missing ownership, a refusal with no alternative, an abrupt close — none of these require reading the customer's messages to detect. So a coaching tool has a real choice: ingest the whole conversation for marginally more context, or read only the agent's draft and keep every customer word inside the browser.

Support Coach chose draft-only, and it's a hard boundary, not a setting:

• What leaves the browser when a check runs: the agent's draft text, which local rules it triggered, a local score, and a content hash. That's the whole payload.
• What never leaves: the customer's messages, names in the thread, the conversation history. The AI cannot see them — so no policy, employee, or breach can expose what was never sent.
• What's stored afterwards: no draft, anywhere. Drafts are processed in flight and discarded; the statistics kept are counts, not content.
• What's embedded in the extension: no third-party analytics SDK, no trackers. It talks only to its own API.

The full version of this — payload fields, subprocessors, retention, and what we don't have (no SOC 2 certification yet, no SSO/SAML, no data-region choice — stated plainly) — is public at supportcoach.io/security, written for IT reviewers.

Why this matters more for chat than anywhere else

Live-chat transcripts are dense with personal data — names, order numbers, addresses, payment fragments, health details, whatever the customer typed in frustration at 11pm. A tool that ingests conversations inherits all of it, and your obligations around it. A tool that reads only what your own employee is drafting inherits almost none — which is why the IT review for draft-only coaching is short, and why it can be deployed to a team without a data-processing negotiation over customer content.

The coach sees the draft. The conversation around it stays in the browser.

What to ask any vendor in this category

Whoever you evaluate — us included — these five questions separate marketing from architecture:

1. Does the tool read the customer's messages, or only the agent's draft?
2. Exactly which fields leave the browser? Ask for the list, not a summary.
3. Is any message content stored after processing? For how long?
4. Does the extension embed any third-party analytics or trackers?
5. Is all of this written down on a public page IT can review — including what the vendor doesn't have yet?

Support Coach's answers: draft only; four fields; nothing stored; no trackers; supportcoach.io/security. $15 per agent per month, 14-day free trial: supportcoach.io/extension.

FAQ

If it can't see the conversation, how good can the coaching be? Good enough that the detection works where it matters: blame, missing ownership, dead ends and cold tone are all visible in the draft itself. What draft-only analysis gives up in context, it returns in deployability — and it will never quietly start "needing" your conversation data.

Can IT deploy it centrally instead of per-agent installs? Yes — the security page includes force-install instructions for managed Chrome, so IT can push it fleet-wide with the same policy tooling as any extension.

Does "no SOC 2" mean it isn't safe? It means we're early and say so. The compensating control is architectural: no customer conversation content is stored anywhere because none is ever sent. Ask the five questions above of anyone, certified or not.