Automate CSAT Follow-Up Workflows by Score

Workflow area: Automation · Tier: Advanced · Hands-on time: ~45 minutes

Speed is the entire game in service recovery — a detractor contacted within a day can often be turned; one contacted next quarter is gone. Automation buys the speed. The human-approval gate on detractor outreach buys the safety, because the moment you most need a personal touch is exactly when a templated AI reply does the most damage.

When to use — and when not to

• Use it when surveys flow in continuouslypost-ticket CSAT, post-purchase NPS, in-app ratings landing in a sheet or helpdesk daily.
• Use it when someone owns recoverythe workflow drafts detractor outreach; a named person must review, personalize and send it within an SLA. No owner, no workflow.
• Skip auto-sending to detractors, alwaysthank-yous to promoters can auto-send; recovery emails to unhappy customers cannot. Ever. One tone-deaf automated apology screenshot on social media outweighs a year of saved minutes.
• Skip it for scores attached to open complaintsif the ticket is still unresolved, the fix is resolution, not a survey follow-up; filter those out.

The recipe

1. 1.

   Land all responses in one sheet. Whatever the survey tool (helpdesk CSAT, Google Forms, Typeform), get every response into a Sheets row: timestamp, ticket/order ID, customer email, score, verbatim comment, agent/product involved. This sheet is the trigger source and the audit trail.

2. 2.

   Define the branches in writing. For a 1–5 CSAT: 5 = promoter (auto-thank, optionally invite a review), 4 = passive (log only, no email), 1–3 = detractor (AI-drafted recovery, human-approved send, target under 24 hours). Also define stop rules now: never email the same customer twice for one survey, skip customers with an open escalation, and honor unsubscribes.

3. 3.

   Build the Zap with paths.

   • •
     Trigger: new row in the responses sheet.
   • •
     Filter: skip rows already marked processed; skip blank emails.
   • •
     Path A (score 5): send the thank-you from a template — warm, two sentences, no marketing pitch — and mark the row processed.
   • •
     Path B (score 4): update the row as logged; no email.
   • •
     Path C (score 1–3): AI step drafts the recovery email, then the draft goes to the recovery owner — as a Gmail draft, a Slack approval message, or a Zapier-interfaces approval — never directly to the customer. On approval, it sends; the row records approver and send time.
   • •
     All paths: append score, theme (step 5) and status to a tracking tab.
4. 4.

   Write the detractor-draft prompt with guardrails.

   Copy

   Draft a service-recovery email to a customer who rated us [SCORE]/5.
   
   Their comment: "[VERBATIM]"
   Context: ticket [ID] about [TOPIC], handled by our [TEAM] team.
   
   Rules:
   - Open by acknowledging their specific issue in their own terms —
     no generic "we're sorry you had a poor experience".
   - Do not dispute their account, make excuses, or blame any agent.
   - Promise only what is in this list of allowed actions:
     [e.g. "escalate to a senior agent", "review the ticket",
     "call within 2 business days"]. NEVER promise refunds,
     credits, or product changes.
   - No discount offers. No survey links. Under 120 words.
   - Sign off from [RECOVERY OWNER NAME, TITLE] — a person, not
     "the team".
   - If the comment mentions legal action, discrimination, safety,
     or data breach: output only "ESCALATE — DO NOT SEND" plus a
     one-line reason.
   

   The allowed-actions list is the critical guardrail — an AI that improvises "we'll refund you immediately" has just made a commitment your company must honor or retract.

5. 5.

   Add the themes feed. A second AI step (any path, or a nightly batch Zap) classifies each verbatim against a fixed theme list — same discipline as pb-49's taxonomy — and writes the theme to the tracking tab. A scheduled weekly Zap then posts a digest to Slack/Teams: response count, average score, theme counts, and the three worst verbatims quoted exactly. That digest is the bridge into pb-51's full voice-of-customer report.

6. 6.

   Test the ugly cases before go-live. Run through: a 1-score with an empty comment (draft should still acknowledge and offer contact, not hallucinate an issue), a comment naming an agent angrily (no blame in the reply), a "call me or I'm going to consumer court" (must hit ESCALATE), a duplicate row (must not double-email), and a promoter who is actually complaining in the text despite a 5 (add a sentiment check or route mismatches to the human queue). Every one must behave before real customers enter the flow.

Make it yours

• NPS version. 9–10 promoter, 7–8 passive, 0–6 detractor; the promoter thank-you can invite a public review or referral — but only promoters, and only after a compliance check on review-solicitation rules in your market.
• Agent-coaching loop. Pipe score-plus-theme by agent into a monthly summary for team leads — praise patterns first, coaching second, and never auto-share individual customer verbatims as performance evidence without context (pb-22's coaching discipline applies).
• Power Automate version. Forms/Dynamics trigger → condition branches on score → Approvals for detractor sends → Teams digest. Identical architecture, M365-native parts.

Quality check

• No detractor email can reach a customer without a named human's approvalverified by testing, not by intention.
• The draft prompt has an allowed-actions listno invented refunds, credits or promises possible.
• Legal/safety comments hard-stop to escalationthe workflow refuses to draft for them.
• Duplicate and unsubscribe protection testedone survey, at most one email, never to an opted-out address.
• Verbatims flow to themes and a weekly digest someone readsthe loop closes into product and process, not just inboxes.