Small business AI quality control
AI Output Acceptance Checklist for Small Businesses
AI drafts can look finished before they are true, policy-safe, or customer-ready. Use this checklist before a team member sends an AI email, publishes a page, updates a CRM note, quotes a customer, stores a procedure, or lets an automation run without another human review.
1. Copy/paste AI output acceptance checklist
Fill one row for each important AI-generated output. If the reviewer cannot verify the source facts, the output is not accepted.
| Output / draft | Where it will be used | Source facts checked | Risk level | Reviewer | Decision | Required edits | Follow-up log |
|---|---|---|---|---|---|---|---|
| [example: customer reply draft] | [email / SMS / CRM / web page] | [order notes, policy page, estimate, source URL] | [low/medium/high] | [name] | [accept / edit / hold / reject] | [specific edits] | [none / add to correction log] |
| [output] | [use] | [source facts] | [level] | [reviewer] | [decision] | [edits] | [log action] |2. Stop rules before accepting an AI output
- STOP AUTOMATION: do not auto-send or auto-publish if the output mentions price, legal rights, warranty, refund, medical/safety advice, employee issues, vendor terms, security, compliance, or customer-specific promises.
- Do not accept unverified output: every claim must match a source record, approved policy, customer note, quote, calendar, or public URL.
- Do not invent missing context: if the AI fills in names, dates, prices, deadlines, service areas, results, or approvals that were not in the source, reject or rewrite the draft.
- Escalate customer-impact changes: anything that changes what a customer expects, pays, receives, or believes needs owner or manager review.
3. Accept / edit / hold / reject decision guide
- Accept: source facts match, tone is appropriate, risk is low, and reviewer can sign off.
- Edit: the draft is useful but needs fact correction, shorter wording, source links, softer claims, or a human-specific detail.
- Hold: the draft depends on a policy, price, owner decision, customer permission, vendor answer, or legal/compliance question.
- Reject: the draft invented facts, overpromised, exposed private data, ignored opt-out/consent rules, or would create a bad customer experience.
4. AI prompt for reviewing an output
You are reviewing an AI-generated output for a small business.
Output to review:
[paste AI draft]
Source facts that are allowed:
[paste source notes, policy, estimate, product details, customer record excerpt, or URL summary]
Review the output and return:
1. Claims that are directly supported by the source facts.
2. Claims that are unsupported or need owner review.
3. Tone or customer-experience risks.
4. Privacy, pricing, refund, warranty, safety, legal, financial, employment, or compliance risks.
5. Decision: accept, edit, hold, or reject.
6. Exact edits needed before use.
Do not invent missing source facts, customer history, prices, deadlines, approvals, guarantees, service areas, legal rights, compliance status, or ROI. If unsure, write "needs owner review".5. Weekly output-quality review
- Move repeated edits into the AI Output Correction Log.
- Update prompts that regularly create unsupported claims.
- Pause workflows where reviewers keep catching high-risk errors.
- Promote only low-risk, repeatable outputs from draft-only to semi-automated workflows.
- Keep a human owner responsible for every customer-facing AI output.
Related Horizon Flow resources
Pair this checklist with the AI Output Correction Log, the AI Workflow Owner Matrix, the AI Tool Usage Policy Template, and the paid Small Business AI Profit Kit.