Review request AI agent for reputation workflows

Install stack

• Trigger only after completed appointments, delivered orders, or closed support cases.
• Filter out refunds, disputes, cancellations, and unresolved complaints before requesting reviews.
• Use OpenClaw for orchestration with cloud routing through OpenRouter or local routing through Ollama.
• Run the gateway on a dedicated VPS, Mac mini, or locked-down local machine with restart monitoring.

Workflow

• Capture the inbound request for review generation follow-up with source, owner, urgency, and missing fields.
• Log who was asked, when, on which channel, and whether they replied.
• Draft or execute the next step only inside approved permissions and rate limits.
• Write the result back to the system of record and send a short operator summary.

Implementation checklist

• Send negative replies to a human recovery workflow instead of a public link.
• Create allowlisted actions, forbidden actions, and escalation phrases.
• Test the agent with real-looking but non-sensitive samples before live credentials are added.
• Record a handoff Loom covering restart, credential rotation, logs, and rollback.

Integrations

• Use channel-specific timing for SMS, email, or WhatsApp follow-up.
• Email, calendar, CRM, or spreadsheet system where the work is recorded.
• Logging destination for transcripts, tool calls, failed jobs, and handoff notes.

Decision criteria

• The workflow repeats often enough that local businesses with repeat service can measure time saved or revenue protected.
• The tools have stable APIs, inbox rules, exports, or admin access.
• A human can define what good, bad, and uncertain outputs look like.

Risks to handle before launch

• The agent can create business risk if it acts without approval on payments, legal commitments, or customer promises.
• Messy source data can cause confident but wrong updates unless the workflow includes verification steps.
• Channel outages, expired tokens, and model latency need a manual fallback path.

Security notes

• Use least-privilege API keys and separate test credentials from live credentials.
• Keep memory, logs, and uploaded files out of public folders and shared drives.
• Rotate credentials after handoff and disable installer access unless ongoing support is contracted.

Acceptance tests

• The agent completes a full review generation follow-up test from trigger to logged outcome.
• A low-confidence or risky request is escalated instead of executed.
• Restarting the gateway does not lose memory, credentials, routing, or scheduled work.