Which workflows can you
Really automate?
OpenClaw workflows combine Gateway, agents, mailbox and browser tools with approvals — not “fully autonomous business operations,” but operational AI for triage, monitoring, CRM steps, and repetitive browser checks in SME processes.
direct answer
What can OpenClaw do in the
Automate practice?
Short answer: Repeatable steps with clear input: classifying mail, summarizing logs/health, populating CRM fields, fixed browser checks — with human approval on everything that affects customers or modifies data.
Nuance: OpenClaw does not automate “the entire business”; it orchestrates agent steps. Workflows fail on DOM drift, injection, and governance holes — not just on model quality.
Works well
Workflows that are feasible in production
Mailbox triage & classification
What works: Labeling inbound mail, priority, draft proposal — a human sends or adjusts.
Why: Structured input (subject/body); low direct side effects with read-only + draft-only.
Approvals: Required on send/delete; read-only classification is possible without.
Risk: Prompt injection via mail body; separate automation inbox.
Monitoring → alert → human
What works: Gateway/cron reads logs of healthz; agent summarizes; human decides action.
Why: Read-mostly; fits operational AI without autonomous fixes.
Approvals: On every remediation action (restart, config patch).
Risk: Alert fatigue if agent generates too much noise.
Lead processing (structured)
What works: Form/CRM trigger → normalize fields → draft in CRM — no auto-publish.
Why: Fixed schedules; less DOM chaos than free browsing.
Approvals: On CRM write and outbound mail.
Risk: Duplicate records on retry without idempotency.
Simple browser tasks
What works: Read status on fixed dashboard, download export, check off checklist at fixed URL.
Why: Repeatable selectors; limited navigation.
Approvals: On submit and navigation outside whitelist.
Risk: DOM deploy breaks flow — see Browser automation.
Support routing
What works: Categorize ticket, internal note, proposed response — agent does not forward to customer directly.
Why: Human review on customer-facing text.
Approvals: Mandatory on customer reply and ticket close.
Risk: Wrong category → wrong SLA queue.
Risk
Workflows that are dangerous
Short answer: everything involving money, bulk side effects, or customer-facing send without a human — not “because AI can’t do it”, but because governance is lacking.
Financial actions
Risk: Refunds, payments, invoice status — irreversible and compliance-sensitive.
Policy: No autonomous flow; dual control + approvals on every transaction URL.
Destructive terminal/bulk shell
Risk: Scripts on prod host; prompt injection = RCE.
Policy: No shell tool on prod-agent; staged runbooks with human trigger.
Production database writes
Risk: UPDATE/DELETE without human SQL review.
Policy: Read-only agent; exports to review tool.
Bulk browser automation
Risk: Hundreds of submissions; rate limits; account ban.
Policy: Batch with cap; approvals per batch; monitor queue depth.
Autonomous customer replies
Risk: Hallucinated response to customer; trademark damage.
Policy: Always draft + human send; no SMTP without approval.
governance
Human-in-the-loop workflows
Readonly → approval → execute
Agent reads mail/page/logs, proposes a plan; Gateway waits for approve before tool that mutates.
Staged approvals
Classify (auto) → draft (auto) → send (approve). Different policy per environment for each step.
Review flows
Show payload: command, URL, CRM fields, mail body — no generic “allow action”.
Escalation
Timeout pending → ops channel; no silent auto-deny without metric (looks like outage).
Detail: OpenClaw approvals
Combination
Browser + mailbox
in one workflow
Short answer: Mail provides the trigger and injection surface; the browser provides the state and actions — governance must cover both channels.
- → Lead follow-up: mail trigger → browser check CRM/dashboard → approval → CRM update + draft reply.
- → Support: inbox classify → browser status in internal tool → draft → mens send.
- → Scrape internal portal → summary → approval → export to ticket (not open web scraping on product).
- → Governance: same agent with mail + browser = double injection surface — separate profiles and mailboxes.
- Separate automation mailbox
- Separate browser profile per workflow
- No prod admin in agent session
- Audit trail on every mutating step
Limits
Where AI agents still fail
Operational awareness: plan workflows as if these failure modes are normal.
- ⚠️ Hallucinated fields in CRM or incorrect email template.
- ⚠️ DOM change after frontend deploy — selector misses target.
- ⚠️ CAPTCHA / anti-bot — workflow hangs in retry loop.
- ⚠️ Prompt injection in email or web → unwanted tool plan.
- ⚠️ Browser OOM on VPS — Gateway appears “up”, browser tool timeout.
- ⚠️ Edge cases: attachments, i18n, timezone, double threads.
Environment
Production vs. experiment
Experiment / hobby
- →Standalone agent on laptop, looser approvals, no audit SLA
- →Quick selector/tests; no prod-cookies in profile
- →Logging optional
Production / SME
- ✓Docker/systemd Gateway, pinned images, monitoring on healthz + approval queue
- ✓Audit trail: who approved what, when, which URL/command
- ✓Rollback: image tag + config backup before workflow change
- ✓SME: one workflow per business risk class — not one mega-agent
Reference
Example workflow architectures
No autonomous “company brain” — but chains with clear gates.
Inbox → classify → browser → approval → CRM
- 1. IMAP: new mail in automation inbox
- 2. Agent: classification + priority (read-only)
- 3. Browser: customer status in CRM (whitelist URL)
- 4. Approval: CRM patch + draft reply
- 5. Human: send of edit
Monitoring → alert → approval → action
- 1. Cron/health: metric or log trigger
- 2. Agent: summary + proposal (restart / no action)
- 3. Approval: shell or compose restart
- 4. Log audit ID; notify channel
Browser scrape → summarize → approval → export
- 1. Browser: fixed internal page (no open crawl)
- 2. Agent: extract + summary
- 3. Approval: write export / ticket attachment
- 4. No autonomous submission on forms
Short answer: Structured, repetitive flows with clear inputs: email triage, monitoring summaries, normalizing CRM fields, fixed browser checks — with a human on mutating steps.
Cause: Chaos grows with free browse + autonomous send.
Advice: Start with one workflow; measure approval denial rate.
Short answer: No — readonly/classify/draft could be lighter; send, submit, shell, DB write yes.
Cause: The risk lies in side effects, not in thinking.
Advice: approval policies per tool type.
Short answer: It can generate drafts and send them with approval — autonomous responses to customers are a governance risk.
Cause: hallucination + injection in inbound mail.
Advice: human send; separate mailbox.
Short answer: Yes, as a workflow step via browser tools — restrict URLs and approval on submit.
Cause: DOM and session instability.
Advice: Browser automation hardening.
Short answer: DOM changes, CAPTCHA, session timeout, OOM, wrong selectors — not “the model is stupid”.
Cause: operational drift
Advice: monitoring + staging replay; problems.
Short answer: Rarely on production with email/browser/CRM — autonomy belongs in read-only or sandbox steps.
Cause: No human gate on irreversible acts.
Advice: human-in-the-loop on mutating chain.
Short answer: At send mail, form submit, shell, file delete, API mutation, prod-DB, financial URLs.
Cause: compliance and prompt injection.
Advice: Document per workflow in the runbook; audit trail is mandatory.
workflows
Designing a workflow with
governance?
We help SME teams map out workflows: what can be read-only, what requires approval, how email and browser work together, and how to monitor on Docker/Gateway. No “AI replaces your team” — but a runbook that fits your risk class.