A simple test for each step
For every action an automation takes, ask four questions:
- Does it involve money? Charging, refunding, invoicing, discounting.
- Does it make a commitment? Confirming a booking, promising a date, accepting a scope change.
- Is it hard to undo? Sending a message to a customer, deleting a record, cancelling a service.
- Does it touch sensitive data? Sharing details with a third party, exporting records.
If the answer to any of these is yes, consider an approval step — at least until you've seen the step work correctly many times.
Steps that usually run without approval
- Internal notifications and task creation.
- Acknowledging receipt of an inquiry (see inquiry acknowledgements).
- Reminders for confirmed bookings, based on verified status.
- Updating a CRM field from a verified event.
- Logging and reporting.
Steps that usually deserve approval
- The first invoice for a new client, or any invoice with non-standard terms.
- Refunds and credits.
- Messages that quote prices, timelines or availability not already published.
- Bulk messages to many customers.
- Deleting or merging records.
- Anything generated by AI that goes directly to a customer.
How to add approvals without friction
- Batch them. One daily digest of drafts to approve is easier than twenty separate notifications.
- Make approve and reject one click, with the key facts visible in the message.
- Set a default for no response. For most steps, the safe default is "don't act and remind me", not "act anyway".
- Record who approved what, and when. This helps when a client asks.
- Remove approvals deliberately. If a step has been approved without changes many times, consider letting it run — and document that decision.
A note on AI in workflows
AI can draft replies, summarise calls or classify inquiries. Those outputs can be wrong in ways that look right. Frameworks like the NIST AI Risk Management Framework emphasise human oversight proportional to impact. Practically: review AI-drafted customer messages before sending, and don't let an AI classification alone decide pricing, eligibility or urgency.
Keep the human in charge of exceptions
The automation should recognise when it's out of its depth — missing data, unusual values, customer complaints — and hand over instead of guessing. That's not a failure of the automation; it's the design working.
Questions
Won't approvals slow everything down?
Only if they're in the wrong places. Put them where a mistake is costly or hard to undo, and let routine, reversible steps run on their own.
Does using AI in a workflow change this?
It raises the bar. Generated text or classifications can be wrong in plausible-sounding ways, so customer-facing AI output and decisions based on it deserve review until you have evidence they're reliable for your use.
Sources and further reading
- NIST AI Risk Management Framework nist.gov
Official documentation changes. Check the current version before you rely on a rule, limit or setting.
Next step
Start with one workflow worth fixing.
Bring the real steps, the tools you already use and the place the handoff gets stuck. We review feasibility, access and approval points before quoting — no instant integration promise.
See Business Automations for what we scope, or prepare a custom quote brief.