Answering questions is half of what customers need. The other half is getting something done: tracking an order, swapping a size, booking a fitting. Workflows are how an AI agent does those jobs reliably, step by step, with rules you can read.
Start from one intent
Don't try to build everything at once. Pick the single most common task that needs data or an action. For most stores that's order tracking; for service businesses it's booking. One well-built workflow teaches you the pattern for the rest.
The building blocks
Most workflows combine the same steps:
- Trigger: a message that matches the intent ("where's my order").
- Collect inputs: order number, email, size, date.
- Verify: check the inputs belong together before sharing data.
- Look up: read order, product or availability data.
- Decide: branch on the result (shipped, delayed, delivered).
- Confirm: ask "shall I go ahead?" before any change.
- Act: create the exchange, book the slot, capture the lead.
- Hand off: when a branch needs a person.
LiveCart's AI Studio offers around fifty of these blocks, from AI decisions to API calls.
Example: exchange a size
- Trigger: "These are too small"
- Collect: order number and email
- Verify: they match
- Look up: order date and item
- Decide: within the return window? If not, hand off
- Look up: stock in the next size
- Confirm: "Swap UK 8 for UK 9? We'll send a returns label."
- Act: create the exchange
- Reply: next steps and timeline
Keep humans in the right places
Put a handoff branch wherever judgement is needed: damaged items, out-of-policy requests, high values. A workflow that hands off cleanly is better than one that tries to cover every case.
Test like a customer
Before publishing, run the workflow with real messages, including awkward ones:
- Missing order numbers
- Typos in emails
- Changing their mind halfway through
- Asking an unrelated question mid-flow
Traces that show each step's inputs and outputs make it obvious where a branch went wrong.
Version everything
Publish workflows as versions. If a change causes problems, restore the previous version while you fix it. Small, frequent improvements beat a big rewrite.
Where to go next
Once one workflow is solid, add the next most common task. Order tracking, returns, product finder and lead capture cover most of what online stores need. See AI agent guardrails for the safety rules every workflow should follow.
