A workflow runs a sequence of steps when something happens — greeting someone who opens the messenger, routing a conversation, collecting a detail, following up later. They're straightforward to build, which is exactly why most teams end up with more of them than they need.
The reliable candidates share a property: the right answer doesn't depend on judgement.
Greeting someone and offering the options they might need.
Routing by something factual — the channel, the plan, the page they were on.
Collecting a detail you always need before anyone can help.
Asking for a rating after a conversation closes.
Tagging and prioritising, so the queue is sorted before anyone looks at it.
These work because doing them by hand is pure overhead. Nobody is better at asking for an order number than a form is.
Automation is bad at the things support is actually for: deciding whether to make an exception, recognising that someone is upset, working out that the question asked isn't the question meant.
The failure mode is specific and expensive. A customer with an unusual problem meets a flow built for the usual one, answers three questions that don't apply, and arrives at your team more annoyed than when they started — having spent four minutes proving your product doesn't understand them.
A workflow follows the path you drew. Kai reads the question and answers from your content, with no path at all.
That makes workflows right for process — what happens, in what order, with what collected — and Kai right for questions. Most good setups are a short workflow that hands over to Kai quickly, rather than a long flow trying to anticipate what someone will ask.
Every step is a chance for someone to give up. A flow with three steps is invisible to customers; one with eight feels like an obstacle course, and the people who abandon it are people who wanted to tell you something.
When you're tempted to add a step, ask what you'd do differently with the answer. If nothing changes, the step is costing you conversations for information nobody uses.
