Iro AIWhat you'll be able to do
- Draft a support reply that matches your brand voice and the customer's tone
- Summarize a long, multi-message ticket into the facts and the ask
- Turn a good reply into a reusable macro or a help-doc draft
- Triage a ticket and decide what to escalate — and to whom
- Get AI to flag missing policy instead of confidently making it up
Inside the path
A focused set of five-minute lessons — each one ends with a hands-on exercise, not a quiz you can guess.
Replies in your brand voice 6 min
Give the model voice examples and tone rules so drafts sound like your team, not a robot.
Summarize a messy ticket 5 min
Condense a 40-message thread into the customer's problem, what's been tried, and the ask.
Macros and help docs, fast 5 min
Turn your best replies into reusable macros and first-draft help-center articles.
Triage and escalation 5 min
Use AI to classify urgency and route tickets — while you keep the judgment call.
Never invent policy 5 min
The most important support skill: make the model flag what it doesn't know instead of guessing.
Try a sample exercise
This is the kind of card you'd practice inside Iro — you do the thinking, then get feedback.
◆ Sample exercise · Prompt practice
A frustrated customer wants a refund on an annual plan they bought five months ago. You're not sure your policy allows a prorated refund, and you want AI to help draft a reply while you check.
Your task: Pick the prompt that keeps you fast without risking a made-up promise.
- "Write a reply telling this customer whether they can get a prorated refund."
- "Here's our refund policy text: [paste]. Here's the customer's message: [paste]. Draft an empathetic reply in our brand voice. Use ONLY what the pasted policy states. If the policy doesn't clearly cover a prorated refund at five months, do not promise or deny one — instead acknowledge their frustration, say I'm confirming the exact terms, and mark [CHECK: prorated refund eligibility] where I need to verify before sending."
- "You're a customer support expert. Write the most helpful refund reply possible for this upset customer."
- "Draft a reply and just assume we allow a full refund to keep them happy."
See why the second prompt wins
The winning prompt grounds the model in real policy text and forbids it from going beyond it, sets the brand voice and empathy, and — most important for support — tells it to flag uncertainty instead of inventing an answer with a clear [CHECK] marker for the human. The losing options all let AI commit the company to a promise it can't verify: deciding refund eligibility, being maximally "helpful," or assuming a full refund. In support, a confident wrong answer becomes a promise you have to honor. In Iro you'd write your own version and get feedback on whether it stays grounded or quietly guesses at policy.
The one rule that makes AI safe in support
AI can draft a warm, on-brand reply in seconds — which is exactly why it's dangerous in support if you skip one rule: it must never invent policy. A model asked whether a refund is allowed will happily produce a confident, specific, and completely made-up answer. Send that, and you've made a promise your company now has to keep.
The fix is to ground it. Paste the actual policy or help-center text, tell the model to use only what's there, and instruct it to flag anything it can't confirm with a marker like [CHECK: …] instead of guessing. That single habit turns AI from a liability into a genuinely fast, safe drafting partner.
Where AI speeds up a support queue
- Brand-voice replies: feed it tone rules and a few example replies so drafts sound like your team and match the customer's mood.
- Ticket summaries: condense a long thread into the problem, what's already been tried, and the actual ask — great for handoffs and escalations.
- Macros and help docs: turn your best answers into reusable macros and first-draft help-center articles.
- Triage: classify urgency and suggest routing, while a human keeps the final judgment call.
Everything AI touches here is a draft or a summary. The reply that goes out, the policy that's quoted, and the escalation that's made still belong to a person who checks the work.