Customer support
Steer AI drafts with a few words
Your help center can't cover every edge case—type a few keywords into initdesk's draft instructions and get a send-ready reply in seconds, with you still in control.
Most support tools now suggest a reply before you type. That is useful when the question matches something you have already documented—a billing FAQ, a setup step, a policy you wrote down last quarter.
It falls apart on the threads that do not have a clean doc: one-off exceptions, angry customers who need empathy before facts, a bug you can acknowledge but not fix today, or a reply that must sound like you after six months of founder-led email.
You could stare at a blank compose box. Or you could give the AI three words of direction and edit something good in seconds.
When the knowledge base is not enough
initdesk AI drafts pull from your help center, past tickets, and connected context—but no library covers every judgment call. The gaps show up in predictable places:
- Policy exceptions — “We usually do not refund after 30 days, but this customer had a documented outage.”
- Tone shifts — the thread needs warmth before policy, or the opposite: short and factual after too much back-and-forth.
- Partial answers — you know the next step (escalate, wait for engineering, offer a credit) but the customer has not asked the exact question your article titles.
- Relationship context — a long-time customer, a trial about to convert, a message that is really a cancellation risk in disguise.
In those moments, a generic draft is a starting guess. A few instructions turn it into a draft you would actually send.
Add optional instructions, then Draft
In the reply box, initdesk shows Use the AI Draft or reply here. Below that, Add optional instructions to AI draft is where you steer.
You do not write the full email. You write keywords or a short phrase—what you would mutter to a teammate sitting next to you:
| Situation | Example instructions |
|---|---|
| Upset customer, fair ask | offer 1-week extension, no refund |
| Feature not on roadmap | decline, will not be available in the next 6 months |
| Bug acknowledged | confirm we're investigating, no ETA on fix |
| Set date expectation | next update Tuesday |
Click Draft. The model uses the ticket, customer history, help center, and your line together. You read, tweak a sentence or two, send. Human review stays in the loop; the blank-page minute disappears.
What makes instructions work (and what does not)
Do:
- Be specific about outcome — “offer credit” beats “be nice.”
- Name constraints — “do not promise a ship date,” “no discount above 10%.”
- Say who is speaking — “founder voice, first person” vs “team we.”
Skip:
- Pasting the entire reply you want (you are drafting, not dictating).
- Instructions that contradict your help center without saying so—if you are making an exception, say
exception to refund policy, outage on our side. - Letting AI invent commitments you have not approved—same rule as any draft; you still press Send.
If you are handing a thread to someone else, instructions help the new owner land in the right voice faster—pair with the handoff habits in Hand off a support thread without losing the customer.
Keep it human and efficient
The goal is not auto-send. It is good-enough in one pass:
- Open the ticket; skim what the customer actually needs.
- Type a short instruction if the first suggested draft misses the mark.
- Draft → edit → send.
Teams that document repeating answers in the help center still win—AI drafts get sharper when docs are fresh. Instructions are for the long tail where docs will never be perfect. See Self-service that actually gets used for keeping that base layer trustworthy; use instructions for everything else.
When tone drifts across people, not just across AI runs, the fix is shared voice rules—From founder-led support to a repeatable voice covers the one-page guide that makes both human and AI replies sound like one company.
A quick habit for busy inboxes
Before you rewrite a draft from scratch, try one instruction line. If the second draft is close, you saved five minutes. If it is still wrong, you learned the ticket needs an internal note or escalation—not more prompting.
Fifteen minutes once a week:
- Pick three tickets where you heavily edited an AI draft. What one instruction would have fixed most of the edit?
- Turn any repeat answer into a help center article so next time needs fewer keywords.
- Check two drafts you sent unchanged. Still accurate? Good sign your docs and defaults are aligned.
initdesk is built for small teams who run email support as a shared inbox—AI drafts with optional instructions, help center context, tags, internal notes, and Linear when product needs the story. See Product Updates for recent AI improvements. If you want to compare draft workflows or tooling, say hello on X @initdeskhq.