The Go-Live Checklist for Your AI Assistant

Assistants rarely fail loudly. They fail by being live on a page nobody checked, answering a question nobody tested, with a hand-off nobody wired up. Everything works; nothing is right.

This is the checklist to run before you point real traffic at one — about forty minutes, ordered so it catches the most problems soonest.

1. Knowledge coverage

Write down the twenty questions you actually get — the ones in your inbox, not the ones you wish people asked.

• Every one of those twenty has a document that answers it, or an instruction that deliberately declines it.
• Every document in the Knowledge Base card shows status indexed. Anything at error is invisible to the assistant, silently, forever.
• No superseded version of a live policy is still uploaded. Faced with two contradictory passages, the search cannot tell which is current.
• Nothing internal or confidential is in there. There is no "staff only" flag — assume every sentence is quotable.
• Anything time-sensitive has an owner. Hours and policies drift, and the assistant quotes the stale version confidently.

Then test it the way a stranger would: ask ten of the twenty in a visitor's words. Look for the Referenced knowledge base marker on answers that should be grounded — if it is missing, retrieval did not fire and the document is the problem. Building a Knowledge Base Your Assistant Can Actually Use covers what to fix.

2. Tone and boundaries

• Instructions on the General tab open with a role and an audience, in that order.
• Answer length is capped. "Under 60 words unless asked for detail" is a good default.
• The topics it must refuse are named explicitly, with the redirect wording you want.
• Nothing in the prompt is a fact that will expire. Facts belong in documents.
• If voice is on, the instructions avoid formatting that only works on screen — bullets and tables read badly aloud.

Then run the five-question test, reading each reply as a customer would: one easy question, one it should refuse, one frustrated and badly punctuated, one adjacent but out of scope, one it has no business answering. Set Up Your Assistant's Persona and Tone has the detail.

3. Set a goal — before launch, not after

Goal, on the General tab, needs a sentence a reader could verify from a transcript.

This is the most commonly skipped item and the only one that cannot be fixed retroactively. Grading runs when a conversation ends and is never backfilled, so a launch week with no goal is a week with no outcome data at all.

4. Escalation

• At least one escalation channel exists and is Enabled. Zero channels means the assistant has no way to reach you, whatever the instructions say.
• Email escalation points at a shared inbox, not one person's address, with the transcript included.
• The channel's Description says when to use it, in language the model can act on.
• Your instructions tell the assistant to collect a name and email before escalating, and not to promise a response time.
• You have triggered one for real and confirmed it arrived, with a summary a colleague could act on.

Untested escalation is the worst failure here: you discover it from the customer who says nobody replied. Set Up Team Handoff So a Real Person Gets the Context walks through the setup.

5. Domain safelist and embed

This is where most launch-day failures actually live.

• Authentication Mode on the Deploy tab is Allowed Websites, unless you deliberately built the server-signed flow.
• Every host you embed on is in Allowed Domains as its own entry, including your staging host and the www. variant. Matching is exact and there are no wildcards — example.com covers neither www.example.com nor shop.example.com.
• The script tag sits before </body> on every page where the assistant should appear.
• The tag's src really points at /static/va-wave-widget.js. The bundle finds itself by looking for its own filename, so a renamed or proxied copy loads and then does nothing.
• Every page is served over HTTPS. In production, requests from a plain HTTP origin are rejected.

You do not need to re-paste the tag when settings change — colours, greeting, voice and knowledge are fetched on each page load. Your First 15 Minutes With Hiroi covers the embed end to end.

6. Appearance on your real site

The preview is not your site. Load the real one.

• The orb does not collide with a cookie banner, a back-to-top button, a sticky purchase bar or a toast.
• Checked on a phone and a desktop. Most collisions only appear on one.
• The Welcome Message beside the orb and the opening chat message both say what you think — they are two different settings on two different tabs.
• If voice is on, Listening and Speaking are visibly different colours. They carry the whole "is it my turn?" signal.

7. Voice, if you are enabling it

• Enable Voice Input and Enable Text-to-Speech are both on. Voice is refused with only one of them.
• You have heard the chosen voice read a real answer — one containing your product name, a number and an abbreviation.
• Your product names, surnames and model numbers are in the Speech vocabulary (key terms) field.
• You have tested on a phone with real background noise, not just a quiet desk.

Choosing a Voice for Your Assistant has the full pass.

8. Limits, credits and the switch

• Rate Limit on the General tab is high enough for the page you are embedding on. It is a per-assistant hourly budget, not per visitor — raise it before a busy launch, not after visitors start getting refused.
• Your credit balance covers the traffic you expect. At zero the assistant stops answering until you top up; nothing is lost. See the pricing page, and consider auto top-up.
• If you need visitors to accept terms first, the terms gate is on and the terms text is filled in.
• The editor header badge reads Live, not Off.
• Everything is saved. Preview and the test page load the last saved version, so an unsaved edit is invisible everywhere.

9. The final ten minutes

Sit with the assistant on the real page, as a stranger would.

1. Ask your five most common questions and read the answers as a customer.
2. Ask one thing it must refuse, and read the refusal out loud — a refusal is the most emotionally loaded thing it says all day.
3. Ask for a human and follow the escalation all the way to the inbox.
4. Type something rude, then something empty, then something enormous.
5. Reload and check the conversation behaves as you expect on a second visit.

If any of those made you wince, fix that one thing and run them again. Do not launch on a wince.

Week one: what to watch

Ten minutes a day is enough.

Change one thing at a time. Two changes in the same week and the following week's reading tells you nothing. Reading Your Assistant Analytics sets out the loop in full.

If something goes wrong

The fastest rollback is one switch: Assistant Status on the General tab. Turning it off takes the assistant down everywhere immediately, without deleting anything and without touching the embed code on your site. Visitors stop being served, and every setting, document and conversation stays where it was.

Fix the problem, turn it back on. Better than pulling script tags out of your templates at 9pm.

Everything on this list lives in the assistant editor at hiroi.ai.