Break it in preview.
Not in production.
Check a procedure on your last 50 conversations, replay a past one with the draft, and roll out inbox by inbox. After go-live, a scorecard grades Aide beside your own team every night.Part of the Agent Governance Engine: nothing is sent to a customer or written to your helpdesk while you preview or watch.How do you test an AI agent before customers see it?
Run it on conversations you already have: in Aide you check a procedure against your last 50 conversations, replay a past conversation with the draft, and chat with the agent in a preview that never reaches customers. Then roll out inbox by inbox, and a scorecard grades Aide beside your own team every night.UpdatedWatch it work before customers do
Every example below is a preview or a staged rollout, not a reply to a customer. This is what your team sees first.
When to use
The customer says an item arrived broken or damaged. Not when the parcel is only late.
Scoring the latest conversations…
Matches of the latest 50 conversations
Matched
- Kettle arrived dented96%Nora Feld · 30 Oct
- Grinder lid cracked93%Ravi Menon · 24 Oct
Declined
- Still waiting on my mugs31%Lucia Romero · 29 Oct
The parcel is only late. Nothing arrived broken.
Use for this
Test
- Kettle arrived dentedNora Feld · 30 Oct
- Grinder lid crackedRavi Menon · 24 Oct
Answering from 42 pages of hearthandbean.com.
My grinder arrived with a cracked hopper.
Sorry it arrived cracked, Rosa. We’ll send you a new hopper this week.
Sorry it arrived cracked, Rosa! Could you send a photo of the hopper? Then we’ll get a new one out to you.
- Watch only: Aide drafts in Aide, writes nothing to Zendesk
- Test senders only
- A share of conversations
- Everyone
| Write | Status | Sent as | Created |
|---|---|---|---|
| Reply by email | SkippedWatch only: not sent | — | 4 minutes ago |
| Reply by email | SkippedWatch only: not sent | — | 11 minutes ago |
| Reply by email | SkippedWatch only: not sent | — | 26 minutes ago |
When to use
The customer says an item arrived broken or damaged. Not when the parcel is only late.
Scoring the latest conversations…
Matches of the latest 50 conversations
Matched
- Kettle arrived dented96%Nora Feld · 30 Oct
- Grinder lid cracked93%Ravi Menon · 24 Oct
Declined
- Still waiting on my mugs31%Lucia Romero · 29 Oct
The parcel is only late. Nothing arrived broken.
Use for this
Tested before launch. Graded every night after.
Testing decides what goes live. The Quality page keeps it accountable once it does.Replies judged acceptable, matched on 17 topics and weighted to Aide's mix of them, against 5 teammates
- Aide79–91%
- Team median72–89%
- Top quartile83–95%
Bars are 95% intervals: the range the true share likely sits in.
Open only the first reply with a greeting
Applied by Aide 3 Nov 2026 at 04:12
Your team greets the customer once, then goes straight to the answer. Aide greeted every reply.
Before
Open every reply with "Hi {first name}".
After
Open the first reply with "Hi {first name}". In follow-ups, answer straight away.
Aide can propose actions, but a person confirms them before anything happens.
- None
- Drafts
- AI Agent
Test on the conversations you already have
Connect your helpdesk or add Aide Chat, and testing runs on your own conversations. No migration and no engineering.
Drafts on recent tickets as soon as you connect.
Watch only, then test senders, per inbox.
Per-inbox rollout from Watch only to everyone.
Order and shipment conditions in every preview.
Contact fields in a procedure's Only if conditions.
Test the agent with Claude during setup.
Frequently asked questions
AI agent testing shows you what an AI support agent would do before your customers see it. In Aide, testing runs on your own conversations: Check trigger tries a procedure's When to use on your last 50 conversations, Test replays a past conversation with the draft procedure, and the agent preview lets you write in as a customer.
In Aide you test on your own conversations. Check a procedure's When to use on your last 50 conversations, replay a past conversation with the draft, and chat with the agent in its preview. Then keep an inbox in Watch only before widening it to test senders, a share of conversations and everyone.
Aide tests on your real conversations. Check trigger tries a procedure on your last 50 conversations, and Test replays a past conversation with the draft. When you connect Zendesk, Gorgias or Front, Aide also drafts replies to up to 10 recent open tickets and 3 recently solved ones inside Aide.
Before launch, Aide is judged procedure by procedure on your own conversations. Check trigger shows which recent conversations a procedure would pick up and why it passed on the near misses, and Test replays a past conversation with the draft. On recently solved tickets, Aide's draft sits beside the reply your team sent, so you grade it against your own standard.
Testing in Aide sends nothing to customers. A preview shows the reply Aide would write, or why it would hand off or hold back, and nothing is sent or written. An inbox in Watch only drafts inside Aide and writes nothing to your helpdesk.
Each helpdesk inbox in Aide moves from Watch only to test senders, where only the addresses you list get Aide's replies, then to a share of conversations and everyone. You can also put one topic at AI Agent and keep the rest at Drafts or None.
After go-live, Aide's Quality page shows a pass or fail scorecard against a written quality bar. Evals replay against the agent every night, and Aide's replies are graded beside your team's median and top quarter on the same scorecard. Quality also says, topic by topic, when Aide's drafts are good enough to send.
In a conversation in Aide, a teammate marks a draft Good or Wrong with a reason, and that verdict overrides Aide's own review. A Wrong draft becomes an eval case that replays against the agent every night, so every later change is checked against it.
Owners and admins run Check trigger in Aide and set a procedure live with a note. Edits to a live procedure stay a draft while the live version keeps running, and version history restores an earlier one.
Previews and spam are never counted as AI replies in Aide. Replies to test senders count like any reply, and drafts kept inside Aide during Watch only count once per customer message, like any draft.
Set it up
Governance and trust, defined
The vocabulary of testing and trusting AI agents, defined.
Test Aide on your own conversations, free
Connect your helpdesk and see Aide's drafts on your recent tickets in minutes.