No published standard for BYOA exists that I could find. The only documented version I verified, as read in November 2026, is Front’s external AI agents. I did not check Zendesk, Intercom or other inbox vendors, so every example here is Front’s.
Setup order and what the inbox exposes
- Create the agent under Settings → Company → Agents with “Create AI agent”.
- Set Escalation, the inactivity minutes before Front unassigns the thread.
- Grant permissions. Front groups them into categories covering triaging, assisting, resolving and managing resources, and does not define each or map them to tools. Its example says a draft-only agent “doesn’t need permission to send or close conversations.” Read the wizard, then test that send_reply is rejected before you rely on it.
- Connect webhooks for assign, inbound, mention and unassign. Delivery is at-least-once, so dedupe on event_id.
- Verify signatures on the raw body: HMAC-SHA256 over the timestamp header (milliseconds), a colon and the body, keyed with the signing secret, base64-encoded, compared with x-front-signature. No tolerance is set; I would reject anything older than five minutes.
- Run the review below.
Tokens last about 15 minutes with no refresh token, so request a new one early (I would use the 10-minute mark). Operation limits are 50, 200 and 800 a minute on Starter, Professional and Enterprise, shared by all AI teammates and separate from your API rate limit, so budget them on their own.
Slack’s help describes Slack Connect approval as invitation requests reviewed by Workspace Owners and Admins, or people they delegate, on Pro and Business+. It says nothing about apps joining shared channels, and Front’s guide never mentions Slack. Do not assume an agent needs or skips approval; ask your Slack admin and Front before promising a customer anything.
What the agent must never do without approval
Keep with a person: promised dates, refunds, credits, changes to account data, adding anyone to a shared channel, and closing threads. update_conversation_status closes and reopens, so withhold it.
Review rule (my suggestion): a lead who did not write the prompt reviews 30 consecutive drafts per thread type, and a draft passes only if the sent text is identical. Then enable send_reply for that type.
Revoking quickly, an interim procedure because the guide does not cover deactivation or queued drafts: remove permissions (they can be adjusted after setup), unassign open threads, stop the webhook endpoint, then delete queued drafts by hand. Rehearse it once in a test inbox.
Data-handling questions
- Where does the model run, and in which region?
- Does the model provider retain prompts, for how long, and may it train on them?
- Which customer fields reach the agent through read_contact and read_message, including attachments?
- Who can read your tool-call logs?
If the provider cannot answer the second question in writing, keep the agent on draft-only and send it no attachments.
Failure modes and fixes
- Stuck threads. If your slowest turn takes 4 minutes, a 10-minute Escalation leaves 6 minutes of buffer (my numbers).
- Work after handoff. The docs say to stop working on unassign. I found nothing saying add_comment still works afterwards, so post any summary before releasing the thread.
- Unclear sender. The agent’s Name “may be shown to external users depending on the channel”. Check a scratch channel; the guide does not say how replies appear there.