An AI agent belongs in a shared Slack channel or an email queue only if, for each request type, you can say whether it closes alone or goes to a person, and what it leaves on record either way. The thresholds marked as suggested are ours, not vendor or standard figures, so replace them with your own replay results.
What an agent can close alone
An agent should close a request only when it can name the source of the answer. Three kinds of request meet that test in most B2B teams.
Order or account status. The agent reads one record and quotes the status with its timestamp, for example “Order 4812 shipped on [date] and is with the carrier.” It does not estimate arrival dates it cannot see.
Password and login resets. The agent triggers your existing self-service reset email to the address on file and posts only a confirmation, such as “I’ve sent a reset email to the address on file.” It never posts the reset link in the channel, because every member of a shared channel, including other organizations, can read it. It does not set passwords, change email addresses or ask for data it would then have to store.
Answers covered by one published policy. The agent quotes the rule and links the article, for example “Our return window is 30 days from delivery, as set out in the returns article.” The article needs an owner and a last-reviewed date.
Close rule (suggested): the answer is one fact from one current source, and the customer does not write again on the same point within 24 business hours. Intercom uses a related definition for billing: its help article defines a resolution as a conversation where “No further help is requested after the last AI answer.” Anything that fails the rule stays open for a person.
What must route to a person
Route anything that spends money, binds your company or promises a date.
- Refunds above a limit you set. Pick the limit at the amount a team lead would approve without checking the margin on that account. Under it, a refund that matches a written rule can be started by the agent, and a person confirms before money moves.
- Partial refunds, credits and policy exceptions, at any amount.
- Legal, security and contract questions, such as data retention terms, sub-processor lists or indemnity wording. The agent routes the question with the customer’s exact words attached and does not answer it.
- Any promise with a date. Only a person commits your company to a date.
- A customer who asks for a human, or whose next message on the same point shows the first answer did not help.
For the human-request check, match whole phrases such as “talk to a human”, “real person”, “speak to someone” or “a person”, and do not match the bare word “agent”, because customers write “AI agent” and “support agent” about the product itself.
The handoff message should name a person and say what happens next: “I’m passing this to Dana on our support team, who can approve that. She’ll reply in this thread.” A routed thread gets a human reply within one business hour (suggested). If it does not, the agent stays silent in that thread and the support lead gets a notice with the thread link.
Capacity decides whether that hour is real. Take last month’s routed threads, multiply by the minutes a person needs per thread, and compare the result with the hours one named person is actually on shift. If routed volume exceeds what one on-call person can clear in an hour at peak, name a second person for those hours or widen the hour to a window you can keep.
Decision rules in one table
| Request | Agent may close | Route to a person | Rule |
|---|---|---|---|
| Order or account status | Yes | No | One record, quoted with its timestamp |
| Password or login reset | Yes | No | Triggers reset email to address on file; no link posted |
| Policy answer | Yes | If the article is ambiguous or unowned | One current article, linked |
| Refund under your limit that matches a written rule | Starts it, never closes it | Confirms before money moves | Agent starts, person confirms |
| Refund above your limit, partial refund, credit | No | Yes | Any amount outside the rule |
| Legal, security or contract question | No | Yes | Route with the customer’s words attached |
| Dated promise | No | Yes | Only people commit dates |
| Customer asks for a human, or the first answer failed | No | Yes | Immediate handoff |
Running it in Slack Connect and email
In Slack Connect the agent is an app installed in your workspace. The Slack docs say slash commands can be invoked only by users belonging to the workspace where the app is installed, and that they, and message actions, are limited to the team that installed the app. The docs do not describe a customer-side button, so test that before relying on one. Build the handoff as a typed phrase and try it from a customer’s account before launch.
The docs also say messages from incoming webhooks are visible to all members of a shared channel. Treat whatever the agent posts as readable by everyone in the channel, not only the person who asked. The help article says a channel can host up to 250 organizations and that people can only access the channels they have been invited to join, but it does not say who can read messages.
If a customer runs its own bot or AI in the same channel, your agent ignores other bots, never takes instructions from them and routes a human to the thread when another bot’s message looks like a request. Otherwise two automated senders can answer each other.
Disclosure: tell the customer in the first reply that they are talking to an assistant. Suggested wording: “I’m an AI assistant for [Company]. I can check order status and trigger a password reset email to the address on file. For anything else I’ll pass this to a person on our team.” Put the same sentence in your contract’s support schedule so the customer’s legal team has seen it.
Email follows the same rules. Skip automatic replies and mail from your own domain. When a request moves between Slack and email, match it on account and order number. If two open threads match, the agent routes the request with both threads linked. After a handoff, the agent stays out of that thread until a person marks it done.
Audit trail, data and review
Log each reply with:
- the channel or mailbox, the thread ID and the customer message it answers
- the reply text exactly as posted
- the source article, its URL and its last-reviewed date
- the rule that allowed the reply, with its version
- whether a person edited or approved it, and who
Store this in your own system, not in Slack, because the posted message carries nothing about the article or rule behind it. Reconcile the log against posted messages: a log entry with no matching message is a failed send and must be visible and retryable.
Post the minimum. Quote the order status, not the whole account record. Before launch, ask each AI vendor for its data processing agreement, where conversation logs are stored, what it retains and for how long, and whether your data trains its models. Set your log’s retention to match your ticket system, and assume the customer’s organization keeps its own copy of anything posted.
Review by sampling. Each week the reviewer reads 30 random replies (suggested; adjust for low volume) and checks three things: the source article is current, the rule fits the request, and the customer did not have to repeat themselves.
Rule ownership (suggested split): the support lead writes the rules, finance approves any raise of the refund limit, and a person who did not write the rules does the weekly sample. Review the rule set monthly and after any change to a policy it cites. Record each change with its date, approver and a version number, because the audit log points to that version.
Measure outcomes, not just cost
Track these per request type:
- Wrong-close rate: closes that were reopened or contradicted by a person, divided by all agent closes.
- Containment: agent closes with no handoff, divided by all conversations the agent touched. High containment with a high wrong-close rate means the agent is closing too much.
- Repeat-contact rate: the same customer writing again on the same point within 7 days (suggested window).
- Satisfaction: compare survey scores on agent-closed and human-closed threads, and read the comments on the low ones.
What it costs per closed request
Compare cost per closed request, not list price. As of 2026-11-10, two vendor pricing pages show two charging models:
- Zendesk lists AI agents at $19 per agent per month on Support Team, $55 on Suite Team and $115 on Suite Professional, billed annually. Suite Team includes 5 automated resolutions per agent per month, and resolutions beyond that cost $1.50 on a committed basis or $2.00 pay-as-you-go. Zendesk says you pay only for requests the AI agent resolved with no escalation to a human.
- Intercom lists $0.99 per resolution, charged at most once per conversation. It does not bill when a customer asks for a human, but its article says a Procedure handoff, where Fin follows a Procedure’s instructions to pass a conversation on, is billed at $0.99.
Worked example on Zendesk Suite Team. Four agents cost 4 × $55 = $220 per month and include 20 resolutions. If 300 conversations close by agent, 280 are overage: 280 × $1.50 = $420 committed, or 280 × $2.00 = $560 pay-as-you-go. Totals are $640 per month committed and $780 per month pay-as-you-go, which is $2.13 and $2.60 per closed request.
The two are not like for like. Zendesk charges a seat fee plus overage per resolved request. Intercom charges per outcome on top of its plan price: 300 × $0.99 = $297 covers outcome fees only. Add the plan price, and any Procedure handoffs you use, before comparing.
Choose a vendor on four tests rather than price: it can read your order and account records through an API, it lets you write the route-to-person rules above, it exports the log fields listed earlier, and it works in your shared channels as well as email. Check Slack Connect support in a trial workspace, because help-desk AI products are often built around email and chat widgets first.
Failure modes and fixes
The agent closes too early. The request is marked resolved, the customer writes “that’s not it”, and your reports already count it as done. Fix: count a close only after 24 business hours with no customer reply (suggested), and reopen the thread when the customer writes again.
The agent quotes a stale policy. A return window changed and the agent still quotes the old article. Fix: every citable article has an owner and a review date, and the log stores that date so the weekly sample catches it.
The customer loops. They ask for a person and the agent answers again. Fix: the phrase matching above routes at once, and so does a customer reply showing the first answer failed. There is no second attempt.
The wrong reply has already gone out. Fix, in this order: the support lead pulls the agent from that thread; a named person posts a correction in the same thread within the hour, such as “Correction: the earlier message gave the wrong return window. The right answer is [fact]. I’m sorry for the confusion.”; the person finds every other reply that cited the same article using the log; and anything that cost the customer money or time gets an incident record with the cause, the rule version and the fix. Tell the customer’s contact directly when they acted on the wrong answer.
The SLA looks met when it is not. If agent replies stop the first response clock, your human SLA can look healthy while nobody has read the request. Write the choice into the SLA. Example wording: “First response time is measured from the customer’s message to the first reply from a named member of our team. Automated acknowledgements and AI-drafted answers do not stop the clock, and AI answers are reported separately.” Or, if you let them count: “An AI answer stops the clock only when it resolves the request and the customer does not reply within 24 business hours.” The Slack Connect first response time guide covers how to define the clock.
Rollout order
Move to the next stage only when the last 50 closed requests of the current type have no wrong closes (suggested threshold).
- Replay 50 closed conversations from last quarter through your rules, with the acceptable error count set before you look.
- Run shadow mode for two weeks (suggested): the agent drafts, a person posts, and you compare.
- Switch on order and account status.
- Add password resets.
- Add policy answers, one named article at a time.
- Last, allow refunds under the limit, with a person confirming before money moves.