Full automation removes the person from a class of questions. A drafting copilot keeps one: someone reads the draft and clicks send. This page covers where that click happens in a shared Slack channel, which drafts need a second reader, what to do when a draft cites a stale policy, and how to tell whether drafts help.
What this page covers
Zendesk is the worked reference, because its help center documents the button-level behavior of its copilot features. The review rules apply to any copilot that drafts replies, in a help desk or in a shared channel. This page does not assess Slack’s own AI features. Dates and limits are as of 2026-11-11.
How a draft reaches a Slack Connect thread
Most copilots do not live inside the customer’s channel. They live in your help desk or shared inbox, and the draft appears in a composer only your team sees. Three setups are common, and only two are copilots.
- Draft in the inbox composer. The agent reads the draft in the thread, edits it and sends. Check which name the reply appears under in your own tool. If it posts under the sender’s name, the customer sees a teammate and that teammate owns the text.
- Draft in an internal triage channel. A bot posts a suggested reply as an internal message, a person edits it, then pastes it into the shared channel. It is slower, and the paste step is where tier 3 checks get skipped.
- Bot posts straight into the shared channel. That is automation, whatever the vendor calls it. Apply the automation rules at the end of this page.
Test the setup before launch. Open a test shared channel with a second workspace you control, trigger a draft, and confirm that nothing reaches the customer-visible side until a named person clicks send. Check which name the message appears under.
Zendesk’s suggested first replies work like the first setup. The agent presses Tab to accept the draft or keeps typing to edit it, and presses Escape to reject it. The help center says agents should review the text in the composer before clicking Submit, and that rejecting a suggestion doesn’t retrain agent copilot or make it suggest replies less often. So rejections will not improve the tool. Your own edit labels and reopen counts will.
Zendesk’s auto assist is a separate path. Its suggestions appear in the ticket interface in place of the composer and need agent approval before they go to a customer. The help center also says auto assist doesn’t work on AI agent tickets, so decide queue by queue which path a ticket takes.
Which drafts need a human read before posting
Assign each draft a tier by what it says, not by how valuable the customer is.
- Tier 1, acknowledgements and handoffs. “Thanks, we have your report and are checking it now.” A skim is enough.
- Tier 2, how-to answers. The draft names a setting or a button, with no figures. Open the linked article, confirm the setting still exists, then send. If the draft states a limit or any other figure, treat it as tier 3.
- Tier 3, anything with a number, date, price, refund, credit or promise. A second person reads the exact text that will post.
- Tier 4, anything a person has to write: escalations, data deletion, security questions, outage updates, and any customer who has said they are considering leaving. Use the draft only as a list of what the customer asked.
If a draft contains a figure, a date or a commitment, it is tier 3.
The second reader must be a teammate on your side who can see the channel. Slack’s help center says a Slack Connect channel can host up to 250 organizations, your own included, so check the member list rather than assuming. Record the review in an internal note on the thread, never in the shared channel. Copy this note for every tier 3 draft:
Draft: [paste the exact text] Source: [policy or article URL, with its last-updated date] Reviewer: [name], reviewed at [time] Label: draft-sent-as-is or draft-edited
When the reviewer is away. Name one reviewer and one backup per shift, in the topic of your internal support channel. If the reviewer hasn’t picked up a tier 3 draft within 15 minutes, ping the backup. If the backup is silent for another 15 minutes, tell the support lead and have the sender post a holding reply without the figure. For example: “I’m checking the exact limit for your plan now, and I’ll reply in this thread by 3 pm your time.” If the check isn’t done 30 minutes before that time, post the new time in the same thread. A promised time that passes without a word is the failure to avoid.
When the copilot cites an outdated policy
Suppose a refund window moved from 30 days to 14. The copilot pulls a snippet that still says 30, writes a confident draft, and a reviewer who tiered it as an acknowledgement skims it. The copilot isn’t lying. It reads the page it is given, and nothing checked that page’s date.
- Give every policy page an owner and a review date. Refunds, SLA terms, pricing and data retention drift most.
- When a policy changes, search your macros and snippets for the old figure before the change is announced, and edit the source once.
- Require the source link in the tier 3 note. If the page predates the change, don’t send. Fix the source, regenerate the draft and send that one.
If a wrong figure has already gone out, say so once: “I gave you the wrong refund window earlier. The current window is [figure], as set out in [link]. I’m sorry for the confusion. If the earlier date affects your order, send me the order number and I’ll check it today.”
If a source went untouched through a policy change, treat every draft that used it in that period as suspect and re-check them from a list you keep from the start.
What the copilot sees
Before you switch a copilot on, get written answers from the vendor on what text goes to the model for each draft (thread, customer name, snippets, account fields), how long it is kept, and whether it trains models. Decide who owns the question of telling customers a reply was AI-drafted, and check it with whoever handles your contracts. This page gives no legal advice. Set a redaction rule too: card numbers, passwords and API keys never go into a draft. If a customer posts one, ask them to remove it, rotate the credential if live, and don’t repeat it.
Measuring handle time without hiding reopens
Track four numbers per queue, per week.
- Handle time, from assignment to resolved. Pick one definition.
- Reopen rate: resolved threads with a new customer message within 7 days, counted by thread. Report a count and a base.
- Edit share: the share of sent drafts edited before posting, from the “draft-sent-as-is” and “draft-edited” labels. Without the labels this number doesn’t exist.
- First response time, by your SLA’s definition. The first response time guide covers what counts.
Worked example, with made-up numbers. Before drafts, a queue resolved 200 threads at a median of 22 minutes, and 14 reopened within 7 days, a rate of 14 in 200. Four weeks after, it resolved 230 threads at a median of 15 minutes, and 27 reopened, a rate of 27 in 230. Reopens nearly doubled while volume rose by 30 threads.
Compare rates on the same base, not raw counts. The old rate applied to 230 threads gives about 16 reopens (14 × 230 ÷ 200), so the baseline to beat is 16, not 14. The 27 is well above it.
Split reopens by label. If they cluster in “draft-sent-as-is” threads, fix the review rule. If they spread evenly, the drafts are answering the wrong question. Read 10 unedited sent drafts a week and mark each answered, partly answered or missed.
The decision rule: if median handle time falls and the reopen rate rises, tighten the tiers before adding a queue. In the example, the team applies the tier 3 second read to every draft with a figure and adds a queue only when the reopen rate is back near 14 in 200.
Copilot or automation, queue by queue
Automate a queue when its answers come from one fact that changes rarely and can be checked automatically, such as where the invoice PDF lives, and a wrong answer is fixed by one follow-up. Keep a person with a copilot when the answer depends on the customer’s contract, environment or history. “Can we get a credit for last month’s outage?” is tier 3 and needs account context.
To sort a queue, pull its last 100 threads and count how many turned on one unchanging answer. If most did, it is a candidate. Re-run the count each quarter, because a policy change can turn a fixed fact into an account-dependent one.
Moving off a retiring feature
Zendesk’s help center says it is discontinuing suggested first replies on January 19, 2027. Move the parts that outlast the feature before that date.
- Export macros and snippets to a sheet with columns for owner, source URL, last review date and tier.
- Keep the tier 3 note and the two labels as plain text in internal notes, so any tool can carry them.
- Record your edit share and reopen rate for the last four weeks as a baseline.
- Run the replacement on one queue, compare against the baseline for four weeks, and only then move others.