The common way to report deflection counts bot answers served. That number rises every time the bot talks, whether or not the customer’s problem went away. Below are the terms, the counting rules that hold up in a budget review, the questions a bot should ask before it answers, and how to see when deflection costs satisfaction.
Three terms and one count
Containment means the bot handled the conversation with no human handoff. Resolution means the customer confirmed the issue is fixed. Deflection, as used here, is stricter than both: a new request where an automated answer was the last substantive reply, that never reached a person, and that had no new message about the same issue for the whole hold window.
Zendesk’s tiers article defines Contained resolution as “The AI agent handled the interaction to completion without the customer requesting further assistance.” That is a vendor billing tier. It can be a useful input, but it does not tell you whether the problem came back.
Deflection rate is deflected requests divided by all new requests in the period whose 7-day window has closed. The denominator includes requests the bot never touched. Otherwise a bot that only sees easy questions looks better than it is.
A request counts as deflected only when all four hold:
- The request is new. It starts with the first top-level message in a shared channel or a new email. Use the same start time as your first response clock, as set out in Slack Connect SLA: how to measure first response time.
- An automated answer was the last substantive reply.
- No person was involved at any point in the conversation.
- The customer sent no new message about the same issue for 7 days after the bot’s last reply.
Same issue. A message counts as the same issue when it comes from the same contact, or from anyone on the same account, and carries the same product area from your question set (see below). A message in the original thread always counts. A message in a new thread on a different product area starts its own request.
Open windows. A request whose 7 days have not ended by the report’s end date is open. Requests still in their window are reported separately as open, with their count shown beside the rate, and they are left out of the denominator. A request from 3 days before month end is open, not deflected and not failed.
Returns after day 7. Decision rule: a message from the customer on day 6 reopens the original, which is then not deflected. A message on day 8 is a new request in the period it arrives, and it enters that period’s denominator. The original stays deflected.
A worked example with 1,000 requests
Take a month with 1,000 new requests across email and Slack Connect, reported 7 days after month end so that every window has closed. The bot sent 1,850 messages in 600 conversations. Of those 600, 180 went to a person. The other 420 ended without a handoff. In 52 of those 420, the customer wrote again about the same issue within 7 days. Those 52 are exactly what the same-issue rule matches, so anyone with a message export can rebuild the number.
- Bot messages sent: 1,850. This describes bot activity, not customers.
- Conversations the bot touched: 600. Still not deflection, because 180 of them reached a person.
- Deflected requests: 420 minus 52, which is 368 out of 1,000.
Report 368 out of 1,000. The gap between 600 and 368 is what a finance reviewer will ask about.
Now the cost side. Intercom’s pricing page says “Fin is priced at $0.99 per outcome.” Its page also says “You’re only charged once per conversation, even if multiple questions are answered,” but an outcome is counted on how the conversation ended, so 600 touched conversations is not the same as 600 billed outcomes. Suppose your invoice shows 600 outcomes for the month, as a hypothetical. That is $594, and $594 divided by 368 deflections is about $1.61 per deflected request. Divide the bill by 368, not by 600, or the unit cost looks better than it is.
The minimum questions before a bot answers
A bot should not answer until it has five things. Ask for the missing ones one at a time, not as a form.
- Who: the account or workspace, matched to your customer record.
- What: the product area, picked from your list of 10 to 20 support topics rather than free text. This field is also what the same-issue rule matches on.
- The exact error: the message text or code, copied rather than paraphrased.
- Environment: plan, app or browser version, and operating system where it matters.
- When: when it started, and whether it is still happening.
Decision rules:
- If who, what or the exact error is missing, the bot asks for it. Wording that works: “Which workspace is this for?” and “Can you paste the exact error message?”
- If the message contains words such as “outage”, “down”, “data loss”, “refund”, “invoice” or “security”, route it to a person with no bot answer.
- If the topic has no approved answer, route it to a person.
- If all five fields are present and the topic has an approved answer, answer with the article, then ask: “Did that fix it, or is something still happening?”
A wrong answer sent fast costs more than one question asked first.
When deflection lowers CSAT
Deflection lowers satisfaction in three recurring ways.
- The answer is wrong or half right, and the customer gives up. They send nothing for 7 days, so the request counts as deflected while the customer is unhappy and may leave.
- The request is urgent, and the customer spends two rounds with the bot before a person sees it.
- In a shared channel, a weak bot reply lands in front of the customer’s colleagues, not only the one person who asked.
To see it, send the CSAT survey at close and split requests into three buckets: deflected and held, handed to a person, and reopened after a bot answer. Report the response count and the response rate beside each score. A bucket with 30 answers out of 400 sends reflects the people who bothered to answer, so read its comments before you act on the score. Below 30 responses, do not compare buckets at all.
Decision rule: if the deflected and held bucket trails the person-handled bucket by 0.5 points on a five-point scale for two weeks, with both buckets at 30 or more responses, take that topic off the bot. The 0.5 gap and the 30-response floor are starting settings. Tune them on your own history.
Why email comes first, and what Slack needs
Plan on the assumption that customers in shared channels self-serve less than email customers, then test it. Take the last 50 questions from one channel, paste each into your help center search, and count how many one clear article answers. If the count is low, that channel stays with people.
Slack’s Slack Connect guide says “When using Slack Connect, only the organization that owns a channel can do the following:” and lists adding other companies to the channel, removing them, and controlling who can post. Agree on any bot with the channel owner before you plan around one.
For the Slack part of the number to hold, every top-level customer message needs a request and a start time. Without that list, the Slack share of the deflection rate is a guess.
What the budget should buy
Fund the budget per real deflection, not per answer. Work out the human side from your own numbers: minutes per conversation, multiplied by loaded hourly cost, divided by 60.
Break-even works in reverse. At the $1.61 figure from the worked example, a 6-minute human conversation breaks even at a loaded cost of about $16.10 an hour. If your loaded rate is above that, a bot that closes at $1.61 per deflection is cheaper. If it is below, it is not. Check the rate on your own payroll figures, not on a benchmark.
Zendesk’s tiers article says it describes the tiers “introduced on May 18, 2026”. It defines Verified resolution as “The AI agent successfully resolved the interaction.” The sentence “This tier does not count against your resolution allowance.” appears under Contained resolution and Assisted escalation, not under Verified resolution, so check which tier your invoice counts before you build a figure on it.
The allowance article describes the platform in place before May 18, 2026. On that platform, Zendesk lists 10 automated resolutions per agent per month for Professional and Growth, 15 for Enterprise and 5 for Team. Light agents are not included in the default allocation calculation. The allocation expires annually or at the end of the subscription term, and the account chooses how it responds when it reaches the limit.
Worked example on the older allowance, as illustrative arithmetic: 12 full Professional agents get 120 allocated resolutions a month. At 300 resolutions a month, the allocation runs out after about 12 days of a 30-day month (120 divided by 300, times 30).
Before a budget review, confirm which count the vendor bills and which count your report uses. If they differ, the budget owner needs both lines.
Failure modes and fixes
- Article views counted as deflection. Fix: count a request only when no new message about the same issue arrives during the 7-day window after the bot’s last reply.
- The bot closes the thread itself. Fix: a closing message from the bot is fine, but the report reads the window, not the close, so the request counts only after 7 days with no new message from the customer.
- Vendor dashboard and your report use different windows. Fix: write the window into the report title, for example “deflected, 7-day hold.”
- Open requests included in the rate. Fix: exclude them from the denominator and show their count next to the rate.
- Slack requests missing from the denominator. Fix: create a request for every top-level customer message in a shared channel.
- A month-end spike hides a drop in deflection. Fix: compare the same weekday over the last four weeks, and the same week last month.