Customer support · Agent workflow

Audit support ticket deflection before you automate it

ZeroTwo reads recent Intercom conversations, checks Notion help content for coverage, sorts repeated topics into safe, fix-first and keep-with-a-person, and posts the result to Slack for approval. Nothing about your automation changes until a person says so.

The output is a reviewed audit, not a bot. You decide what gets automated, and what success means.

  • Intercom
  • Notion
  • Slack

Three lanes for every repeated topic

Example topics
  • Safe to deflect

    Repeated, low-risk, clear answer

    • Where to download an invoice
    • Connecting a calendar
    • Exporting a report
  • Fix the article first

    Customers ask because the answer is stale or hard to find

    • Setup steps from an older flow
    • Article written in internal terms
  • Keep with a person

    Judgment, identity or an open defect

    • Charged twice
    • Locked out after a phone change
    • Export fails on large files

Ticket deflection fails when teams automate before they audit

Two things go wrong. The team picks topics from volume alone, and the answer customers get is only as good as the article behind it.

  • Volume is not a verdict

    A queue can hold hundreds of repeated questions, and still most should not be automated. Some are simple setup questions. Others involve a billing exception, a security concern, a broken feature, or a customer who already tried the documented answer.

  • A weak article becomes a weak bot

    People ask the same thing because the article is missing, written in internal language, or stale after the workflow changed. Automating on top of that does not fix the answer. It hides the gap and sends the customer around again.

Choosing what to automate
Without an audit. Leaders guess from dashboard volume and anecdotes from agents.
With the audit. Repeated tickets are grouped, examples are linked, and each topic shows why it is or is not safe.
Knowledge coverage
Without an audit. Articles are checked by hand after the queue has already grown.
With the audit. Notion pages and drafts are compared with how customers actually phrase the problem.
Risky topics
Without an audit. Automation rules can bury urgent, billing or account issues behind a generic answer.
With the audit. High-risk topics are excluded by policy and low-confidence items wait for Slack review.
Follow-through
Without an audit. Insights end up in spreadsheets and one-off threads.
With the audit. Approved article fixes, routing rules and owners sit in one reviewable queue.

What a sampled audit table should show

Every topic gets the same five facts: how often it appears, what a customer actually wrote, whether the help content covers it, a verdict with the reason, and the proposed fix. If a row cannot be filled from source tickets, it is not ready to approve.

Illustrative support deflection audit: five example topics with a representative ticket, knowledge coverage, a verdict and a proposed fix
TopicRepresentative ticketTickets in sampleKnowledge coverageVerdictProposed fix
Find or download an invoice“Where do I get last month’s invoice as a PDF?”[count]Covered, but hard to findCandidatePut the invoice location in the article title and link it from the billing help page.
Set up a chat integration“I approved access but nothing shows up in the channel.”[count]Stale: written for an older flowFix the article firstRewrite the steps against the current flow and retest with a fresh account before approving.
Refund for a double charge“I was billed twice this month.”[count]Not applicableExclude: billing exceptionKeep with a person. Add a routing rule and an escalation brief instead of an article.
Locked out after changing phones“My new phone does not receive the login code.”[count]PartialExclude: account accessRoute to a person who can verify identity. Review the wording of the existing article only.
Export fails on large projects“The export spins and never finishes on my biggest project.”[count]Missing: probable open defectExclude: unresolved defectSend to product feedback triage. Write nothing customer-facing until the cause is known.

Illustrative layout only. The topics and tickets are invented and the counts are blank on purpose. This is not output from a real support queue, and it is not a result.

What the audit cannot tell you

A last-500 sample is a starting point for a decision, not proof. State these limits next to every recommendation.

  • A sample is not the whole queue

    The last 500 conversations may cover one busy week. A launch, an outage or a billing cycle can dominate it. Record the date range, and rerun on a different window before you approve anything.

  • It only contains people who wrote in

    Customers who found the article, or gave up, never opened a ticket. The audit sees demand for answers, not every confusion.

  • Tags and clusters can be wrong

    Grouping starts from conversation text, and agent tags shape what the agent sees. Spot-check each cluster against its source tickets before trusting the label.

  • One conversation cannot show a recontact

    Whether the customer came back needs a later look, which is why the guardrails below are measured after a change, not in the audit.

Topics that stay with a person

Excluded by policy, with the reason written next to each one. The strongest candidates are the boring ones: repeated, low-risk and answered by an article someone has tested.

  • Billing exceptions

    Refunds, disputed charges and credits need account context and authority.

  • Account access

    Identity checks cannot be handled by a help article.

  • Security concerns

    A suspected compromise should reach a person immediately.

  • Angry customers

    A canned answer to a frustrated person usually escalates the situation.

  • Outages and unresolved defects

    The honest answer changes hourly, or does not exist yet.

  • Legal or policy judgment

    The answer depends on interpretation, not on a documented step.

How ZeroTwo audits tickets for safe deflection

It runs weekly before your support operations review, or when a ticket tag crosses a volume threshold you set. Slack carries the review queue.

A starting prompt

It asks for the date range, the reasons for each exclusion and a stop at the queue. Adjust the sample size and the exclusion list to your own policy.

Every Friday, review the last 500 Intercom support conversations and state the date range they cover.

1. Group repeated questions by customer problem. For each group give the number of conversations and two paraphrased examples with ticket links.
2. Compare each group with our Notion help content. Mark coverage as clear, stale, missing or hard to find.
3. Exclude billing exceptions, account access, security, angry customers, outages, legal or policy judgment and unresolved product defects. State the reason for each exclusion.
4. For the safe candidates, draft the article fix and a short answer snippet.

Post a Slack queue with approve, revise, assign and reject for each item. Do not change any automation rule.
  1. Read recent conversations

    Review conversations, tags, ticket state, customer wording and repeated agent replies, so the audit starts from real support work.

  2. Check what the help content already says

    Search Notion pages, runbooks and help center drafts for a matching answer, then mark coverage as clear, stale, missing or hard to find.

  3. Separate candidates from exclusions

    Group repeated problems, set aside anything on your exclusion list with a stated reason, and keep low-risk questions with clear answers.

  4. Draft the article fix

    Turn support examples into plain-language updates, suggested answer snippets, open caveats and an owner note.

  5. Post the approval queue

    Send support leaders a Slack queue with source tickets, the knowledge gap, the proposed edit, risk flags, and approve, revise, assign or reject.

Define success before the first change

Ticket reduction alone can look like a win while customers write back, escalate or give up. Write each measure and its rollback trigger down before anything goes live. The thresholds are yours to set.

Five measures for judging a deflection change, with how to define each, why it matters and a rollback trigger to set in advance
MeasureDefine it asWhy it mattersRollback trigger
ResolutionThe customer did not contact support again about the same topic within a window you choose.A closed ticket is not a solved problem.Resolution drops below a baseline you wrote down before launch.
RecontactsShare of customers who return on the topic after seeing the article or automation.Fewer tickets can just mean the same customer is stuck in a loop.Recontacts rise above the baseline for a period you set in advance.
EscalationShare of affected conversations that still end up with a person.Ticket counts hide the cases that bounce back to an agent.Escalations rise on the topic, or a new risky subtopic appears.
SatisfactionYour own rating, where you collect one, for the affected topic against before the change.Volume can fall while the experience gets worse.Satisfaction on the topic falls below the baseline you set.
Ticket volumeConversations on the topic. Read it last.On its own, it can make a harmful change look like a win.Not a trigger by itself. Only meaningful while the four above hold.

Start with one reviewed topic

Once a topic is approved, the response workflow lives elsewhere. For an ecommerce store, that is Shopify customer support, which owns the individual reply. This page decides whether a topic is ready for it.

  1. Approve one topic

    Pick a single low-risk topic whose article a person has read and tested. Write down the baseline for each guardrail.

  2. Pilot with a reviewer

    Expose it to a narrow slice, such as one channel or one segment, and have a named person read outcomes on a fixed rhythm.

  3. Widen or roll back

    Compare against the baseline and your rollback criteria. If any trigger fires, revert the change and keep the Slack thread as the record.

Where this stops and the next workflow starts

This page owns the audit and the approve-or-hold decision. Browse every workflow in agent use cases, or see how the Work agent and scheduled tasks run them.

Questions about auditing support deflection

What is an AI support ticket deflection agent?

An AI support ticket deflection agent reviews support conversations to find questions that can be answered safely through self-serve content or automation. In ZeroTwo, the workflow checks Intercom tickets against Notion help content and sends a Slack review queue before any recommendation is approved.

How is this different from a support chatbot?

A chatbot answers customers in the moment. A deflection audit helps the support team decide what should be automated at all. It reviews ticket evidence, checks article coverage, flags risk, and asks humans to approve article updates or routing changes before customers are affected.

Can ZeroTwo identify knowledge base gaps from tickets?

Yes. ZeroTwo can compare repeated customer questions with existing Notion content and mark whether the answer is missing, stale, unclear or hard to find. The workflow can draft updates, but support owners should review them before publishing.

Which tickets should not be deflected automatically?

Tickets involving account access, billing exceptions, security concerns, angry customers, legal or policy judgment, outages and unclear product defects should usually stay with human agents. Ask for these as explicit exclusions, with a stated reason, instead of treating every repeated topic as an automation candidate.

How does Slack approval fit into the workflow?

Slack gives support leaders a review surface for each recommendation. The message can include source tickets, proposed article copy, confidence notes, risk flags, and actions to approve, revise, assign or reject the candidate.

How do we know deflection helped and did not just hide tickets?

Define success before launch. Track whether customers came back on the same topic, how many still reached a person, and satisfaction where you collect it, and read ticket volume last. Write a rollback trigger for each measure and start with one reviewed topic.

What are the limits of auditing the last 500 conversations?

It is a sample. It may cover one unusual week, only includes customers who wrote in, and depends on how tickets were tagged. Record the date range, spot-check clusters against source tickets, and repeat the audit on another window before approving a change.

What should support operations review before approving deflection?

Review the source tickets, whether the proposed answer fully resolves the issue, whether the article is readable by customers, which segments are affected, and whether the topic could hide urgent or sensitive issues. Approval should mean the team is comfortable exposing the answer to customers.

Find the support questions that are actually safe to deflect

100 free credits to start on the Free plan. Work agent access starts on Plus at $14.99/mo. See pricing.

Get started with ZeroTwo