thirdbrain.

What's possible with AI in customer support and success

Evidence current to mid-2026.

In customer support and success, AI is ready today for the writing and the admin, like drafting replies, summarising tickets and reading customer satisfaction (CSAT) scores and review feedback at scale, always with the support agent reviewing. Refund and credit decisions, consumer guarantee conversations, complaints and vulnerable-customer handling stay with people. The tables below show exactly which is which.

Each task below sits in one of three bands: a strong fit today, worth a careful pilot, or keep with people for now.

A clear breakdown of where AI does and does not fit across the core tasks of a customer support and success team. A practical starting point, not the last word.


Where AI is mature: strong fit today

Proven and available today. AI does the bulk of the work and the support agent reviews it.

FunctionThe job todayWith AI
Drafting replies and macrosWrite and update response templates and individual replies to common queriesAI drafts from the ticket and your knowledge base; the support agent edits, personalises and sends
Ticket and conversation summarisationSummarise long threads for handover, escalation or case notesAI produces the summary and bullet actions; the support agent checks it before passing on
Knowledge base drafting and maintenanceTurn resolved tickets and product updates into articles; audit for gaps and stale contentAI drafts articles from resolved ticket patterns and flags outdated content; the support manager checks accuracy and publishes
Call and chat transcription and case notesTranscribe calls and chats and write them up as case notes with agreed actionsAI transcribes and drafts with consent; the support agent checks attribution and owns what goes in the system
Customer satisfaction (CSAT), net promoter score (NPS) and review analysisRead every comment, group the themes, report patterns to the businessAI clusters themes and drafts the summary; the support manager checks it and guards customer anonymity

Where AI is emerging: consider piloting with a human gate

Promising but not yet proven at this scale. AI assists and the support agent stays in the loop, so trial it on a contained scope first.

FunctionThe job todayWith AI
Ticket triage and routingClassify every ticket by type and urgency, assign to the right team or agentAI classifies and routes on settled patterns; the support manager checks edge cases and keeps escalation routing
FAQ deflection and self-service chatbotAnswer routine questions before a ticket is raisedAI handles the common query types from your knowledge base; the support agent handles anything involving rights, refunds or disputes
Ticket tagging and categorisationTag tickets for reporting, quality review and analysisAI auto-tags on trained patterns; the support manager spot-checks and maintains the taxonomy
Success outreach and check-in draftingDraft onboarding emails, milestone check-ins and renewal touchpointsAI drafts from account context; the customer success manager (CSM) reviews, personalises and owns the send
Account and case researchPull account history, ticket patterns and usage data before a success callAI assembles the brief; the CSM reads it and verifies before the conversation
Reporting and metricsCompile ticket volumes, resolution times, CSAT and handle time into regular reportsAI assembles the data and drafts commentary; the support manager checks the definitions and owns the narrative
Voice-of-customer synthesisAggregate themes from tickets, chats, surveys and reviews for product and leadershipAI surfaces the themes; the support manager judges signal versus noise and owns the recommendation

Where AI is not ready or suitable today: keep with people

These stay with people, either because the tooling is not reliable enough or because controls and compliance rule it out.

  • Refund, credit and goodwill decisions. The authority to grant a refund, credit or goodwill gesture stays with the support agent or team lead. AI can identify signals that a customer may be entitled; it does not decide or commit.
  • Consumer guarantee and ACL rights conversations. Telling a customer what their legal rights are under the Australian Consumer Law is your business’s legal liability. A misleading statement by an AI chatbot is a misleading statement by the business, so the support agent owns every consumer rights conversation.
  • Complaint and escalation resolution. Deciding how to resolve a formal complaint, what remedy is appropriate and whether to accept or reject the complaint, stays with the support manager. Judgement and empathy cannot be delegated.
  • Vulnerable-customer handling. Customers in financial hardship, distress, grief or crisis need a human conversation. AI can flag signals; it must not handle the conversation.
  • Autonomous outreach at scale. Letting AI generate and send success or renewal outreach without review creates Spam Act exposure, incorrect entitlement claims and brand risk. The support manager owns who is contacted and what is sent.
  • Churn commitment and save decisions. The retention rate you commit to leadership and the decision about which at-risk accounts to prioritise for a save stay with the support manager. AI surfaces risk signals; the support manager makes the call.

Worth weighing: for any of these, the upfront setup (loading your knowledge base, agreeing the tagging taxonomy, defining the scope for the chatbot) is a one-off, separate from the ongoing effort, which on the strong-fit items is mostly review. The time that comes back goes to the cases that need real attention: the escalations that need empathy, the consumer rights conversations that need human judgement, and the customers worth keeping. AI takes the drafting and the tagging; the support team stays where it matters. One practical compliance note: from 10 December 2026, automated systems that make decisions about refund eligibility, escalation priority or account status are likely to need disclosing in your privacy policy under the Privacy Act amendments; audit those decision points now rather than in December.

This guide sits on top of the things that stay the same whatever your function. See the ground rules

Questions

The questions leaders ask.

Will AI replace my support team?

No, it changes what they spend time on. The drafting, the tagging and the reading of CSAT and review feedback shrink, and their judgement, their conversations and their ability to read a situation matter more. The work that carries legal weight and the decisions that affect customer rights and remedies stay firmly with your support team. The goal is a support function that resolves more with the same people, not fewer.

Is it safe to let AI talk to our customers?

It is safe for well-defined, bounded queries where the answer is deterministic and the support manager reviews the system. Order status and password resets are a different category from refund eligibility and consumer rights. Under the Australian Consumer Law, what an AI chatbot tells your customer is a statement by your business. A confident but wrong answer about consumer guarantees, refund entitlements or service terms is potential misleading conduct, and the liability is yours, not the tool's.

Where should we start?

Pick something narrow, repetitive and easy to check. Reply drafting and ticket summarisation are the common first step: AI produces a draft, the support agent edits and sends, and the improvement is visible in the time it takes to handle a ticket. CSAT and review analysis is the other strong one, because reading hundreds of verbatim comments by hand is the real chore and the support manager still makes sense of the themes. Prove one before adding the next.

What about chatbot deflection, does it actually work?

For a narrow, well-defined set of routine queries, yes. Vendor claims of 60-80% deflection are self-reported and unaudited, and real-world averages on broader scope are lower. The failure mode matters as much as the successes: when a chatbot fails on a complex query or a consumer rights question, the customer experience is often worse than if the customer had reached a support agent. The tool earns its place as a triage aid, not a replacement for your team on anything that matters.

What can AI do in customer support now that it could not six months ago?

Quite a bit, which is why this is dated. Through early 2026 the major helpdesk platforms made their AI agents action-taking rather than answer-only: Zendesk launched resolution-based pricing at its May 2026 conference, where customers pay per confirmed resolved interaction rather than per seat. That at least aligns the vendor's incentive with your outcomes. The honest other side: several autonomous deployments were walked back after quality dropped, and one CEO publicly said cost-first thinking produced lower quality and committed to always-available human agents. The assisted model keeps proving out; the fully autonomous model keeps running into trouble on anything complicated.

How current is this, and what if the tools have moved on?

The evidence is current to mid-2026 and we refresh it as the tools and the rules change, which in support they do often. Where a finding comes from an independent source we lead with it; where it comes from a vendor we treat it as marketing. If something has shifted since you read this, the fastest way to get the current picture for your own setup is a quick chat.

Book a discovery call

A first conversation

See how this applies in your world.

Book a confidential chat. Tell us where you see the opportunity, in customer support and success or anywhere else, and we will agree a sensible next step. No deck, no pitch.