AI Agents for Customer Support: What They Can Do and What to Keep Human
Help desks now sell AI agents that resolve tickets on their own and charge per outcome. Here is what they can do, what they cost and where humans still matter.
In this article
- What can AI agents do for customer support teams?
- How do AI agents for customer support work?
- How much do AI agents for customer support cost?
- What does the Air Canada chatbot case mean for support teams?
- Do you have to tell customers they are talking to AI?
- A starter workflow: AI agents for customer support in five steps
- Which customer support tasks should AI agents leave to people?
- Data protection and security checks
- Internal support: AI agents in Slack and Teams
- Who should use AI agents for customer support?
- AI agents for customer support: FAQs
- Sources
AI agents for customer support: the short answer
AI agents for customer support are AI systems that answer customer questions, triage tickets and take simple actions, such as checking an order, inside your help desk. Vendors such as Zendesk and Intercom now charge per resolved conversation. They work best on repetitive questions with good help articles behind them, while people keep refunds, complaints and judgement calls.
AI agents for customer support have moved past the old scripted chatbot. Today’s help-desk agents can read your knowledge base, ask follow-up questions and, with permission, act in other systems. That makes them useful for small support teams. However, it also makes mistakes more costly, because an agent’s promise to a customer can bind your business.
This guide explains what support agents can realistically do, how the main help desks describe and price them, what a 2024 tribunal ruling means for your chatbot, and how to start without damaging customer trust.
It draws on publicly documented capabilities and sources checked in October 2026. TheJusGrow has not tested these products, and nothing here is legal advice.
What can AI agents do for customer support teams?
First, separate three jobs that people often blur together. For example, some tools face customers directly. Others, by contrast, help human agents work faster. A third group runs behind the scenes on routing and reporting.
| Job | What the AI agent does | Typical risk level |
|---|---|---|
| — | — | — |
| Answering customers | Replies to common questions using your help articles and policies | Medium: wrong answers reach customers directly |
| Taking actions | Looks up an order, updates an address or starts a return in connected systems | Higher: changes real records |
| Triage and routing | Tags, prioritises and sends tickets to the right person or queue | Lower: a human still handles the reply |
| Agent assist | Drafts replies, summarises long threads and suggests articles for staff | Lower: a person reviews before sending |
| Quality review | Flags conversations for review and spots repeated issues | Low: internal only |
For most small teams, the back-office jobs are the easiest place to start. Meanwhile, customer-facing answers deliver the biggest time savings but need the most preparation.
How do AI agents for customer support work?
Zendesk’s help centre describes its AI agents as AI-powered bots that talk with customers on messaging, email and voice channels. They can hold a conversation and, in Zendesk’s words, “perform actions in authorized systems autonomously”. Zendesk also describes goal-oriented “procedures”, scripted dialogues and API integrations with other systems.
Similarly, Intercom’s Fin AI Agent follows the same broad model. It answers customer questions and can run what Intercom calls Procedures, which are step-by-step workflows for more complex requests.
In practice, three ingredients decide how well AI agents for customer support perform:
- Knowledge. The agent can only answer well from accurate, current help articles and policies.
- Permissions. Each connected system, such as your shop or billing tool, needs a clear limit on what the agent may read or change.
- Handoff rules. The agent needs clear triggers for passing a conversation to a person, such as anger, legal threats or money disputes.
How much do AI agents for customer support cost?
Pricing, however, has shifted from per-seat fees towards paying for outcomes. As of October 2026, here is what the two best-documented vendors say.
| Vendor | How AI is billed | What counts | Seats |
|---|---|---|---|
| — | — | — | — |
| Intercom Fin | From $0.99 per outcome | The customer confirms the answer helped, or leaves without asking for more help | Essential $29, Advanced $85, Expert $132 per seat per month |
| Zendesk AI agents | Automated resolutions, with an allowance included in eligible plans and more available to buy | An issue resolved without live-agent intervention, checked by an AI model | AI agents need an eligible Suite or Support plan |
Intercom’s help centre adds useful detail. For instance, Fin only bills when it gives a real answer, not a reply to a greeting. In addition, if the customer returns to the same conversation for more help, Intercom says it removes the charge.
Zendesk says a resolution counts only after 72 hours of inactivity on messaging and email. It also runs an AI check to confirm that the issue was actually resolved. Zendesk moved to a new resolution-tier system in May 2026, so check its current pricing page for exact rates. Also note that Zendesk’s older “AI agents – Essential” product reaches end of availability on 10 December 2026.
Outcome pricing for AI agents in customer support sounds fair, but it has a catch. An “assumed” resolution may simply mean a customer gave up. So compare billed resolutions with your satisfaction scores and repeat contacts.
What does the Air Canada chatbot case mean for support teams?
Above all, the most cited warning comes from Canada. In Moffatt v. Air Canada, 2024 BCCRT 149, a customer asked the airline’s website chatbot about bereavement fares. The chatbot said he could claim the discount after travel, although another page on the same website said otherwise.
The British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation. According to an analysis by McCarthy Tétrault, the tribunal rejected the idea that the chatbot was a separate entity, because it was part of Air Canada’s own website.
The lesson for anyone using AI agents for customer support is simple. If your AI agent tells a customer something, customers may reasonably rely on it. Therefore, the agent’s sources must match your official policy pages, and anything involving money or exceptions should go to a person.
Do you have to tell customers they are talking to AI?
In the European Union, the AI Act’s transparency rules in Article 50 became applicable on 2 August 2026. According to a September 2026 analysis by Baker Botts, providers of AI systems that interact directly with people must design them so users know they are dealing with AI, unless that is obvious.
Admittedly, that duty falls mainly on the companies that build the systems. Even so, clear labelling is good practice for any business. A short line such as “You’re chatting with our AI assistant. Type ‘agent’ to reach a person” sets expectations and gives customers an exit. For advice on your own obligations, speak to a qualified lawyer.
A starter workflow: AI agents for customer support in five steps
Here is a cautious, step-by-step way to introduce AI agents for customer support. It reflects TheJusGrow’s editorial judgement, not a vendor’s setup guide.
Step 1: Audit your help content
First, list your top 20 questions by volume. Then check that each has an accurate, current help article. Fix contradictions first, because an agent will repeat them.
Step 2: Start with agent assist
Next, let the AI draft replies and summarise threads for your staff. People review every draft for two to four weeks. As a result, you learn where the AI goes wrong before customers see it.
Step 3: Turn on answers for a narrow topic
After that, allow the agent to answer customers directly on one low-risk topic, such as delivery times or password resets. Meanwhile, keep everything else routed to people.
Step 4: Add actions slowly
Only then connect systems that let the agent change records. In other words, begin with read-only lookups, such as order status. Add changes, such as address updates, once the read-only phase runs cleanly.
Step 5: Review weekly
Finally, read a sample of AI conversations every week. In particular, look for wrong answers, missed handoffs and customers who gave up.
Here is an example instruction you could adapt for a customer-facing agent:
Answer only from our published help articles and policies. If the answer is not there, say you are not sure and offer to connect the customer with a person. Never promise refunds, discounts, compensation or exceptions. Hand the conversation to a person immediately if the customer is upset, mentions a legal issue or asks for a manager.
Which customer support tasks should AI agents leave to people?
Some tasks should remain with people, even though AI agents for customer support could attempt them. The list below is an editorial view, not a regulatory requirement.
- Refunds, credits and compensation above a small, fixed limit.
- Complaints, cancellations from long-standing customers and anything emotionally charged.
- Legal threats, data-protection requests and safety issues.
- Policy exceptions, because an AI agent cannot weigh goodwill against cost.
- Vulnerable customers, such as people in financial difficulty or bereavement.
Data protection and security checks
Support conversations often contain names, addresses, order details and sometimes payment or health information. So review your vendor’s data-processing terms, retention settings and storage locations before launch.
Also think about attackers. AI agents for customer support face the public, and customers can type anything, including text designed to manipulate the agent. Our explainer on prompt injection explains the risk. Give the agent the narrowest permissions possible, and require human approval for any action that moves money or changes account ownership.
For a wider view of agent risks, see our guide on whether AI agents are safe.
Internal support: AI agents in Slack and Teams
Customer support also has an internal side. For example, IT help desks and HR teams answer the same questions from colleagues every week. OpenAI announced at DevDay 2026 that business users can mention @ChatGPT in Slack and Microsoft Teams channels, using connected tools. Our guide to an AI agent for Slack covers that setup in more detail.
If your support team works from a CRM, our AI agent for HubSpot guide explains how CRM-based agents handle tickets and contacts.
Who should use AI agents for customer support?
Good fit: online shops, software companies and service businesses that see the same questions repeatedly and already keep a tidy help centre. Similarly, teams with seasonal spikes may benefit, because AI handles volume without overtime.
Wait or go slowly: businesses with complex or regulated products, very low ticket volumes or poor documentation. In those cases, therefore, the setup effort and risk may outweigh savings.
Checklist before launch:
- Top questions mapped to accurate help articles.
- Clear AI disclosure and an easy route to a person.
- Handoff rules for money, complaints and legal issues.
- Read-only permissions first, with approvals for changes.
- A weekly review of AI conversations and billed resolutions.
Key takeaways
- AI agents for customer support can answer questions, triage tickets and take simple actions in connected systems.
- Intercom charges from $0.99 per Fin outcome, while Zendesk bills by automated resolutions on eligible plans.
- A 2024 Canadian tribunal held Air Canada responsible for its chatbot’s wrong advice.
- EU transparency rules for AI that talks to people have applied since 2 August 2026.
- Start with agent assist, then narrow customer-facing topics, and add actions last.
AI agents for customer support: FAQs
Not fully. Help-desk AI agents can resolve many repetitive questions, but people are still needed for complaints, refunds, exceptions, vulnerable customers and reviewing the AI’s work.
As of October 2026, Intercom lists Fin from $0.99 per outcome plus seat fees, and Zendesk bills AI agents by automated resolutions on eligible plans. Check each vendor’s pricing page for current rates.
In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable for its chatbot’s wrong advice. Rules differ by country, so ask a lawyer about your situation.
Begin with AI-drafted replies that staff review, then allow direct answers on one low-risk topic, and only later connect systems that let the agent change records.
Sources
- Zendesk Help, “About AI agents” (accessed October 2026)
- Zendesk Help, “About automated resolutions for AI agents” (accessed October 2026)
- Intercom, “Pricing” (accessed October 2026)
- Intercom Help, “Fin AI Agent resolutions” (accessed October 2026)
- McCarthy Tétrault, “Moffatt v. Air Canada: A misrepresentation by an AI chatbot” (19 February 2024)
- Baker Botts, “EU AI Act Article 50 transparency obligations go live” (10 September 2026)
- OpenAI, “DevDay 2026 Recap” (29 September 2026)
