AI agents for customer support
Answer the routine, hand off the rest.
A support agent that reads the inbox, answers the questions your help centre already covers, and leaves refunds, complaints and anything unclear to a person. Every reply is on the record.
The job
The questions repeat, and the answers already exist.
Most support email asks something your help centre answers. A person still reads each one, finds the article and writes the reply. It is the most common first job teams give an agent, and the one where an overconfident agent does the most damage: a promised refund, or a help article sent to someone who is upset.
How it runs
Split the inbox in two.
The agent sorts each message. Questions your documents answer get a reply. Refunds, billing, complaints and anything it is unsure about are flagged for a person, or wait for one: any tool can require an approval before it runs.
- Grounded answers. Upload your help articles and policies as documents, and the agent answers from them.
- A line it does not cross. Sending money, changing an account or replying to a complaint can each require a person's approval.
- Tested before it goes live. Test cases for the messages that would hurt, run against every change to the prompt.
- Read what it did. Every run keeps the message, the steps, the tools it called and the reply it sent.
From our own inbox
What makes it safe to trust with more.
Every reply the agent sends should be on the record: the email it read, the article it found, what it sent, and why it left the others alone.
Step by step
From an inbox to a supervised agent.
Install the template
Inbound Support Triage, or start from your own prompt.Upload what it answers from
Help articles, policies, the product's FAQ.Gate the sensitive tools
Sending to a customer, refunds, account changes.Read the first runs
Every message, the steps it took and the reply it drafted.
Questions
What teams ask first.
Does every reply need an approval?
What does it answer from?
Can I see why it replied the way it did?
Start from a template
Agents for this job, ready to install and adapt.
- EmailInbound Support TriageReads the inbox, classifies inbound support emails, auto-replies to how-to questions, flags bugs and feature requests for manual follow-up.Connect an appGmail
- Help & SupportCustomer Refund HandlerFor inbound refund requests under a configurable threshold, drafts the Stripe refund + customer confirmation email and waits for one-click human approval before executing.Advanced setupGmailStripe
- Help & SupportStatus Page ResponderWatches the public status pages of your critical vendors. When a degraded or major incident appears, posts a single notice to your #incidents Slack channel naming the affected services so the team isn't surprised by user reports.Advanced setupSlack
- Help & SupportGitHub Issue ResponderFor open-source repos: replies to common new issues (env / repro requests / duplicate links / RTFM cases), labels them for maintainers, and skips anything that needs human judgment.Advanced setupGitHub
- Help & SupportWhatsApp After-Hours SupportResponds to incoming WhatsApp customer messages outside business hours using a pinned knowledge base. Answers what it can; escalates everything else with a next-business-day promise and an internal email alert.Advanced setupWhatsAppGmail
What it runs on
The parts of AgentOS this job leans on.
Go deeper
How it works, in detail
From the blog
Guide · 30 August 2026How to build a support inbox agent that knows when to stop
A worked example of a support agent that answers the routine questions, leaves the sensitive ones to a person, and shows its work.
Guide · 26 September 2026How to add human approval to AI agents
Which agent actions need a person's sign-off, how an approval gate should work, and how to keep the queue from turning into a bottleneck.
Pilot
Try it on one real workflow.
We build the agent with you, with an approval gate on the risky step, and report what it did.
Integrations
18 apps, your models, your APIs.
Connect GitHub, Gmail, Slack and more once, describe any HTTP API as a tool, and choose the model per agent.
Give one job to an agent this week.
Start from a template, gate the sensitive step behind your approval, and read the first run end to end.