AI customer service agents: what they do and how to deploy one.
An AI customer service agent is software that reads a customer's message, looks up the answer in your own systems and takes the action the customer asked for — checking an order, rebooking an appointment, opening a return — then hands the conversation to a person when it should. What to give it, what to keep human, the guards it needs, the US rules to know and how to roll one out.
AI customer service agent
- Does
- Answers and acts in your systems
- Acts through
- Tools your application runs
- Keep human
- High-impact actions and exceptions
- Must have
- A clear route to a person
- US voice calls
- AI voices fall under the TCPA
Answers versus actions: what makes it an agent
A support chatbot answers questions. An AI customer service agent also does things: it looks up the order, changes the booking, opens the ticket. The difference is not a smarter model but access to tools. In Anthropic's documentation, for example, the model decides when a tool is needed and returns a structured call — the tool's name and its inputs — that your application runs; your code then sends the result back and the model writes the reply from it.
That split matters for support. The agent never touches your order system directly: it asks, and your code decides what it is allowed to do, with which customer's data, and logs it. Everything that follows — permissions, approvals, handover — is built in that layer. For how to pick a team to build it, see our guide to choosing an AI chatbot development company.
What to give the agent, and what to keep human
A starting split for a typical support inbox. Move a task to the left only once the agent has handled it correctly on your real conversations.
| Request | Agent alone | Agent drafts, person approves | Person only |
|---|---|---|---|
| Opening hours, policies, how-to questions | Answer from your own content | — | — |
| Where is my order? | Look up the status for the verified customer | — | Lost or damaged parcel |
| Book, move or cancel an appointment | Within the rules you set | Outside those rules | — |
| Refund or credit | Collect the details, open the request | Issue the refund | Disputes |
| Account or billing change | Explain the steps | Make the change | Fraud or identity doubts |
| Complaint, angry or vulnerable customer | Acknowledge and hand over | — | Handle the conversation |
One support request, end to end
A customer of an online shop writes on WhatsApp: “My order hasn't arrived and I need it for Saturday. Can you send a replacement?”
- 01
Identify
The agent matches the WhatsApp number to a customer record. No match: it asks for the order e-mail before showing anything.
- 02
Look up
It calls the order-status tool. Your code returns only that customer's orders, not the whole database.
- 03
Answer
It gives the carrier status and the tracking link in plain words, from the data it received — not from memory.
- 04
Prepare the action
A replacement costs money, so it is a high-impact action: the agent drafts the request with the order and the reason, and does not send anything itself.
- 05
Hand over
The draft lands with a person on your team, who approves or changes it. The agent tells the customer someone is on it, and the person picks up the same conversation with its history.
The guards an agent that acts needs
The OWASP Top 10 for LLM Applications (2025) names the risk as LLM06, Excessive Agency: damaging actions taken because of an unexpected or manipulated model output. It traces it to three causes — excessive functionality, excessive permissions and excessive autonomy — and lists the counter-measures that map directly onto a support agent:
- Give it only the tools it needs. A support agent that reads orders does not need a tool that deletes them.
- Least privilege, per customer. Each call runs with the rights of the conversation's customer, not an admin account.
- A person approves high-impact actions. Refunds, credits, account changes: the agent drafts, someone signs off.
- Monitor and rate-limit. Log every tool call, watch the downstream systems, and cap how fast the agent can act so a fault cannot spread.
US rules to know before launch
Voice. In a declaratory ruling of February 2024, the FCC confirmed that the Telephone Consumer Protection Act's restrictions on “artificial or prerecorded voice” cover current AI technologies that generate human voices. A call that uses one therefore needs the prior express consent of the called party, unless it is an emergency or an exemption applies. This concerns calls your agent places; check outbound voice campaigns with counsel before launch.
WhatsApp. The WhatsApp Business Messaging Policy lets a business use automation when it replies within the 24-hour customer service window, which opens and resets with each customer message, but requires prompt, clear and direct escalation paths: a transfer to a human agent in the chat, a phone number, e-mail, web support, a store or a support form. Outside that window, only approved templates can be sent. Meta's terms on general-purpose AI assistants are covered in our article on WhatsApp bots for business.
Deploy one, and measure it, in five steps
- 01
Export real conversations
Take a sample from your inbox, chat or call notes and sort them by request type. That sample becomes your test set.
- 02
Pick one channel and a few request types
Start with questions answered from your content and read-only lookups — the “Agent alone” column of the table above.
- 03
Run it in dry-run first
The agent says what it would reply and do, a person checks it, and nothing reaches a customer until the answers hold up.
- 04
Go live with handover on
Every conversation can reach a person, and the agent hands over by itself on complaints and on anything outside its scope.
- 05
Measure on your own numbers
Before launch, note your current response times and workload. After, track how many conversations end without a person, how many answers your team corrected, and how many actions were reversed. Add a tool only when those hold.
How we build them
Our WhatsApp AI agents run on Meta's official platform with a daily usage cap, a dry-run mode, an off switch, human handover and a log of every action — the same guards as in the list above, described with the rest of our AI agents and automation work. For example, our WhatsApp and voice-note agent for gyms in East Africa, tested on a real phone, lets members book and ask questions by text or voice note; it hands the conversation to staff when needed and alerts the owner.
Sources (checked 25 October 2026)
- Anthropic — Tool use with Claude — the model returns a structured tool call, the application runs it and sends back the result
- OWASP — LLM06:2025 Excessive Agency — excessive functionality, permissions and autonomy; human approval, least privilege, monitoring, rate limiting
- FCC — Declaratory Ruling on AI-generated voices in robocalls (February 2024) — TCPA restrictions on artificial or prerecorded voice cover AI-generated human voices; prior express consent
- WhatsApp Business — Business Messaging Policy — automation within the 24-hour customer service window, escalation paths to a human, templates outside the window
Related
Frequently asked questions
What is an AI customer service agent?
Software that reads a customer's message, looks up the answer in the business's own systems and takes the requested action — checking an order, rebooking an appointment, opening a return — and hands the conversation to a person when it should. Unlike a chatbot that only answers, it acts through tools that the business's application runs and controls.
Which support tasks should an AI agent not do alone?
High-impact actions such as refunds, credits and account changes. The OWASP Top 10 for LLM Applications (2025) recommends requiring a human to approve high-impact actions, giving the agent only the tools it needs and running them with the least privilege.
Do AI voice agents need consent in the US?
For calls they place, yes in most cases. In February 2024 the FCC confirmed that the TCPA's restrictions on artificial or prerecorded voice cover AI-generated human voices, so such calls need the called party's prior express consent unless an emergency purpose or an exemption applies.
Can an AI agent answer customers on WhatsApp?
Yes, within the rules of the WhatsApp Business Messaging Policy: automation is allowed when replying within the 24-hour customer service window, as long as there are prompt, clear and direct escalation paths to a human, such as a transfer in the chat, a phone number, e-mail, web support, a store or a support form.
Which requests would you hand to an agent first?
Bring a sample of your real support conversations and we will tell you honestly what an agent can take on, and what should stay with your team.