AI customer service tools: the five types, and which ones you need.
AI customer service tools fall into five types: chatbots that answer customers, AI agents that also act in your systems, agent-assist tools that draft for your team, AI features inside your help desk or CRM, and conversation analytics. Most support teams need two or three of them working together, not one of each. What each type does, when it is worth it, how they fit on one request, the US rules to check, and when to buy rather than build.
AI customer service tools
- Customer-facing
- Chatbots and AI agents
- Team-facing
- Agent assist, help desk AI, analytics
- Key difference
- Answering versus acting in your systems
- Must keep
- A clear route to a person
- US law
- No AI exemption, says the FTC
The five types of AI customer service tools at a glance
Vendors bundle these differently and give them different names. Sort what you are offered by what it does, not by what it is called.
| Type | What it does | Who sees it | Worth it when | Watch out for |
|---|---|---|---|---|
| Chatbot | Answers questions from your help content, in chat or messaging | Customers | Many questions already have a written answer | Loops with no way out; answers that drift from your policies |
| AI agent | Answers and also acts: looks up an order, rebooks, opens a return | Customers | Requests are repetitive and your systems have usable APIs | Too many permissions; actions nobody approves |
| Agent assist | Drafts replies, summarises the thread, suggests the help article | Your support team | Answers need judgement but follow known patterns | Drafts sent unread; suggestions from stale content |
| Help desk or CRM AI | Tags, routes and prioritises tickets, fills customer fields | Your support team | Volume is high and requests land in the wrong queue | Wrong routing that nobody notices |
| Conversation analytics | Groups conversations by topic and sentiment, flags trends | Managers | You need to know why customers contact you | Dashboards that nobody acts on |
Chatbot or agent: the line that matters
Most of what is sold as AI customer service sits on one side of a single line: does the tool only answer, or can it also act? A chatbot reads your help content and replies. An agent is given tools — check an order, move a booking, open a refund request — and uses them. In Anthropic's documentation, for example, the model returns a structured call naming the tool and its inputs, and your application runs it and sends back the result. Whoever builds that layer decides what the agent may touch.
That is why the same request can be cheap to automate with one type and risky with another. Answering “what is your returns policy?” needs only good content. Processing the return needs access to your order system, rules on who can do what, and a person signing off on anything that costs money. What to hand an agent and what to keep human is covered in our guide to AI customer service agents.
One request, four tools
A customer of an online shop writes on WhatsApp: “I moved last week. Can you send my order to my new address?” Here is how the types share the work.
- 01
Help desk AI routes it
The message is tagged “change of address” and linked to the customer record that matches the phone number.
- 02
The agent checks what it can do
It calls the order-status tool. The parcel has not shipped yet, so changing the address is allowed by your rules — the agent asks the customer to confirm the new address and updates the order.
- 03
Agent assist takes the exception
Had the parcel already shipped, the agent would not change anything. It hands the conversation to a person, with a drafted reply and the carrier's redirection options in the summary.
- 04
Analytics spots the pattern
At the end of the month, address changes after dispatch show up as a recurring topic. That is a checkout fix — asking customers to confirm their address — not a support problem.
When to buy a packaged tool, and when to build
Most teams do both: a packaged help desk with its built-in AI, plus a custom agent where the work depends on their own systems.
| Your situation | Lean towards buying | Lean towards building |
|---|---|---|
| Where your customers write | Web chat and e-mail, already covered by your help desk | WhatsApp or voice notes, in several languages |
| What the AI must do | Answer from your help centre | Act in your own booking, order or policy systems |
| Where your data lives | In the help desk or a CRM it connects to natively | In in-house or sector software with its own API |
| Who should own the rules | The vendor's default workflows fit | Your approval steps and limits are specific to your business |
| What you need to see | Standard reports are enough | A log of every action, with a dry-run mode and an off switch |
US rules and evidence to check before you sign
The law applies to the bot as it does to your staff. In its June 2023 issue spotlight on chatbots in consumer finance, the Consumer Financial Protection Bureau found that each of the top 10 commercial banks uses chatbots and that about 37% of the US population is estimated to have interacted with a bank's chatbot in 2022. It described customers stuck in “doom loops” with no offramp to a human, and stated that chatbots must comply with all applicable federal consumer financial laws, and that companies may be liable when they do not. The report is about banking, but the failure it describes — a bot that blocks the way to a person — is the one to test for in any sector.
So do vendors' claims. In September 2024 the Federal Trade Commission announced Operation AI Comply, a sweep against deceptive AI claims; its chair stated that “there is no AI exemption from the laws on the books.” One case concerned DoNotPay, which marketed “the world's first robot lawyer” without testing that claim; the order included $193,000. For a buyer, the lesson is practical: ask a vendor to show its tool on a sample of your own conversations rather than relying on a headline resolution rate.
Messaging platforms set their own conditions. The WhatsApp Business Messaging Policy, last updated 23 September 2026, allows automation when a business replies within the 24-hour customer service window, as long as there are prompt, clear and direct escalation paths: an in-chat transfer to a human agent, a phone number, e-mail, web support, a store visit or a support form. Any tool you put on WhatsApp has to support one of them. Meta's limits on general-purpose AI assistants are covered in our article on WhatsApp bots.
Choose your tools in five steps
- 01
Sort a sample of real conversations
Export a few weeks of tickets, chats and messages and group them by what the customer wanted: an answer, an action, or a judgement call.
- 02
Match each group to a type
Answers to a chatbot or help-centre search, actions to an agent, judgement calls to agent assist. Routing and analytics serve all three.
- 03
Check what your current help desk already does
Many help desks and CRMs include AI features. Use those before adding a separate tool that duplicates them.
- 04
Test on your data, not a demo
Run the shortlisted tool on your sample in a mode where nothing reaches a customer, and have your team mark every wrong answer and every action it should not have taken.
- 05
Launch with the exit built in
Every channel offers a person in one step, conversations arrive with their history, and you track corrected answers and reversed actions before widening the scope.
Where we fit
We build the custom side of that stack: WhatsApp AI agents that act in a business's own systems, on Meta's official platform, with a dry-run mode, an off switch, a log of every action and handover to staff. One example from our work is a booking and guest platform for a restaurant group in South Africa, where guests book online or on WhatsApp through an AI booking assistant and every guest's history sits in one CRM — described in our restaurant booking platform case. The rest of the scope is on our AI agents and automation page; if you are comparing outside teams, see how to choose an AI chatbot development company.
Sources (checked 26 October 2026)
- CFPB — Chatbots in consumer finance (issue spotlight, June 2023) — top 10 commercial banks all use chatbots; about 37% of the US population in 2022; doom loops; chatbots must comply with federal consumer financial law
- FTC — FTC announces crackdown on deceptive AI claims and schemes (September 2024) — Operation AI Comply; “no AI exemption from the laws on the books”; DoNotPay order with $193,000
- WhatsApp Business — Business Messaging Policy — last updated 23 September 2026: automation within the 24-hour window, prompt, clear and direct escalation paths
- Anthropic — Tool use with Claude — the model returns a structured tool call that the application runs, then receives the result
Related
Frequently asked questions
What are AI customer service tools?
Software that uses AI to handle or support customer service work. They fall into five types: chatbots that answer customers, AI agents that also act in the business's systems, agent-assist tools that draft replies for the support team, AI features inside a help desk or CRM that tag and route tickets, and conversation analytics that show why customers get in touch.
What is the difference between an AI chatbot and an AI agent in customer service?
A chatbot answers questions from your content. An AI agent also takes actions — checking an order, moving a booking, opening a return — through tools that the business's own application runs and controls. That access is what makes an agent more useful, and why it needs limits on what it may do and a person approving high-impact actions.
Should I buy an AI customer service tool or build one?
Buy when customers write through web chat and e-mail already covered by your help desk and the AI only needs to answer from your help centre. Build, or customise, when customers use channels such as WhatsApp or voice notes, or when the AI must act in your own booking, order or policy systems with approval rules specific to your business. Many teams combine both.
Are there US rules for AI in customer service?
Existing laws apply. The FTC stated in September 2024 that there is no AI exemption from the laws on the books, and the CFPB wrote in June 2023 that chatbots must comply with all applicable federal consumer financial laws, warning about customers stuck in loops without a route to a human. Channels add their own rules: WhatsApp allows automation only with clear escalation paths to a person.
Which of your requests need an answer, and which need an action?
Bring a sample of your real support conversations and we will tell you honestly which tools fit, what your help desk already covers, and where a custom agent would help.