The 3 Types of AI Agents in Customer Service (and What Each Level Takes)
The same two words describe very different products. A founder's take on the three levels of AI agents for customer service, from answering questions to issuing refunds, and what each one demands from your integrations, security and team.
Tommi Koirikivi
Every support vendor sells an “AI agent” now. We do too. The problem is that the same two words describe products with almost nothing in common. One answers questions from a help center. Another issues refunds straight from your billing system. Both carry the same label on the website.
I think about it as three levels. Each level includes the one below it, and each one asks for a very different amount of integration work, security thinking and readiness from the company running it. If you’re comparing AI agents for customer service, the first question isn’t which vendor. It’s which level you’re buying, and which level your company is ready to run.
Level 1
Answers
Reads your help center and documents
- Integrations
- Security
- Maturity
Read only
Level 2
Skills
Works the ticket: fields, notes, routing
- Integrations
- Security
- Maturity
Writes in the helpdesk
Level 3
Actions
Refunds, account changes, record updates
- Integrations
- Security
- Maturity
Reads and writes in your backend
To make the difference concrete, I’ll run one customer message through all three levels:
I was charged twice for my subscription this month. Can I get a refund?
Level 1: The AI agent that answers
A level 1 agent works from a fixed body of knowledge: your help center, plus whatever documents you feed it. A question comes in, the agent finds the relevant content and writes an answer.
In essence, this is a glorified chatbot. I don’t mean that as an insult. A lot of support volume is questions that already have a written answer somewhere, and a good level 1 agent gives that answer instantly, in the customer’s language, at three in the morning.
But look at what happens to our double charge.
The answer is correct, and the customer still has the problem. A level 1 agent can explain your refund policy. It can’t see the charge, and it can’t do anything about it.
What it takes: a knowledge base worth reading. Integration is light and the risk is low, because the agent can only talk. The worst case is a wrong answer. Quality depends almost entirely on your content, which is why I’d tell most teams to fix their knowledge base before they evaluate anything else.
Level 2: The AI agent with skills
Level 2 adds skills on top of the knowledge. A skill is a customizable behavior tied to a scenario: when the customer wants this, do that. The agent stops treating every conversation the same way and starts working the ticket like a good triage person would.
In practice that means it can:
- Recognize what the customer is trying to do, and pick the right scenario
- Enrich the ticket with the details a human would otherwise have to ask for
- Add a private note with a summary and a suggested next step
- Assign the ticket and route it to the right team with the right priority
These are simple, tool-based actions, and they all happen inside the support system.
The customer still waits for a human. But that human opens a ticket that is categorized, summarized and sitting in the right queue, and nobody asks the customer to repeat themselves.
What it takes: someone has to design the scenarios. You need a clear taxonomy, clear ownership between teams, and written rules for what goes where. Companies that never documented their routing logic find that out at level 2. The stakes are still modest, since a misrouted ticket or a clumsy private note is easy to undo. But you now have to review what the agent does, and not only what it says.
Level 3: The AI agent that acts
Level 3 is where the agent performs real actions in your backend: refunds, account-level queries, updating customer records, changing customer information. It has read and write access to the systems where your customer data lives.
This is the first level where the customer’s problem is gone, with no ticket and no human involved. It’s also what most people picture when they hear “AI agent”, and what a lot of demos imply.
What it takes: almost everything.
- Integrations. The agent needs API access to billing, accounts and orders. If your backend has no clean way to issue a refund programmatically, the agent can’t issue one either.
- Identity. Before the agent touches an account, it has to know it’s talking to the account’s owner.
- Boundaries. Which actions, up to what amount, how often, for which customers. A duplicate charge under a set limit gets refunded automatically. Anything above it goes to a human.
- Audit. Every action is logged, attributable and, where possible, reversible.
- Ownership. Someone is accountable for the agent’s permissions, the same way someone is accountable for a new hire’s.
The technology is rarely the blocker here. The models are capable enough. What holds companies back is that write access to customer data is a security decision, and most companies haven’t decided how far they trust software to make judgment calls with money.
What each level requires
| Level 1: Answers | Level 2: Skills | Level 3: Actions | |
|---|---|---|---|
| Works with | Help center and documents | Plus ticket fields, notes and queues | Plus billing, accounts and backend records |
| Access | Reads content | Writes inside the support tool | Reads and writes in your backend |
| Integration work | Connect a knowledge base | Configure skills and routing | An API integration per system |
| Worst case | A wrong answer | A misrouted ticket | A wrong refund or a changed record |
| Guardrails | Content quality and tone | Scenario rules and review | Identity checks, limits, approvals, audit log |
| Who’s involved | Support | Support operations | Support, engineering, security and finance |
Read the last row twice. The jump from level 2 to level 3 turns a support project into a company project.
AI agent vs chatbot
This is the question I hear most, and the levels answer it. A chatbot follows a script or retrieves an answer. An agent decides what to do, and then does it.
By that definition, level 1 is a chatbot with much better language skills. Level 2 is where the agent label starts to be earned, because the system chooses between actions. Level 3 is where it’s fully true.
So when a vendor says “AI agent”, ask one thing: what can it change? If the honest answer is “nothing, but it answers really well”, you’re looking at level 1. That can be exactly the right product for you. Just know that’s what it is.
Which level do you need?
My take is that the levels are a path, and you shouldn’t skip steps.
- Get level 1 right first. Every level builds on the one below. A level 3 agent with bad knowledge will confidently take the wrong action.
- Move to level 2 when you know your top contact reasons. If you can name the ten things customers ask for most, you can write the scenarios.
- Enter level 3 one action at a time. Pick a high-volume, low-risk action, set a tight limit, watch it, then widen it. Read access before write access. Small refunds before large ones.
The higher the level, the more it takes from your integrations and from your organization. The limit on your AI agent is rarely the AI. It’s how much of your business is reachable through an API, and how clearly you’ve written down the rules.
Where AI agents are going
On the technical side, the gap between the levels is closing. Open standards like Model Context Protocol make it much cheaper to connect an agent to the systems it needs, so level 3 will stop being a custom engineering project.
The trust side will take longer. I expect level 3 to become normal the same way it did for human agents: scoped permissions, approval thresholds and audit trails. You already run these controls for people. You’ll run them for software too.
The companies that get there first will be the ones that did the boring work: clean knowledge, documented rules, and a backend the agent can reach.
That’s the work we care about at Theymes. Our AI agent runs decoupled from the ticketing system, so your team only sees the conversations that need a human, whichever level you’re running. If you want to talk through which level your setup is ready for, get in touch.