The AI chatbot maturity scale sorts customer-facing bots into four levels: static FAQ bots (scripted buttons), interactive assistants (understand free text and hand off to people), transactional agents (take actions inside your systems) and autonomous agents (plan and act on their own). Gartner estimates that only about 130 of the thousands of vendors marketing “agentic AI” deliver real agentic capability, so the level a vendor claims and the level it ships often differ. Ollo Agent works at Level 2: it answers from your knowledge base and hands a structured request to your team.

Almost every chatbot sold in 2026 is called an “AI agent.” Gartner has a name for this: agent washing, the rebranding of chatbots, assistants and automation tools as agentic without the capability behind the label. In a June 2025 forecast, the firm estimated that only around 130 of the thousands of agentic AI vendors are the real thing, and predicted that over 40% of agentic AI projects will be canceled by the end of 2027.

For a business comparing tools, that creates a practical problem. A scripted FAQ bot, an AI assistant and an autonomous agent can sit behind the same demo video and the same pricing page. The difference shows up after you sign: when a customer asks something off-script, when the bot is supposed to book a slot, or when it should have passed the chat to a person and didn’t.

This article lays out a four-level scale you can use to place any customer-service bot, including ours. Each level is defined by three things: what the bot can do on its own, what it needs from you, and how it typically fails.

What is the AI chatbot maturity scale?

The maturity scale ranks chatbots by how much they can do without a human: from following a script (Level 1), to understanding and routing conversations (Level 2), to completing tasks in other systems (Level 3), to pursuing goals with little supervision (Level 4). Each level adds a capability; none of them is “better” in the abstract. The right level depends on what your customers ask and what your systems allow.

Level

 Name What it does What it needs from you

Typical failure

1

 Static FAQ bot Shows buttons and scripted answers; matches keywords A decision tree for every expected question

Dead end the moment a customer types something off-script

2

 Interactive assistant Understands free text, answers from your knowledge base, asks follow-ups, qualifies and hands off A maintained knowledge base and clear handoff rules

Confident wrong answer if the knowledge base is stale or handoff rules are vague

3

 Transactional agent Reads and writes to other systems: books into a live calendar, checks an order, changes a booking Systems with working APIs, clean data, defined permissions

Wrong action in a real system (double booking, wrong refund)

4

 Autonomous agent Plans multi-step work toward a goal, picks its own tools, handles exceptions with minimal approval Mature data, monitoring, risk controls, budget for oversight

Costly, hard-to-audit mistakes; projects stall before value appears

The scale is cumulative. A real Level 3 agent still needs everything Level 2 does well, because it has to understand the request before it can act on it.

What does a Level 1 static FAQ bot actually do?

A Level 1 bot follows a script. It recognizes button clicks or keywords and returns pre-written answers, which makes it predictable and cheap, and useless the moment a customer leaves the script.

Most “chatbots” from the 2015–2022 era live here: a welcome message, three buttons (“Prices”, “Opening hours”, “Talk to us”), and a keyword list behind a text box. They work well for a narrow, stable set of questions where customers are happy to tap rather than type.

The failure is easy to spot. Ask a Level 1 bot “do you have anything for two adults and a toddler in late May, under 2,000 euros?” and it either shows the generic prices menu or says it didn’t understand. Messenger customers write in full sentences, often with several questions in one message. A script can’t keep up with that.

What changes at Level 2, the interactive assistant?

A Level 2 assistant understands free text, answers from the business’s own knowledge base, asks clarifying questions and decides when to hand the conversation to a person. It reads information; it does not change anything in your other systems.

This is where large language models made the biggest practical difference for small businesses. The assistant can take that messy “two adults and a toddler” message, pull the relevant options from your knowledge base, ask the one missing detail (departure city, say), and pass the whole conversation to a manager as a structured request.

Three things separate a good Level 2 assistant from a weak one:

  • Grounding. Answers come from your documents, price lists and policies, not from the model’s general knowledge. When the answer isn’t there, the assistant says so.
  • Qualification. It collects the details your team needs before a person steps in: what, when, how many, budget, contact.
  • Handoff. It knows its limits. Complaints, unusual requests and anything involving money or personal circumstances go to a human, with context, so the customer doesn’t repeat themselves.

The typical Level 2 failure is not a dramatic one. It’s an outdated price in the knowledge base, or a handoff rule so loose that the assistant keeps chatting when it should have escalated.

What makes a Level 3 transactional agent different?

A Level 3 agent completes tasks in other systems on the customer’s behalf: it checks real availability, writes a booking into the calendar, looks up an order, or changes a reservation, within rules you define.

The jump from Level 2 to Level 3 is not about smarter conversation. It’s about access. To book a slot, the agent needs read and write access to your calendar, a way to handle conflicts, and a reliable way to confirm, reschedule and cancel. To answer “where is my order?”, it needs your order system and your carrier’s tracking data. Every action needs permissions, logs and an undo path.

This is also where marketing claims drift furthest from the product. A common pattern: a bot described as offering “booking” turns out to send the customer a link to a separate scheduling page. That can be a perfectly good design. It just isn’t a transactional agent, and the difference matters when a buyer expects the bot to fill the calendar by itself.

Level 3 makes sense when three conditions hold: the task is high-volume and repetitive, the underlying system has a working API, and a wrong action is cheap to reverse. If any of the three is missing, a strong Level 2 assistant with a fast human handoff usually delivers more for less risk.

Is a Level 4 autonomous agent real yet?

Level 4 is mostly a forecast, not a product category a small business can buy off the shelf today. Analysts expect it, but early deployments show how expensive autonomy is to get right.

Gartner’s March 2025 prediction is the headline case for Level 4: by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention, cutting operational costs by 30%. The examples Gartner gives are telling: canceling a membership, negotiating a shipping rate. Those are actions taken on the customer’s behalf, not answers.

The same firm’s June 2025 forecast is the counterweight. It attributes the expected cancellation of over 40% of agentic AI projects to escalating costs, unclear business value and inadequate risk controls, and notes that many use cases don’t need an agentic approach at all.

Klarna is the best-known real-world example of the gap. In 2024 the company said its AI assistant was doing the work of 700 customer service agents. By May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company was hiring human agents again, because the AI-first approach had produced lower-quality service and customers should always be able to reach a person.

How can you tell which level a vendor is really selling?

Ask questions that force a vendor to describe mechanisms, not outcomes. A demo shows the best conversation; these questions show what happens in the other ones.

  1. “Show me the bot answering a question that isn’t in the knowledge base.” A Level 1 bot fails. A good Level 2 assistant admits it doesn’t know and offers a person. A risky one invents an answer.
  2. “When the bot books something, which system does the booking land in, and who can see it?” If the answer is “the customer gets a link,” you are looking at Level 2 with a handoff, not Level 3. That may be fine. Just price it as what it is.
  3. “What happens when two customers try to book the same slot?” Real transactional agents have an answer involving live availability and conflict handling. Link-based flows push the problem to your team.
  4. “How do you define a resolved conversation?” Definitions vary widely. Intercom’s Fin, for example, counts a conversation as resolved if the customer confirms the answer helped or simply stops replying for 24 hours, per a third-party summary of Intercom’s documentation (Gleap, 2026).
  5. “What resolution rate do your customers actually see?” Vendor-published figures are a useful reality check. Freshworks publishes a 23–30% resolution rate for its Freddy AI Agent, and Tidio’s top-tier “Resolution Rate Guarantee” commits to 50% for its Lyro AI, according to third-party pricing analyses from eesel AI and Chatarmin (both August 2026).
  6. “How does a conversation reach a human, and what does the human see?” At every level, the handoff is where customers are won or lost.

Ollo Agent’s take: We expect “agent” to stop meaning anything in buyer conversations within the next year, the way “AI-powered” already has. The buyers who do well will be the ones who ask where an action lands and who checks it, not whether the product is agentic. For most service businesses, a Level 2 assistant that answers well and escalates fast will beat a Level 3 claim that quietly turns out to be a link.

Which level does a small service business actually need?

Most small service businesses get the most value from a well-built Level 2 assistant, because their bottleneck is answering and qualifying inquiries fast, not executing transactions. Work through these steps to check where you fit:

  1. List your top 20 inbound questions. If most are about prices, availability, policies and “can you do X?”, that’s Level 2 territory.
  2. Mark which ones end in an action (a booking, a change, a refund). Then check whether the system behind that action has an API your vendor actually connects to.
  3. Estimate the cost of a wrong action. A double-booked haircut is annoying. A wrong refund or a missed medical appointment is not. Higher stakes favor a human confirmation step.
  4. Look at when inquiries arrive. If many come in evenings and weekends, the immediate win is answering and qualifying them then, whatever level you choose.
  5. Start one level lower than the vendor suggests. It’s easier to add actions to an assistant that already answers well than to fix an agent that acts on misunderstood requests.

Where does Ollo Agent sit on the scale?

Ollo Agent is a Level 2 interactive assistant with structured handoff. It answers customers across messengers and the website from your knowledge base, qualifies the request, and passes a ready-to-act lead card to your team.

We’d rather say this plainly than claim a level we don’t ship. Ollo does not offer native booking integration today: it qualifies the request in conversation, then hands a structured request to the business’s own booking page or team. That design is deliberate. It keeps a person in charge of every action that touches money or a calendar, and it lets Ollo launch in about 24 hours, without connecting to your internal systems first.

At Level 2, the numbers come from answering well and escalating at the right moment. In conversations handled by Ollo Agent, 60–65% of initial inquiries are resolved without a person stepping in (Ollo Agent internal data, 50+ clients, 2025–2026). For one travel agency client, ALF, that meant +27% bookings on the same traffic, −35% manager workload, and 65% of first inquiries handled without human involvement. The customer support assistant covers Instagram, WhatsApp, Telegram, Facebook Messenger and website chat in one place, with automatic escalation to your team.

Source: Ollo Agent internal data, ALF client, 2025–2026.

See a Level 2 assistant answer and hand off your real inquiries: book a demo →
In the demo, we run your own common questions through Ollo and show exactly what your team receives