Chatbot, agent assist or full automation?

What are the three models for AI in customer service?

An AI customer service model is the operating set-up that defines who answers a request and what role AI plays in that answer. In practice three models stand out, with the classic rule-based bot as a fourth reference point:

• Chatbot (self-service): the customer talks to AI directly. Modern versions use RAG to find answers in the company’s own content. For a definition, see our chatbot glossary page.
• Agent assist: the customer talks to a person. In the background, AI offers the human agent a draft reply, the relevant procedure and a summary of the customer’s history.
• Action-taking agent: AI does not just answer; it checks order status, changes a booking or opens a return. An AI agent reaches your systems through tools and permissions.

Anthropic’s article on agents names customer support as a good fit for agents: it combines conversation with actions such as pulling order history or issuing refunds, and success can be measured by whether the user’s issue is resolved. The same article warns that autonomy raises costs and the risk of compounding errors.

Aspect	Rule-based bot	AI chatbot (RAG)	Agent assist	Action-taking agent
Who answers the customer	Bot, from a fixed script	AI	Human agent	AI, which also acts
Main risk	Dead end on off-script questions	Inventing answers not in the sources	Agent sends a draft without checking	Wrong or unauthorised action
Data needed	Decision tree	Current FAQ, policies, product content	Content plus customer history	Content plus system access (CRM, orders)
Human role	Maintains the script	Takes over, fills content gaps	Approves every reply	Approves risky actions, audits logs
Suitable requests	Simple menu-driven flows	Information questions (delivery, returns, opening hours)	Questions touching account or contract	Clear-rule, reversible actions

Which requests suit self-service, and which do not?

Sort your own support history along two axes: how often a question repeats and what a wrong answer costs.

Suited to self-service: frequent questions whose answer is in a written policy and does not vary by customer. Delivery times, return terms, opening hours, payment options, check-in and check-out times. In ecommerce and retail and in tourism and hospitality, these questions repeat constantly and often arrive through messaging channels such as WhatsApp Business.

Suited to agent assist: questions where the answer is specific to the customer or a wrong answer has financial consequences. Policy coverage, loan application status, invoice disputes. In finance and insurance, AI drafts these replies and a person sends them.

Stays with people: complaints, legal notices, health or safety matters, angry or vulnerable customers. Here AI’s job is at most to route quickly to the right person.

How do you write hand-off rules?

Hand-off is the most neglected part of chatbot projects. Write the rules up front and test them:

1. Topic rule: hand off directly when a complaint, refund amount, legal wording or another defined topic appears.
2. Confidence rule: if the assistant cannot find the answer in the content, it says “I don’t know” and hands off. It does not guess.
3. Customer rule: hand off when the customer asks for a person, repeats the same question or signals frustration.
4. Context transfer: the conversation summary and customer identity pass to the agent, so the customer does not have to explain again.
5. Out of hours: if nobody is available, tell the customer clearly when they will hear back and log the request.

How do you tell customers they are talking to AI?

Article 50 of the EU AI Act requires AI systems intended to interact directly with people to be designed so that those people are informed they are interacting with an AI system. There is an exception where this is obvious to a reasonably well-informed, observant and circumspect person. The article text states that the obligation applies from 2 August 2026.

The practical implementation is simple: a clear line at the start of the conversation, such as “You are chatting with an AI assistant. You can ask for a person at any time,” and an assistant that does not pretend to be human. Confirm the scope with your legal counsel. Under GDPR, and KVKK (Türkiye’s data protection law) where Turkish customers are involved, your privacy notice should also cover how chat transcripts are stored.

How do you measure quality?

Instead of a made-up success percentage, track these metric types on your own data:

• Resolution rate: conversations closed without hand-off. Read it together with whether the customer comes back on the same topic shortly after.
• Hand-off rate and reason: which topics get handed off shows gaps in your content.
• Wrong-answer log: every wrong answer flagged by an agent or customer is logged with its cause.
• Acceptance rate for agent assist: how many drafts are sent unchanged, edited or discarded.

Two frameworks help with risk. NIST’s Generative AI Profile names confabulation, confidently presented false content, as a distinct risk; it is the problem usually called AI hallucination. The OWASP 2025 list covers risks such as prompt injection and excessive agency, which come into play directly once an agent can take actions.

Where should you start?

Sort the last few months of support tickets into the three groups above. For self-service questions, make sure the content is current and in one place, and limit the chatbot to that content. Try agent assist on a queue where wrong answers are expensive. Add an action-taking agent last, and only with reversible actions; document checks during customer onboarding or application review backed by document analysis are good starting points.

We use the same approach in our own product. Promptor is an AI chat and sales assistant we built for WordPress that knows the site’s content. In client projects we write down hand-off rules, disclosure wording and a measurement plan during design. For details see AI customer service, AI chatbot development, RAG enterprise knowledge assistant and, for monitoring in production, managed services.