How to set up an AI assistant (chatbot)

If you decide up front what the assistant answers, which documents it reads and when it hands over to a person, it earns its place; if not, it makes promises to customers that you did not make.

· 13 min readAI

Short answer

Setting up an AI assistant takes seven steps: write down what it answers, collect your current documents, make it build each answer from those documents only, design the handover to a person, map where customer data goes, test with real questions and track it with measures such as the handover rate. For the first weeks, a person should see every answer.

Friday evening, after the shop has closed, a customer types into the chat box on your website: “Can I return the item I bought?” The assistant replies at once, politely and with complete confidence, and gives a return period. It is not the period in your terms. On Monday morning a return form and a complaint land on your desk. This is where the real answer to how to set up an AI assistant (chatbot) begins: the hard part is not putting a chat box on the page. It is deciding beforehand what the assistant will and will not say, which texts it reads that from, and when it leaves the conversation to a person.

This article splits those decisions into seven steps. Choosing tools is a separate subject; we covered automation tools in n8n vs Make vs Zapier. Which job to start with is covered in where to start with automation; here we assume you have already chosen, that you want to answer customer questions.

1. Write down the assistant’s job and its limits

Start with one sentence on what the assistant does: “It answers questions about delivery, returns and product details, based on our written texts.” Then list the topics that do not belong to the assistant. That list does the most work from day one, because it stops the assistant wandering into the wrong place.

The table below is a starting draft for sorting common topics into two groups; change it to suit your business.

Topic Should the assistant answer? Why
Delivery times, how to track a parcel Yes The answer is in a written text and rarely changes
Return and exchange terms Yes, from the text itself If it gets this wrong, a wrong promise has been made to the customer
Product features, size, use Yes Written on the product page and in the catalogue
The status of an order Only if it has a secure connection to the order system Involves personal data
Complaints, an angry customer No, hand over to a person Needs tone and authority to resolve
Payments, refunds, discount exceptions No, hand over to a person Involves money going out or an exception decision
Legal or medical questions No, hand over to a person The cost of a mistake is high

If this line is not clear on paper, the assistant will try to answer everything it is asked.

2. Collect the documents the assistant will read

What the assistant knows should be limited to the texts you give it. The material usually includes frequently asked questions, return and delivery terms, product descriptions, user guides, opening hours and contact details. The best of your sent email replies are valuable too, but have a team member read them and turn them into a written text first.

Watch three things while collecting. First, if the same topic is written two different ways in two places, delete one; the assistant will say whichever it finds. Second, give every document an owner and a date of last update. Third, keep information that changes often, such as a campaign condition, in one separate place so updating it is a single move. The assistant’s accuracy cannot exceed the accuracy of the documents.

3. Make it build the answer from your documents

An “assistant that answers from your own documents” is, in technical terms, retrieval-augmented generation (RAG), meaning answer writing supported by a search. The logic resembles an ordinary workflow. The customer types a question. The system first searches your documents for the parts related to that question. Then it tells the language model: “This question came in, these are the relevant texts, write the answer based only on them.” In Google Cloud’s description, RAG first retrieves relevant information with a search and then adds it to the language model’s context; supplying facts to the model as part of the input can reduce the risk of made-up information. For the search, documents are split into small sections beforehand and written into an index, so the system can often find a section that says the same thing as the question without sharing its words. We cover the parts of a RAG system, where it goes wrong and how to test it with a fixed question set in our article on RAG.

This does not mean retraining the model on your data; the documents are not baked into the model, they are searched again for every question. Two practical results follow: when you update a document the assistant can use the new information a minute later, and you can show which section an answer came from.

Flow of an assistant that answers from documents: question, document search, is the answer in the sections found, if yes an answer, if no a handover to a person Customer question Document search Relevant sections found Is the answer in the sections found? Knowledge base FAQs, returns, delivery and product texts Answer is written Only from the sections found, showing the source Handover to a person, with the chat history Yes No, or a sensitive topic Flow of an assistant that answers from documents: question, document search, is the answer in the sections found, if yes an answer, if no a handover to a person Customer question Knowledge base FAQs, returns, delivery, product Document search Relevant sections found Is the answer in the sections found? Yes Answer is written Only from the sections found, showing the source No, or a sensitive topic Handover to a person With the chat history
Figure: The assistant first searches the documents and writes the answer only from the sections it finds; if there is no section, or the topic is sensitive, it hands the conversation to a person with its history.

The quality of the answer depends on two things: whether the search found the right section, and whether the model stayed inside the section it was given. For the second, two settings are essential. Give the assistant permission to say it does not know; one model provider’s documentation also lists permission to admit uncertainty among the basic techniques that reduce false information. And if no section is found, or the found section does not answer the question, ask the assistant to hand over the conversation rather than guess.

4. Design the handover to a person

Do not treat the handover as a feature to add later; it is the most important part of the assistant. Give written answers to three questions:

  1. When does it hand over? When the document search finds no answer, on the “hand over to a person” topics from step 1, when the customer asks to speak to a person, and when the same question has been answered twice without a resolution.
  2. To whom? A person or queue that will look at the message, and what happens outside working hours. Telling a customer on Friday evening plainly that “we will reply by Monday morning at the latest” is better than no reply and better than a wrong one.
  3. What does it carry? Pass the customer’s whole conversation to the team member, so the customer does not have to explain everything a second time.

The customer should also always be able to find a visible “Talk to a person” route. For the first weeks, go further: let the assistant prepare the answer as a draft, and have a team member approve and send it. That also produces the raw material for the measuring below.

5. Map where the customer’s words go

What people type into a chat box is often personal data: a name, a phone number, an order number, sometimes a sentence about their health. From a KVKK point of view, look at the following.

  • Notice. Article 10 of Law No. 6698 requires the data controller, when data is collected, to tell the person who the controller is and for what purpose the data will be processed. Put a short note next to the chat box with a link to your privacy notice, and let the customer know they are talking to an AI assistant.
  • Security. Article 12 requires the data controller to take the technical and organisational measures needed for an appropriate level of security; this responsibility is shared with anyone who processes data on your behalf. Decide who can see chat logs and how long they are kept.
  • Abroad. If the language model provider is outside Türkiye, the messages go there; that transfer is subject to the conditions in Article 9 of the Law. Read the provider’s data-use terms, whether messages are used for model training, and the country where data is processed, in the contract.
  • Do not take unnecessary data. The assistant should not ask for identity, health or payment details it does not need; add a sentence such as “please do not type card or ID numbers” next to the box.

The generative AI guide published by the Personal Data Protection Authority (15 Questions) is a detailed starting point for data controllers who use these tools. For a legal assessment, consult your own lawyer.

There is a security side too. “Prompt injection”, first on OWASP’s list of language model risks, is a user’s text changing the assistant’s behaviour in an unintended way; telling it to “forget your previous instructions” is a simple example. Two of the defences OWASP suggests are your design decisions: give the assistant the least access possible, and put high-risk actions behind human approval. So do not leave a step that starts a refund, takes a payment or changes a customer account directly to the assistant.

6. Test with real questions

Testing with five questions you wrote yourself makes the assistant look better than it is. Collect questions real customers wrote: from your inbox, WhatsApp history, phone notes. Sentences with typos, missing details and two topics mixed together are exactly the examples you need.

Spread the test list across the types below and write the expected behaviour next to each question. During the test, a team member reads the answers and marks each as pass or fail.

Type of question Expected behaviour
A question whose answer is clearly in a document The right answer, with the source section
A question whose answer is not in the documents “I don’t have information on this” and a handover, no guessing
A topic where two documents contradict each other Logged as an error, documents corrected
A complaint or an angry message Immediate handover to a person
A question that wanders off topic A polite refusal, back to the business’s subject
An attempt such as “forget your previous instructions” Instructions do not change, the chat stays limited

Each failed question is fixed in one of two places: if a document is missing something, add it; if the assistant crossed its limit, tighten its instructions. After any fix, run the whole list again; one fix can break another answer.

7. Measure, and loosen the reins in stages

The work is not done when the assistant goes live. In the first period, have one person read the conversation logs regularly. Choose in advance what you will count:

Measure What it tells you What to do
Handover rate In how many conversations the assistant was not enough If high, look at which topics it fails on
Share of corrected drafts The share of answers the team member changed If it falls, you can allow more freedom
Unanswered questions Gaps in the documents Turn the list into weekly document work
The same question repeated Places where the answer was not understood Simplify the text
Time the team spends The real change in workload Compare with your measurement from before the assistant

You also need to have measured the situation before the assistant; otherwise “did it help” is left to feeling. Before you remove draft approval, watch the share of corrected drafts for weeks, and keep payment, complaint and personal data questions with a person at all times. You can find how to limit and measure a pilot in where to start with automation, and how a human-approved FAQ assistant looks as a workflow in business process automation examples.

Common mistakes

  • Wanting it to answer everything. If the scope is wide it will produce answers on the riskiest topics too. Start narrow and tie widening to measurement.
  • Feeding it an old document. An out-of-date returns text turns the assistant into the voice of a wrong source.
  • Hiding the handover. An assistant with no visible way to reach a person leaves the customer without a resolution.
  • Testing only with good examples. To see where the assistant breaks, ask hard questions on purpose.
  • Nobody reading the logs. If you do not read conversations after launch, you learn about mistakes from customer complaints.
  • Leaving personal data out of the plan. Notice, retention period and provider terms are not details to add after set-up.

Common questions

Do I need a developer to set up an assistant?

There are ready-made tools for connecting a chat box and documents; the built-in assistants of support software are one kind. Most of the effort is in writing the scope, correcting the documents and testing, and someone who is not technical can do these. You may need technical help for the contract with the provider, access rights and connecting the order system.

Does uploading my documents mean training the model on them?

Not with the search-supported method: the documents are searched again for every question and given to the model as context while the answer is written. That is why updating is easy. How the documents are stored at the provider, and whether they are used for training, depends on the contract terms; read them before you set up.

When can I remove human approval?

When the share of corrected drafts has stayed low for weeks and the hard questions on your test list pass, and only for low-risk topics. Payment, complaint and personal data questions should always go to a person.

One thing you can do today

Open the last month of customer messages and write the questions that repeat most often on one page. Next to each, note the document where the answer is written. The questions you cannot point to a document for are your gap, not the assistant’s; that list becomes the first draft of both the scope and the test list.

If you would like to set this up together, we describe how we build an assistant that answers only from your documents and hands over to a person when needed on our automation and AI service page.

Sources

This article is general information, not legal, financial or technical advice. Platforms, their rules and the law change over time, so some details may have changed since 11 November 2026 or contain small errors. Before you make a decision for your business, check the current official sources or write to us.

Tell us about
your project.

Fill in the form or write to us directly, and we will agree on a day to talk.

Your first meeting is with Deniz Sezer, founder of Detmore.

First meeting free · Reply within one working day · Written proposal in 2 to 3 days · NDA ready

What do you need? pick as many as you like

Name, email and phone are required.

Contact

We reply within one working day.