Clear limits, clean sources, handover to a person and measurement: seven steps for launching a support assistant in a small business.
An AI assistant that answers customers can take dozens of repetitive questions off your team’s plate. It can also promise something the company does not offer, answer a delicate question wrongly or leave an angry customer going round in circles. Reputation is lost faster than it is earned, so the launch matters more than the technology. Below is a simple step-by-step method for a small company.
Why it is not only a technical risk
In 2024 a tribunal in Canada (Moffatt v. Air Canada) decided that an airline was responsible for wrong information given by the chatbot on its website; the defence that “the chatbot is a separate entity” was rejected. The case is not a precedent for Romania and the ruling has its limits, but it illustrates a sound idea: what your assistant says on the company’s behalf, the company answers for. The more its statements concern prices, deadlines, warranties or conditions, the greater the risk.
Step 1: decide what it may do
Start with a short, clear list:
- May answer: opening hours, locations, how to order, how to return an item, product descriptions, usage steps.
- Answers only with a reference to a document: warranties, delivery conditions, prices. Ideally it cites its source.
- May not decide: discounts, exceptions, refunds, any new promise, medical or legal advice.
- Always hands over to a person: complaints, threats, sensitive topics, requests concerning personal data.
A stricter list at the start is better than a broad one. You widen it once you see how it goes.
Step 2: feed it from controlled sources
A good assistant answers from the company’s documents, not from a model’s memory. That means having current documents: the returns policy, the price list, the FAQ. If documents contradict each other or are out of date, the assistant will inherit the chaos. We explain how this works technically in the article on RAG.
A good way of working: one person in the company who updates the sources and knows where to fix things when an answer is wrong.
Step 3: say clearly that it is an AI assistant
Customers deserve to know who they are talking to. Besides common sense, the EU Artificial Intelligence Act (AI Act) requires, in Article 50, that people be informed when they interact directly with an AI system, unless it is obvious from the context, an obligation that applies from 2 August 2026. Put a simple welcome message in from the start saying it is an automated assistant and showing how to reach a person. For your specific case, check with a lawyer.
If the assistant handles conversations containing personal data, GDPR rules apply too: inform the person, do not ask for unnecessary data, decide how long conversations are kept.
Step 4: the handover to a person, designed in advance
This is where trust is won or lost. The handover must be easy and quick, and the person should receive the conversation already summarised, so the customer does not have to repeat everything. Useful rules:
- automatic handover when the customer asks for a person, for any reason;
- handover when the assistant cannot find the answer in its sources (better “I do not know” than an invented answer);
- handover after a second failed attempt or when the tone turns tense;
- real hours and a real response time: if people only reply during working hours, say so.
An assistant that does not admit it does not know is more dangerous than one that hands over often.
Step 5: test before launch
Before showing it to customers, test it with dozens of real questions, gathered from past emails and messages, including the awkward ones: off-topic questions, prices, requests for exceptions, attempts to trick it into promising something. Note what it answered well, what it answered wrongly and what it should have handed over. Repeat after every major change to the sources.
Step 6: launch gradually
First version: the assistant proposes the answer, a person approves it. Second: it answers alone within a narrow category of questions, and the rest go to people. Third: you widen it. Each stage lasts as long as you need to see real data. Keep a stop button: if something goes wrong, you should be able to take it away from customers within minutes.
Step 7: measure what matters
Not just “how many conversations it held”. Track:
- what share of conversations were resolved without a person, then how many actually solved the customer’s problem (check by sampling, reading conversations);
- how often it handed over to a person and why;
- wrong answers found in sampling and their cause (outdated source, uncovered question, misinterpretation);
- customer complaints about the assistant;
- how fast people respond to handed-over conversations.
Decide from the start which number shows it is working and which number makes you stop it. Without such criteria, the discussion becomes opinion.
Common mistakes
- launching with outdated or contradictory sources;
- giving the assistant permission to make commercial promises;
- a handover to a person that is hard to find;
- nobody reading conversations after launch;
- measuring only savings, not customer satisfaction.
Instead of a conclusion
A good support assistant is more of a discipline than a technology: clear limits, clean sources, easy handover and continuing attention. If you do that, it gives your people time for the cases that matter.
ME Digital provides AI implementation, including support assistants with limits and handover to a person configured together with you; you can start from the contact page.