What an AI support agent is
Older chatbots followed a script: the customer picked from a menu, and anything outside the menu led nowhere. An AI agent is different. It reads the customer's question in their own words, looks up the answer in the information you have given it, and replies in natural language. Connected to your systems, it can also do things: check an order, book a slot, open a ticket.
The important phrase is "the information you have given it". An agent that answers from your own help articles, policies and order data is useful. One that answers from general knowledge will invent details about your business.
What they do well
- Answer common questions. Opening hours, delivery times, how to reset a password, what a plan includes. If the answer is written down somewhere, an agent can find and explain it.
- Work at any hour. Customers get a reply at night, at weekends and in every time zone, without waiting for an office to open.
- Collect the details first. The agent can ask for the order number, the device and the error message before a person ever sees the ticket.
- Route requests. It can tell a billing question from a technical one and send each to the right team.
- Look things up. With a connection to your systems it can report an order's status or an appointment time.
- Help your staff. Drafting a reply for an agent to review, or summarizing a long conversation, saves time even when a person stays in charge.
What they cannot do, or do badly
- Make judgment calls. Whether to bend the refund policy for a loyal customer is a decision for a person.
- Handle an upset customer. Someone who is angry or distressed wants to be heard by a human. Send those conversations to one quickly.
- Know what you have not told them. If a policy changed last week and the help article did not, the agent will give the old answer.
- Guarantee they are right. AI models can state something false in a confident tone. Good design reduces this; nothing removes it completely.
- Act without access. An agent cannot change a booking unless it has been connected to the booking system and allowed to.
- Give regulated advice. Legal, medical and financial advice need a qualified person.
At a glance
| Give it to the agent | Keep it with a person |
|---|---|
| Questions answered in your help center | Complaints and disputes |
| Order and booking status | Exceptions to policy |
| Collecting details before handover | Anything involving safety, law or health |
| Sorting and tagging incoming requests | High-value or sensitive accounts |
| First reply outside office hours | Cases the agent has already failed to solve |
What you need before you start
- A written knowledge base. If the answers live only in your team's heads, write them down first. This is most of the work, and it improves your human support too.
- Clear policies. The agent can only apply rules that are stated plainly.
- A handover path. Every conversation needs a way to reach a person, with the history attached so the customer does not repeat themselves.
- Access to the right systems, limited to what the agent actually needs.
- An owner. One person who reviews conversations and keeps the content up to date.
How to roll one out safely
Start narrow
Choose one type of question that is frequent and low-risk, such as delivery questions. Get that working well before adding more.
Keep a person in the loop at first
Begin with the agent drafting replies for your team to approve. When the drafts are reliably good, let it send them itself for that topic.
Tell customers what they are talking to
Say clearly that it is an automated assistant and make "talk to a person" easy to find. People are more forgiving of a bot that is honest about being one.
Measure the right things
Track how many conversations the agent resolves without help, how many it hands over, and whether customers were satisfied. A high "resolved" number means little if customers gave up in frustration.
Read the transcripts
Review a sample every week. Wrong answers nearly always trace back to missing or outdated content, which you can fix.
Protect customer data
Know where conversations are stored, who can read them and whether they are used to train anyone's model. Update your privacy policy to match.
What a sensible first project looks like
A typical well-scoped first project runs like this:
- Pick the topic. Go through recent support requests and find the question type that is most frequent and least risky. Delivery and returns questions are a common choice.
- Fix the content. Rewrite the help articles for that topic so they are complete, current and unambiguous. Where staff have been answering from memory, write the answer down.
- Connect one system. Give the agent read-only access to order status, so it can answer "where is my order" with a real answer.
- Run it in draft mode. For the first weeks the agent writes replies and your team approves or edits them. Every edit shows you a gap in the content.
- Let it answer directly for that one topic once the drafts need few changes, with handover to a person always available.
- Review, then widen. After a month of data, decide which topic to add next.
This is slower than switching on a chatbot for everything on day one. It is also how you avoid a public mistake.
Common mistakes
- Launching on every topic at once. You cannot tell which answers are wrong, and customers meet the weak spots first.
- Hiding the route to a person. Customers who feel trapped by a bot become complaints.
- Treating it as finished. Products, prices and policies change. The agent's content has to change with them, and someone must own that.
- Measuring only cost saved. If satisfaction falls, the saving is not real.
- Giving it more access than it needs. An agent that only needs to read order status should not be able to change orders.
Questions to ask anyone building one for you
- Does the agent answer only from our own content, and how do you test that?
- What happens when it does not know the answer?
- How does handover to a person work, and what does our team see?
- Where is conversation data stored and who has access?
- How do we update the content ourselves, without calling a developer?
- How will we see what it is getting wrong?
Next step
An AI agent is a piece of software like any other: it needs a clear job, good inputs and someone responsible for it. If you are weighing a ready-made tool against a built one, our guide to custom software vs off-the-shelf applies here too.
Qorica builds chatbots and automations connected to your own data, for support, lead qualification and repetitive back-office tasks. See our AI agents service or tell us which questions your team answers most often.
