How to start using AI agents in customer support safely
A step-by-step guide for owners and support leads: start narrow, use approved content, limit permissions, disclose the AI, protect data and widen only after testing.
- Author
Umair Mateen- Published
- Reading time
- 6 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 4 understood

Start small and stay in control. Let an AI agent answer a narrow set of common questions from your own approved help content, give it read-only access first, make it say it is an AI, and route anything involving money, accounts or complaints to a person. Widen what it can do only after you have read its real conversations and the numbers look right.
The reason for the caution is simple: your customers will treat whatever the agent says as your company speaking. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible after its website chatbot gave a passenger wrong information about a bereavement fare refund. The airline argued the chatbot was responsible for its own actions; the tribunal disagreed and awarded the passenger roughly C$650 plus interest and fees.
Step 1: Pick a narrow, low-risk starting point
List the questions your team answers most often and sort them by risk. Good first candidates have one correct answer that already sits in your help centre: opening hours, delivery times, how to reset a password, where to find an invoice. Leave out anything where a wrong answer costs money or trust: refunds, cancellations, medical or financial advice, legal complaints and anything about a specific person's account. Write that list down; it becomes the agent's job description.
Quick check · Myth or fact
AI agents: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 3 checked
From the explainer What is an AI agent? Explained with an assistant you send on errands
Step 2: Give it only approved knowledge
The agent should answer from your own help articles and policies, not from what the underlying model happens to "know". Models can produce confident answers that are wrong, known as AI hallucinations. Clean up your help content before launch: remove outdated pages, fix contradictions, and add the policy wording for the questions you chose. If a policy is not written down, the agent should say it does not know and hand over.
Step 3: Limit what it can do, not just what it says
Answering questions and taking actions are different risks. An agent that can issue refunds, change addresses or cancel orders needs much tighter control. The OWASP security project lists "excessive agency" among the top risks for AI applications and traces it to three causes: tools the agent does not need, permissions broader than the task, and high-impact actions that nobody checks. Its advice translates into plain rules:
- Fewest tools. Connect only the systems the in-scope questions need.
- Read before write. Start with read-only access, for example looking up an order status, not editing it.
- Approval for big actions. A person confirms refunds, cancellations and account changes. This is human in the loop.
- Checks outside the AI. Your order or billing system should enforce its own limits, so the AI cannot talk its way past them.
This matters because customers can type anything, including instructions designed to trick the agent. That is called prompt injection, and tight permissions limit the damage if it works.
Step 4: Tell customers it is an AI and keep a clear way out
If you serve people in the EU, Article 50 of the EU AI Act has applied since 2 August 2026. According to the European Commission's guidance, AI systems that talk directly with people must be built so those people are told, clearly and from the first interaction, that they are dealing with an AI, unless that is obvious. The legal duty sits mainly with the provider of the system, so check that your vendor's widget shows the notice, or add one yourself if you built the agent. Always offer an easy route to a person, such as "talk to someone", and pass the full conversation along so the customer does not repeat themselves.
Step 5: Protect personal data
Support chats contain names, emails, order details and sometimes health or payment information. Before launch, ask the vendor where conversations are stored, how long they are kept, whether they are used to train models, and who on their side can see them. Tell the agent never to ask for card numbers, passwords or medical details in chat. If you are in healthcare or finance, check the vendor's contract terms against your own rules before any real customer data flows through it.
Step 6: Test, pilot, then widen
Write 30 to 50 real questions from past tickets, including awkward ones, and check the answers before going live. Then switch the agent on for a small slice of traffic, a single channel or business hours only. Have someone read a sample of conversations every week, track how often customers ask for a person, and keep a log of every action the agent took. The US standards body NIST publishes a free, voluntary Generative AI Profile of its AI Risk Management Framework that you can use as a fuller checklist as you grow.
What it costs and how long it takes
Pricing varies by vendor and model. As one example, Intercom's help page for its Fin AI agent lists $0.99 per resolution, counted when the customer confirms the answer helped or leaves without asking for more help. Other tools charge per seat, per conversation or by usage of the underlying AI model, so compare them on your own monthly ticket volume. Budget staff time too: cleaning up help content, writing test questions and reviewing conversations each week. How long a rollout takes depends mostly on the state of your help content and how many systems the agent must connect to, so ask vendors for a written plan with stages rather than a single launch date.
Common mistakes
- Launching on every question and every channel at once.
- Giving the agent write access to orders or billing before it has proven itself on read-only work.
- Letting it answer from general model knowledge instead of your approved content.
- Hiding the human option to push up the automation rate.
- Measuring only cost per ticket and never reading the conversations.
- Assuming the vendor carries the blame for wrong answers.
A checklist you can use today
- List your 20 most common support questions and mark each low or high risk.
- Pick the low-risk ones that already have a correct, written answer.
- Find and fix outdated or conflicting help articles on those topics.
- Decide which systems the agent may read, and which actions need a person's approval.
- Check that the chat clearly says it is an AI and has a visible "talk to a person" option.
- Ask your vendor where chat data is stored, for how long, and whether it trains models.
- Collect 30 to 50 past tickets as a test set.
- Name one person who reads a sample of AI conversations every week.
If you would like help scoping, building or reviewing an AI support agent with approvals and audit trails, The DevDrift can help: see our AI services.
Sources
- OWASP: LLM06:2025 Excessive Agency
- European Commission: Transparency obligations under Article 50 AI Act (FAQ)
- American Bar Association, Business Law Today: BC tribunal confirms companies remain liable for information provided by AI chatbot
- NIST AI 600-1: AI Risk Management Framework, Generative AI Profile
- Intercom Help Center: Fin AI Agent outcomes and pricing
You’ve read the news. Now understand it.
- Learn · 5 min · Myth or factWhat is an AI agent? Explained with an assistant you send on errandsAn AI agent takes a goal, plans the steps and carries them out with software tools, checking each result as it goes. Explained with an assistant you send on errands.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat are AI hallucinations? Explained with a confident quiz guessAn AI hallucination is a confident answer that is false or made up. Explained with a quiz team that writes down a neat, sure-looking guess when it doesn't know.Start in 5 seconds →
- Learn · 5 min · What would you do?What is human-in-the-loop? Explained with a manager signing chequesHuman-in-the-loop means AI prepares the work but a person approves the important steps before they happen. Explained with a manager who signs every cheque before it goes out.Start in 5 seconds →
- Learn · 5 min · Spot the weak pointWhat is prompt injection? Explained with a forged note in the in-trayPrompt injection is when hidden instructions inside an email, web page or file trick an AI model into following an attacker instead of you. Explained with a forged note slipped into an assistant's in-tray.Start in 5 seconds →


