Customer-Facing AI Agents: Guardrails Before Go-Live

Set the guardrails before a customer-facing AI agent talks to anyone: what it may say and do, when it must hand off to a person, and how you will judge it. In our XRHealth case study, a custom chatbot answered common questions about pricing and technical requirements and passed high-intent prospects to live agents, as one part of a funnel rebuild that cut cost per lead from $150 to $10.

In the EU, Article 50 of the AI Act has also required, since August 2, 2026, that AI systems that talk directly with people be designed to tell them they are dealing with AI, unless that is obvious. Most of the work that decides whether an agent builds trust or costs you trust happens before launch, in five areas: scope, handoff, disclosure, compliance, and measurement.

What Should a Customer-Facing AI Agent Be Allowed to Do?

Start with a written scope that holds on every channel the agent runs on, such as website chat, WhatsApp and other messaging apps, and in-app support. We recommend three lists, agreed by the teams that own the customer relationship:

• Topics it may answer. Product questions, pricing and plan options, how-to steps, opening hours, and order or account status once the customer is verified. Every answer should come from content you have approved, such as your help center, pricing pages, and policies. When no approved source covers a question, the agent says so and offers a person.
• Actions it may take. Reasonable first actions include booking a call, creating a support ticket, or updating a contact record. Give the agent read-only access to customer and account data at first, and write access only for the few actions on this list. Widen its permissions one job at a time. Refunds, cancellations, account changes, and anything that moves money stay with people. Revisit that only with sign-off from the team that owns the risk, which for a regulated firm means compliance.
• Things it never does. Promise outcomes, make exceptions to policy, argue with an upset customer, or guess. Regulated topics have their own rules, covered in the compliance questions below.

Brand voice comes after scope. Write it as rules with examples: how formal the agent is, which words it avoids, how it apologizes, and how it says “I don’t know.” A distinctive voice does not make up for a wrong answer, so test accuracy before tone.

The same rules apply when an agent works inside your CRM rather than in a chat window. We cover that groundwork, including why an agent should start in read-only mode, in our post on AI agents inside your CRM.

When Should an AI Agent Hand Off to a Person?

Whenever the customer asks, and whenever the conversation reaches something the agent is not cleared to handle. Write these down as explicit escalation rules, so the handoff never depends on the agent’s own judgment:

1. The customer asks for a person. Hand off on the first request, without trying to keep the conversation with the agent.
2. The customer shows high intent. Define the signals in advance, such as a pricing request, a request to book, or a completed qualification question, and route those prospects to sales while they are still engaged.
3. Money, access, or a complaint is involved. Refunds, failed payments, withdrawals, disputes, locked accounts, and formal complaints go to a person.
4. The topic is regulated. Anything that could read as a personal investment recommendation or as medical advice goes to a trained person who follows your firm’s rules on what may be said. If your firm is not authorized to give investment advice, that person declines too.
5. The agent is stuck. Two failed attempts at the same question, a customer rephrasing the same request, or no approved source for the answer each trigger a handoff.
6. The customer seems distressed or vulnerable. Hand off to a trained person, and give the agent a fixed, approved message for those moments.

A handoff is only as good as what arrives with it. Pass the transcript and any answers already collected, so the customer does not have to repeat themselves, and route every handoff to a named person or queue. Outside staffed hours, the agent should say when a person will reply instead of ending the conversation.

Our XRHealth case study shows the high-intent rule at work. A custom chatbot handled common questions about pricing and technical requirements and used logic-based escalation to pass high-intent prospects to live agents. The chatbot was one part of a wider funnel rebuild, and the case study credits the biggest gain to no longer forcing people to book a call right away. The same handoff rule applies whether a bot follows a script or runs on AI. We tell the full XRHealth funnel story, and cover first response, qualification, and routing for new leads, in our post on AI sales assistants and new leads.

Do You Have to Tell Customers They Are Talking to AI?

In the EU, the agent must tell them, unless that is already obvious. The legal duty sits with the agent’s provider, which may be your company, as explained below. Article 50(1) of the AI Act requires providers to design AI systems that interact directly with people so that those people are informed they are interacting with an AI system. The exception is when this is obvious to a reasonably well-informed, observant, and circumspect person, given the circumstances and context of use. Article 50(5) requires the information to be clear and distinguishable and to be given at the latest at the first interaction. Under Article 113, these rules have applied since August 2, 2026.

The 2026 amendments to the AI Act did not move that date. The Digital Omnibus on AI (Regulation (EU) 2026/1744), in force since July 27, 2026, pushed back several high-risk deadlines and left the duty to tell people they are dealing with AI as it was. It also added a grace period for a second duty, in Article 50(2): providers of AI systems that generate synthetic audio, image, video, or text content must mark that output in a machine-readable format, so it can be detected as AI-generated. Systems placed on the market before August 2, 2026, have until December 2, 2026, to comply. Systems placed on the market on or after August 2, 2026, have no grace period.

Who counts as the provider matters. Under Article 3(3), a provider includes a company that has an AI system developed and puts it into service under its own name or trademark, so a brand that launches its own agent may hold that role. If your agent writes its replies with generative AI and you are its provider, ask whether Article 50(2) also requires those replies to be marked, and whether your model vendor already handles it. Whatever your role, we recommend disclosing in every market: say in the first message that the customer is talking to an AI assistant, and keep a visible way to reach a person throughout the conversation.

Which Questions Should Regulated Firms Take to Compliance?

If you are a broker or a healthtech company, take these questions to your compliance team before launch. They are meant to start that conversation, and the answers come from your compliance team.

For Brokers and Other Investment Firms

• Is your firm authorized to give investment advice at all? Could any answer about a product, leverage, or a specific trade count as a personal recommendation? For EU investment firms, MiFID II defines investment advice as the provision of personal recommendations to a client (Article 4(1)(4)). Where is the line, and what does the agent say when a customer asks it to cross it?
• Do the agent’s answers meet the standard for your other client communications? MiFID II requires all information that an investment firm addresses to clients or potential clients, including marketing communications, to be fair, clear, and not misleading (Article 24(3)). Who approves the answer library, and how often is it reviewed?
• When the agent discusses CFDs or other leveraged products, does it carry the same risk warnings as your other client communications? MiFID II requires information on financial instruments to include appropriate guidance on and warnings of the risks of investing in them (Article 24(4)(b)).
• If the agent helps people sign up, which fields may it ask for, and which identity, anti-money-laundering, and appropriateness questions must stay in your regulated onboarding flow? We cover how to separate marketing fields from verification data in our broker onboarding post, Fewer Fields, Faster KYC.
• Which conversations must be recorded and kept, for how long, and where?
• Does the agent know which products and offers it may mention to a visitor from each country you serve?

For Healthtech Companies

• Could an answer be read as medical advice, a diagnosis, or a treatment recommendation? What does the agent say at that line?
• What health data may the agent collect, on what legal basis, and where is it stored? Under the GDPR, processing data concerning health is prohibited unless one of the exceptions in Article 9(2) applies, such as the person’s explicit consent.
• Which questions about insurance coverage and eligibility may the agent answer, and which go to a person?
• What does the agent do if someone describes an emergency?

For Every Sector

• Does the disclosure meet Article 50, and what do your markets outside the EU require?
• If your agent generates text and you are its provider, how are its replies marked as AI-generated under Article 50(2), and does your model vendor already handle this?
• Who reviews transcripts, how often, and which named person can switch the agent off the same day?

This is not legal advice. Rules differ by market and by license, and your compliance team decides what applies to your firm. Alchemy Avenue is a marketing agency, not a broker, and does not provide investment advice.

How Do You Measure Whether a Customer-Facing AI Agent Works?

Track three numbers from the first day, and read them together. The definitions below are the ones we use. Check how your chat tool defines each one, because many tools count an abandoned conversation as contained.

• Containment rate: the share of conversations the agent resolves without handing off to a person. We count a conversation as contained only if the customer does not come back about the same issue within a set window, such as seven days, and we report abandoned conversations separately.
• Escalation rate: the share of conversations handed to a person, split by the rule that triggered the handoff. A rising share of “agent is stuck” handoffs points to gaps in your approved content. A rising share of high-intent handoffs can be good news.
• CSAT delta: the customer satisfaction (CSAT) score for conversations the agent resolved, minus the score for conversations people handled on the same contact reasons over the same period. Compare like for like, because people also handle the harder conversations the agent hands off. If people handled too few of those conversations, compare against your pre-launch score for the same reasons. Report conversations that the agent started and a person finished as a third group. A negative delta shows where the agent is costing you goodwill.

Containment on its own rewards an agent that makes handoff hard, so never set it as a target without the other two. Then connect the agent to downstream results, such as qualified leads, cost per lead, repeat contacts, and churn. In our XRHealth case study, tests were judged on downstream customer quality rather than raw lead volume, and the same standard should apply to an agent.

What Should You Check Before Go-Live?

1. The scope is written down: topics, actions, and things the agent never does, with an approved source behind every answer.
2. The escalation rules are written and tested end to end, including outside staffed hours.
3. The first message discloses the AI, a way to reach a person stays visible, and, if Article 50(2) applies, generated replies are marked.
4. Your compliance team has answered the questions above, in writing, for each market you serve.
5. Containment rate, escalation rate, and CSAT delta are tracked from day one, and a named owner reviews transcripts every week.
6. A named person can switch the agent off the same day.

If you are planning an agent like this, our AI-Powered Solutions service covers chatbots for lead engagement and customer service, and AI ticket routing.

Sources: Regulation (EU) 2024/1689 (the EU AI Act), Articles 3(3), 50, and 113, Official Journal of July 12, 2024, EUR-Lex. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of July 24, 2026, EUR-Lex. Directive 2014/65/EU (MiFID II), Articles 4(1)(4), 24(3), and 24(4)(b), EUR-Lex. Regulation (EU) 2016/679 (the GDPR), Article 9, EUR-Lex. Our XRHealth case study, linked above. All accessed October 23, 2026.

Updated October 23, 2026: we rewrote this post, re-checked its facts, and added sources.