Conversational AI greeting

A conversational AI greeting is the first thing an AI agent says when a conversation opens. It usually runs a sentence or two. In that space it has to tell customers they're talking to an AI, hint at what the agent can help with, and get them to describe their problem.

It's also the one turn a team fully controls. The greeting is written ahead of time and runs on every conversation, so one that sets the wrong expectations, runs long, or pushes people into a menu affects containment and escalation before the agent has done any real work.

How disclosure rules shape the greeting

The first job of any greeting is disclosure. Customers should know they are talking to an AI, and in more places that is now the law. The European Commission's guidance on Article 50 of the EU AI Act says that from 2 August 2026, people must be told from the start when they're interacting with an AI system, unless it's already obvious.

California got there earlier. Its bot disclosure law, SB 1001, has since July 2019 made it unlawful to use a bot to mislead someone about its artificial identity in order to drive a sale. A clear, conspicuous disclosure protects the business.

What else a greeting has to do

Next comes scope. A short phrase naming orders, billing and account changes does more than a paragraph, and it keeps customers from spending their first turn on something the agent can't do.

The last two are the invitation and the tone. The greeting should end with a clear prompt for the customer's first turn. It's also the first sample of the brand's tone of voice, and a greeting that sounds different from what follows makes the whole agent feel stitched together.

Channel changes how much fits. In chat people can skim, so a greeting can carry a disclosure line, a scope line and a few suggested topics. On a call every word costs time, so the same content shrinks to who is speaking, that it is an AI, and an invitation to explain.

Open questions versus menus

The biggest design choice is whether to ask customers to describe the issue in their own words or to offer options.

An open prompt, like "What's going on with your account today?", lets people explain naturally, and modern language models handle that well. Someone with an unusual problem doesn't have to pick the closest wrong option.

Research from the older speech-recognition era found that open prompts work well when callers know what to say, but newer or infrequent callers often struggle. The usual fix, described by the Association for Conversational Interaction Design, is to follow the open question right away with one or two natural examples. A third example starts to sound like a menu.

Structured openings take the other side. Some design guidance for text assistants recommends showing a few common requests as buttons, because an open question invites requests the assistant can't handle. The cost is that the buttons shape what people ask for, and anyone whose issue isn't listed has to hunt for the free-text option.

Many teams land in the middle, with an open invitation and an example or two on voice, and a free-text box next to a few suggested topics in chat.

Using what the agent already knows

When a customer is logged in, the agent often knows a lot before the first message. It may see their name, a recent order, an open ticket or an outage in their area. Using that can shrink the first exchange to a single line: "Welcome back. Is this about your order from Tuesday, or something else?"

The working rule is to use context the customer would expect the brand to have and that is likely relevant to this contact. An open order or yesterday's failed payment is fair game. Guesses drawn from browsing behavior are not, and reading out several account details at once feels like surveillance.

Always keep the "or something else" path. The most likely reason for a contact isn't always the actual one.

New and returning customers can reasonably get different openings. A fuller greeting helps first-timers, while regulars usually want the short version.

Common mistakes in AI greetings

Length is the most common problem. Greetings collect content over time, from legal notices and promotions to capability lists, and each line has an internal sponsor. The result is a voice opening so long that callers talk over it.

Overpromising is next. A greeting that offers help with anything sets up the first escalation. An agent that claims broad skills and then deflects the first real request does more damage than one that sets narrower expectations and meets them.

Some greetings rebuild the phone tree. Opening with a numbered list, or only accepting the listed choices, recreates the IVR menu the agent was meant to replace. Customers who learned to say "representative" to old systems will do the same here.

The last is false familiarity. Exclamation marks, small talk and a persona that hides what the agent is can blur the very disclosure the greeting is supposed to make.

The greeting deserves the same measurement as any other turn in conversational AI design. The useful signals come right after it: how often the first message is something the agent can handle, how often people ask for a human straight away, and how often callers talk over the opening.

For a deeper dive, download Decagon's guide to agentic AI for customer experience.