What agent traffic looks like

Depending on how you define it, our data suggests up to 3 percent of inbound customer support traffic today comes from AI agents and other automated systems, and it has grown quickly in the last three months. It takes four main shapes.

• A consumer agent acting for a person explicitly. A consumer AI assistant that has been asked to cancel a subscription, chase a refund or change a booking, and opens a support conversation to do it. Some say so up front. Others relay the request in the third person and hand off when they hit a wall.

• A consumer agent impersonating a user. A consumer agent may send an email on their user’s behalf as the user, or engage a chat line seeking a refund as the user. This presents unique security and authentication challenges.

• A person using AI to write. The customer is human, but the message was pasted from an AI tool. Sometimes the tool's own words are still in it: "Here's a message you can send to support."

• Machine mail that lands as a customer. Auto-replies, receipts, submissions from partners, verification templates and bounces that reach the inbox and get treated as a person asking for help.

Every one of those shapes is growing. Personal AI agents such as Instinct, Meta’s Muse, OpenAI’s Dots and SpaceX’s Grok Bot are growing at a startling rate, and they are gaining the ability to browse, fill forms, call and email on their owner's behalf.

Why agent traffic matters

If agents go unidentified, your team can spend its time on their questions while human customers wait. An agent is a different kind of customer, with different needs and different capabilities, so it makes sense to treat it differently, and that starts with identifying it. Until you do, treating agents as people costs you in four places.

Policy

Your terms were written for people. Can an assistant accept a refund on its owner's behalf, consent to a data request, or agree to a plan change? Most teams have not decided, because they could not see the question being asked.

Verification

Identity checks assume the person answering is the account holder. An assistant that knows the date of birth and the last four digits passes the same check. Whether that is fine depends on the request, and you need to know it is happening to have a view.

Tone and effort

Your team spends care on empathy, reassurance and plain language. An agent needs none of it, only a clear answer, a structured one if possible, and a fast path to the outcome. The same goes for your AI agent's replies.

Measurement

Containment, CSAT and handle time all assume a human on the other end. A polite agent that never rates a conversation skews every one of them. You should at least be able to report agent traffic as its own line.

How agent traffic detection works

Agent traffic detection reads the customer side of every conversation after it ends, on chat, email, voice, SMS and WhatsApp, and asks four separate questions.

Each question gets two passes. A fast screen leans toward flagging so nothing slips through, and a stricter second read then has to agree, with the evidence in front of it, before anything counts. Only hits that survive both are recorded.

A verified hit tags the conversation, one tag per signal, in Lorikeet and in your ticketing system if the conversation syncs there. From that point it is ordinary ticket data: filter on it, report on it, or branch a workflow on it.

The tags say suspected for a reason. These are judgements about text, and a formal human or a pasted policy can look like a machine. During the beta we run detection in observation mode first and review the hits with you before any tag becomes visible.

Where this goes

Seeing agent traffic is the first step. Deciding what to do with it is what we are building next with beta customers.

• Prioritise humans. Once you can see agent traffic, you can put people first in the queue and stop agents consuming the attention meant for them.

• Verify harder where it matters. An assistant that knows the account holder's details passes the same identity check a person would. Knowing an agent is on the line lets you ask for more before you act on an account.

• Route agents to an agent-facing concierge. Lorikeet already offers an agent-facing endpoint that lets AI assistants talk to your support directly, with structured answers and no small talk. Detection is how you catch the ones that arrive through the front door and send them there.

• Report it. Agent traffic as its own line in your containment, satisfaction and volume numbers, so the human numbers stay honest.

Agents increasingly represent your customers, and the teams that notice first will set the terms.

Get started

It is early days and we are learning together. Agent traffic detection is available now in beta: we start in observation mode, share what we find on your traffic, and turn on tagging when you are ready. You can get started free today at lorikeetcx.ai/sign-up.