How Contact Centres Can Detect Personal AI Agents Today

How do we even know when the “customer” on the line is a machine?

Personal AI agents can already search websites, complete forms, start chats and make calls on a customer’s behalf. When I tested two consumer AI agents on my own website, both submitted meeting requests without human involvement. Nothing identified either visitor as an AI agent.

Most contact centres are in the same position today. They may be serving personal assistants, business bots and malicious synthetic callers without knowing which is which. Detecting the machine is only the first challenge. The harder questions are whether the agent genuinely represents the customer, what it is allowed to do, and who is accountable when something goes wrong.

One recent signal of scale comes from Amazon. The company says more than 400 million customers used Alexa for Shopping during the past year, with interactions increasing 570 percent year over year. Alexa for Shopping operates within Amazon, so it is not a Poppy implementation or evidence of cross-company delegated authority. It does, however, show how quickly customers are becoming comfortable asking AI to research, manage and act on their behalf.

Not All Non-Human Callers Are the Same

Before talking about detection, it helps to separate three very different kinds of callers:

  • Legitimate personal agents. A customer's AI acting on their behalf to book, cancel, dispute or ask. These are customers by proxy, and you want to serve them well.

  • Business agents. Another company's AI contacting you, such as a supplier, partner or aggregator.

  • Malicious synthetic callers. Cloned voices and bots trying to take over accounts, probe your IVR or socially engineer an agent.

The goal is not to block machines. It is to recognize which of these you are dealing with and send each down the right path.

Detecting AI on the Voice Channel

Voice is the hardest channel, because a good synthetic voice can fool a human agent. Detection works best in layers:

  • Synthetic voice and liveness detection. This is the most mature option. Specialist vendors analyze the audio itself for signs that it was machine-generated. Pindrop integrates with NiCE CXone, and Resemble AI integrates with Genesys Cloud. Pindrop's newer BotStopper aims to tell human from machine in about two seconds, backed by a registry of more than 5,000 known AI voices, so it can suggest which AI is calling, not just that one is.

  • Behavioural signals. AI callers tend to behave differently: unusually consistent response timing, perfectly timed keypad input, very clean background audio and no hesitation when interrupted. On their own, these are hints, not proof, but they add weight to a score.

  • Network signals. VoIP origination and caller ID attestation (STIR/SHAKEN) add context. Be careful, though. STIR/SHAKEN tells you whether the number is genuine, not whether a human is speaking.

  • Simply ask. Reputable agents generally disclose that they are automated. An opening prompt such as "Are you calling on someone's behalf?" costs nothing and catches the honest ones.

Detecting AI on Digital Channels

On digital channels, most of the work happens at the web layer, before a chat or messaging session even starts:

  • Signed agent identity. This is where the industry is heading. Web Bot Auth, a Cloudflare-led IETF proposal, lets bots and AI agents cryptographically sign their web requests. Visa's Trusted Agent Protocol and Mastercard's Agent Pay both build on it. A separate effort targets customer service. The Personal Agent Protocol, known as Poppy, is an open draft from Sierra and Meta, with Genesys and NiCE among its design partners. It describes how a personal agent can discover a company, identify itself, connect a customer through OAuth and work through the company's website, APIs or company agent. At the time of writing, the specification identifies itself as draft version 0.3 and warns that any part may change before a stable release. Design partnership does not mean that a vendor has deployed a conforming production service.

  • Emerging identity services. The market is beginning to produce services around this problem. Twilio's Agent Identity, currently in private beta, uses OAuth 2.1 and OpenID Connect so an application such as an AI agent can request access on a user's behalf. Twilio cautions that the service may change, that some capabilities are not yet implemented and that it does not carry a general availability service commitment. Even with those limits, it is a practical example of the shift from merely detecting an AI agent to validating whom that agent represents.

  • Bot management. Web security tools already look at mouse movement, typing cadence, form completion speed and browser fingerprints. An agent filling in a contact form in two seconds stands out.

  • Declared identity. Some agents identify themselves in their user agent or request headers. Log it, even if you do nothing with it yet.

One warning: resist the urge to throw a CAPTCHA at everything. It blocks the legitimate agents your customers are choosing to use, and that will increasingly feel like refusing to serve the customer.

Poppy From Detecting an Agent to Trusting One

Recognizing an AI agent does not establish that it is legitimate, represents the customer it names, or has permission to complete a particular task. Poppy offers a technical starting point for carrying identity, a customer sign-in and scoped access across a session. It does not remove the need for business policy, legal analysis or risk controls.

  • Identification. Verify which personal agent is connecting, which company or operator controls it and whether its credentials are valid. Then bind that agent session to the customer through an authentication flow the business trusts. A declared agent identity is useful, but declaration alone is not proof.

  • Delegated authority and consent. Record what the customer allowed the agent to do, for which account, for how long, and under what limits. Permission to read a balance does not imply permission to move money, change an address or cancel a service. Consent should be specific, current and available as evidence after the interaction.

  • Ownership and control. Know who operates the personal agent and who owns the company agent or workflow it contacts. Ownership matters because each party controls different code, models, credentials, policies and logs. A protocol can connect those systems, but it does not automatically allocate responsibility between the customer, agent provider, business and technology vendors.

  • Accountability. Assign an internal owner for the decision to accept an agent's request. The business still needs a policy engine that checks the requested action against the customer's grant, product rules and fraud controls. It also needs an escalation path when the request falls outside those limits.

The Bank Transfer Test

Consider a personal assistant that asks a bank to transfer $5,000. Detecting that the request came from an AI is the least important part of the decision. The bank must first verify the agent's identity and connect it to the right customer. It must then confirm that the customer's grant covers transfers, the source account, the intended recipient, the amount and the time window.

The bank may still require fresh customer confirmation for a new recipient or a high-value transaction. That confirmation should state the material details, not merely ask whether the customer approves "the agent's request." The authorization service should enforce transfer limits before execution rather than relying on the agent to stay within them.

The audit record should preserve the agent identity, customer identity, authorization scope, confirmation event, policy result, transaction reference and any human intervention. If the customer revokes the agent before the bank executes the transfer, the bank needs a reliable way to reject the pending request and invalidate future access. If something goes wrong, liability will depend on the facts, account terms and applicable law. Poppy does not decide that question.

Governance Enforcement and Revocation

Poppy is an emerging draft protocol, not a law, licence, regulator, universal certification service or guarantee that an agent is safe. Its technical rules can help participants identify sessions and apply access scopes. Each business must decide which credentials it accepts, how it proves customer consent, what actions it permits, and which controls apply to higher-risk work.

A practical control model needs more than authentication. It should enforce least privilege, short-lived credentials and transaction limits; log requests and decisions in a form that can be investigated; monitor for unusual behaviour; support immediate revocation by the customer or business; and retain a human path for disputes, vulnerable customers and ambiguous requests. These controls should apply whether the agent reaches the company through a website, an API or another AI agent.

Putting It to Work in Genesys Cloud and NiCE CXone

Neither platform has a native "is this an AI?" switch today. Both vendors are design partners on Poppy, but that is not evidence of a generally available implementation. Watch their product roadmaps. Until something ships, both platforms have the components needed to build an internal control path. The pattern is similar on either platform:

  • Classify. Early in the flow, call your detection service through a data action in Genesys Architect or a REST call in a NiCE Studio script, and return a score.

  • Tag. Store the result as an interaction attribute, for example, callerType = human, verified agent or unknown agent, so routing, reporting and analytics can all use it.

  • Verify identity. For agents, the first question changes from "Is this the customer?" to "Which agent is this, which customer does it represent, and how was that relationship established?" Validate the agent's credentials and bind the session to an authenticated customer before exposing account data.

  • Authorize and enforce. Translate the customer's grant into machine-enforced scopes and transaction rules. Treat read, change and payment permissions separately. Require fresh confirmation or human review for work that exceeds a limit or carries material risk.

  • Record and revoke. Keep an audit trail that links the agent, customer, consent, policy decision and outcome. Make revocation effective across web, API and agent-to-agent sessions, and test that expired or withdrawn credentials fail.

  • Route. Send verified agents to a structured, API-style path designed for machines. Keep them out of queues built for people. Send unknown synthetic voices to step-up authentication or your fraud team.

  • Measure. Report AI-originated contacts as their own category. Gartner warns that when customers can raise requests with near-zero effort, volumes will rise. You want to see that trend in your own data first.

Watch Out for False Positives

Not every synthetic voice belongs to an AI agent. Some people with speech or hearing disabilities use text-to-speech to make calls. A detection policy that automatically blocks or deprioritizes synthetic voices could lock out the very customers who need the most care. Use detection to choose a path, not to slam the door.

Voice analysis can also raise privacy questions, particularly where biometrics are involved. Check your obligations in each region you serve before switching it on.

Where to Start

  • Find out whether you are already receiving AI-originated contacts. Your web and voice logs probably hold clues.

  • Agree on a policy for legitimate personal agents before you choose technology. Define which tasks an agent can perform, which require fresh customer confirmation, and which always require a person.

  • Pilot synthetic voice detection on a high-risk queue, such as account changes or payments.

  • Make your most common tasks available as simple, secure APIs, so good agents have a better option than pretending to be human.

  • Map accountability before launch. Name who owns the policy, who can change it, who investigates a disputed action and who can revoke access during an incident.

The Real Shift

Detection is a stepping stone. Today we infer whether a caller is human from clues. Emerging protocols such as Poppy could let legitimate agents announce themselves and connect to a customer through verifiable credentials. The contact centre would then have a better basis for deciding what the agent may do. That still leaves the hard work with each organization: setting policy, enforcing limits, recording consent, monitoring behaviour, revoking access and accepting responsibility for the decisions made by its own systems and staff.

Agent-led journeys may also change how customers discover companies. If a personal assistant begins the search, a business that blocks legitimate agents, or cannot serve them safely, may never enter the customer's consideration set.

Next
Next

Native or Best-of-Breed: Where Should Your CCaaS Platform Stop?