Knowledge-Based Authentication in Voice AI

Before a voice AI agent can help a customer, it has to know who the customer is. That sounds simple. In practice, it is one of the most consequential design decisions in a contact centre deployment.

Get it wrong in one direction, authentication that is too weak, and you expose customer accounts to social engineering, fraud, and regulatory sanction. Get it wrong in the other direction, authentication that is too friction-heavy, and you lose customers before the interaction has even started. The Ponemon Institute has found that customers who experience a failed or frustrating authentication process abandon at rates that dwarf nearly every other drop-off point in the contact centre journey.

Knowledge-based authentication (KBA) is the industry’s answer to this balance. Not a perfect answer, and not the last word. But the answer that has been validated against real fraud patterns, real regulatory requirements, and real customer behaviour across industries, and the one that remains the foundation of identity verification in voice channels for good reason.

Knowledge-based authentication is the verification of a caller’s identity through information that only the legitimate account holder should know. In practice, this typically means a combination of static identifiers (account number, date of birth, postcode) and dynamic or transactional data (last payment amount, recent transaction details, a one-time code sent to a registered device).

The specific combination used in any given deployment is not arbitrary. It reflects a risk-based design decision guided by several industry frameworks:

These frameworks converge on a shared conclusion: effective voice channel authentication requires layered verification, something the customer knows, corroborated by something transactional or dynamic, calibrated to the risk level of the interaction being accessed.

KBA is the implementation of that principle at the point of call.

Authentication is not a hurdle between the customer and the agent. It is the mechanism that makes it safe to give the agent access to that customer’s account.

Authentication in a voice channel is harder than authentication in a digital channel. That is not a UX observation. It is a technical and fraud-risk reality that shapes how KBA is designed.

In a digital channel, you can layer biometrics, device recognition, behavioural analytics, and passive signals into the authentication stack with relatively low friction. In a voice channel, you are working with an audio signal and a conversation. The same richness of passive verification is not available, or at least, not available in the same low-friction form.

At the same time, voice channels are disproportionately targeted by social engineering. The 2024 UK Finance Annual Fraud Report identified authorised push payment fraud, much of it originating through phone-based impersonation, as the dominant fraud vector by volume. Fraudsters specifically target voice channels because the authentication mechanisms have historically been weaker than digital equivalents.

KBA addresses this by requiring callers to demonstrate possession of information that a fraudster is unlikely to have acquired, particularly when transactional or dynamic data is included in the verification set. A fraudster who has acquired a customer’s name and date of birth from a data breach still cannot answer questions about the last three transactions on the account, or respond to a one-time code sent to the registered mobile number.

This is why the industry has standardised on layered KBA for voice channels, not because it is perfect, but because it provides meaningful resistance to the specific fraud vectors that voice channels face.

One of the most important and least visible aspects of KBA design is calibration, the process of matching authentication strength to interaction risk.

Not every call requires the same verification depth. A caller checking their account balance faces a different risk profile than a caller requesting a large payment, updating contact details, or resetting account credentials. Applying maximum friction to every interaction creates unnecessary abandonment. Applying minimum friction to high-risk interactions creates unnecessary exposure.

This calibration is not a product decision. It is a risk management function, and in regulated industries, it is typically governed by a combination of internal risk policy and external regulatory expectation.

The principle behind it is recognised across frameworks: authentication should be proportionate to the risk of what it is protecting. NIST 800-63B frames this as Identity Assurance Levels. PCI DSS frames it through its tiered control requirements. The FCA frames it through its Consumer Duty proportionality obligation. The language differs; the underlying logic is the same.

When a voice AI agent handles authentication, something changes in the experience. The interaction is faster. It is consistent. There is no human variability in how the questions are asked or how the responses are validated. The process is the same at 2am on a Sunday as it is at 9am on a Monday.

What does not change is the standard the authentication needs to meet.

A voice AI agent processing KBA is still operating within the same regulatory and risk framework as a human agent processing the same verification. The automated nature of the interaction does not reduce the authentication requirement. In some jurisdictions, it increases the scrutiny, because automated systems that fail to authenticate correctly at scale represent a systemic risk, not an individual error.

This has a practical implication for how voice AI KBA is designed and operated. The authentication logic cannot be simplified for automation convenience. It needs to be at least as robust as the human equivalent it is replacing, and it needs to handle the edge cases that human agents handle through judgment.

A caller who misremembers their postcode. A caller whose registered details are out of date. A caller who is the legitimate account holder but cannot pass static KBA because they recently moved. These situations require fallback handling, a graceful path to identity verification that does not simply lock the caller out, but also does not abandon the verification requirement.

A caller contacts a utilities company to report a billing dispute. They pass the first KBA factor (account number) but cannot recall the postcode on the account. They moved six months ago and the update was not processed. A well-designed voice AI KBA system does not fail the authentication at this point. It offers a dynamic fallback: a one-time code to the registered mobile, or a warm transfer to a human agent with the partial authentication context preserved. The authentication standard is maintained. The customer is not abandoned.

At Voicing.ai, our KBA implementation follows the layered, risk-calibrated model that industry frameworks specify, because that is what responsible deployment in a voice channel requires.

Every deployment is configured with an authentication design that reflects the risk profile of the interactions it will handle. Low-risk enquiry flows use streamlined verification. High-risk transactional or account-change flows use layered verification, including dynamic factors. The configuration is not a default. It is a design decision made in collaboration with each client, informed by their regulatory obligations, their fraud risk profile, and their customer population.

We also build fallback paths into every KBA implementation. A caller who cannot pass primary KBA has a route to identity verification that does not simply terminate the interaction, but that route maintains the authentication standard rather than bypassing it.

Our voice AI agents ask authentication questions in the same way for every caller, every time. There is no variability in phrasing, no variation in the information required, no inconsistency based on who is handling the call. This consistency is one of the structural advantages of automated KBA: the authentication process is exactly what it was designed to be, every time it runs.

Beyond the interaction itself, our KBA implementations are logged and auditable. Every authentication attempt, successful or not, is recorded in a way that supports compliance review, fraud investigation, and quality assurance. In regulated industries, this audit trail is not optional. It is the evidence base that demonstrates the authentication standard was met.

Authentication design is an area where the gap between providers who understand the standards and providers who have implemented a surface-level version can be significant. These questions will help you assess which side of that gap you are on:

Authentication is not a feature. It is a risk management function. A provider who cannot answer these questions with specificity is a provider who has not engaged seriously with what authentication in a voice channel actually requires.

Voicing.ai’s authentication implementations are designed against the same industry standards that govern enterprise contact centre deployments, because that is the standard. If you want to review your current authentication design against those standards, our team can walk through what a risk-calibrated KBA configuration looks like for your specific deployment context.