Voice AI Caller ID Authentication: Stopping Number Spoofing

The number on the screen is the first thing a person trusts and the last thing most enterprises think to protect. A voice AI agent inherits that same instinct: it answers a call, reads the presentation number, and, unless told otherwise, treats a familiar looking caller ID as a signal of who is on the line. It is not. Caller ID can be forged in seconds, and fraudsters do it precisely because it works. Ofcom research from February 2025 found that two in five UK phone users, 42%, had received a suspicious call in the previous three months, and that people are far more willing to answer a number that looks local and legitimate.

That trust asymmetry is the whole problem. When an enterprise runs voice automation at scale, two risks sit on top of each other: the agent may act on a spoofed inbound number as if it were verified, and criminals may spoof the enterprise's own outbound number to defraud its customers. Both erode the brand faster than any product defect. This guide sets out what caller ID authentication means for enterprise voice AI, why the UK blocks spoofed calls instead of running the US STIR/SHAKEN framework, how Ofcom's rules actually allocate responsibility, and how a voice agent should behave when the displayed number cannot be trusted.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What is caller ID authentication for enterprise voice AI?

Caller ID authentication is the practice of establishing whether the calling number a voice AI agent sees is genuine before that number is used to make any decision. It sits at the telephony layer, not the identity layer: it asks whether the number can be trusted, not whether the caller is the right person. Dilr Voice treats the presentation number as an untrusted input by default and records its trust state on every call.

The distinction matters because caller ID was never designed as a security control. The presentation number, the caller line identity or CLI, is set by the originating network and can be populated with almost any value as a call crosses provider boundaries. That is what makes number spoofing possible, and it is why suppression controls such as Do Not Call logic in AI voice diallers must key off verified numbers, not the displayed CLI. As Ofcom puts it plainly:

A common tactic used by scammers is to 'spoof' their telephone number to disguise the origin of the call, or to make their call appear to be from a trusted person or organisation.

For a voice agent, the practical consequence is that the CLI is a hint, not a credential. It is useful for routing, personalisation and fraud scoring, but it cannot on its own gate access to an account or authorise a payment. That is a separate control, covered in our guide to voice AI caller identity verification, and the two should never be confused in a production enterprise voice AI design.

Why can't the UK use STIR/SHAKEN like the US?

The UK cannot yet run STIR/SHAKEN because most calls still traverse the legacy analogue network, and STIR/SHAKEN only works end to end over internet protocol. In the US, the FCC required major voice providers to implement the framework in the IP portions of their networks by 30 June 2021 under the TRACED Act. The UK's public switched telephone network is not fully digital, so cryptographic attestation cannot survive the journey from origin to destination intact.

STIR/SHAKEN, short for Secure Telephone Identity Revisited and Signature-based Handling of Asserted information using toKENs, lets an originating provider cryptographically sign the calling number and assert how confident it is that the caller is entitled to use it. The FCC framework defines three attestation levels: full (A), where the provider knows the customer and confirms their right to the number; partial (B), where the provider knows the customer but cannot confirm the number; and gateway (C), where the provider has no direct relationship with the originator. A terminating network, or a voice AI platform sitting behind it, can then treat an A-attested call very differently from a C-attested one.

Will STIR/SHAKEN come to the UK after the 2027 switch-off?

It becomes technically possible once the migration completes. According to the House of Commons Library, by January 2027 landline telephone services in the UK will switch to a fully digital network, retiring the analogue PSTN. An all-IP network removes the non-IP tandem switches that today break the signature chain, so attestation could in principle be carried and checked. Ofcom has been cautious, judging that CLI authentication alone would not stop scam calls originating overseas.

How does Ofcom's CLI blocking regime actually work?

Rather than wait for full authentication, Ofcom took a blocking approach. Its updated Calling Line Identification Guidance came into effect on 29 January 2025 and expects providers to block international calls that display a UK number as the presentation number, except in limited legitimate cases. The logic is simple: a call from abroad showing a UK landline is almost always spoofed, so the network drops it before it reaches an inbound voice agent.

The scale is already meaningful. Ofcom reports that industry measures are blocking around a million calls a day, and on 21 July 2025 it proposed a further step: withholding the caller ID on calls that appear to come from a UK mobile roaming abroad unless the network can verify the number is genuine. The direction of travel is clear. Unauthenticatable numbers are increasingly filtered, withheld or blocked upstream, which changes what an enterprise voice agent actually sees on the line.

Who is responsible when a voice AI number is spoofed?

Responsibility splits between the telecoms provider and the enterprise, and enterprises routinely misread where their own duty sits. Ofcom's blocking rules bind the communications providers that carry the call, not the business running the voice agent. The enterprise cannot make its network block spoofed traffic; that is the provider's job under the General Conditions. What the enterprise owns is the integrity of its own outbound identity and how its voice AI agent reacts to inbound numbers it did not authenticate.

On the outbound side, the enterprise's exposure is practical as well as legal. Presenting clean, valid, recognised numbers is how a business keeps its legitimate calls from being filtered out alongside the spoofed ones once providers apply their blocking rules. There are statutory obligations here too: under the Privacy and Electronic Communications Regulations, a direct marketing caller must identify itself and give either an address or a free telephone number on which it can be reached, and the ICO enforces PECR. These sit alongside the wider GDPR and PECR obligations for outbound calling that any outbound voice AI programme has to satisfy.

On the inbound side, the enterprise decides how much its agent trusts an unverified number. That decision is a governance choice, not a telephony default, and it belongs in an AI operating model with a named owner. Getting it wrong is how a spoofed number ends up treated as a returning customer.

How should a voice AI agent treat an unauthenticated inbound number?

A voice AI agent should escalate its caution as the displayed number becomes less trustworthy, never granting sensitive access on the strength of caller ID alone. Dilr Voice records a trust state for every inbound call and maps it to a defined behaviour: a known-good number flows straight into self-service, an unverified UK number proceeds with sensitive actions gated behind step-up, and an invalid or withheld number is limited to information that carries no risk.

The point is to make the number's trust state an explicit input to the conversation, not an invisible assumption. The ladder below shows how the same agent can serve a verified caller quickly while refusing to leak account data to a spoofed one.

This is confidence-gated design applied to the transport layer rather than the speech model. It pairs naturally with the audio quality and infrastructure controls that decide whether the agent can hear the caller at all, and with behaviour-level handling such as abusive-caller de-escalation, which governs conduct rather than number trust.

Why does protecting your own number matter so much?

It matters because the numbers people trust are exactly the ones criminals want to wear. Ofcom's research shows the asymmetry starkly: a quarter of people, 26%, said they were likely to answer an unrecognised UK mobile, against just 9% for an international number with an unknown country code. A recognised, local-looking number is a currency, and when a fraudster spoofs an enterprise's published line to impersonate its fraud team, every answered call spends that brand's reputation.

Practical protection for an enterprise number is unglamorous but effective. Register the numbers you actually use for outbound so providers can recognise them, retire numbers you no longer originate calls from so they cannot be quietly reused, keep presentation numbers valid and diallable, and add numbers you will never call from to a Do Not Originate list so any call claiming to be from them is treated as fraud. The same discipline that scopes a voice deployment should confirm these controls exist before go-live, and the execution office should keep them under review as the estate changes.

The same trust-state logic runs through Dilr Voice itself, where the number behind every inbound call is recorded and governed rather than quietly assumed.

What is the best voice AI platform for caller ID trust in 2026?

The best platform depends on where your calls originate and how regulated your outcomes are, so the honest answer names trade-offs rather than a single winner. For US-heavy outbound, where STIR/SHAKEN attestation is available end to end, a platform wired tightly to a carrier like Twilio can read attestation signals directly and will suit teams that want to build the logic themselves.

Among self-serve builders, Vapi, Retell AI and Synthflow give the most direct access to those signals. If your priority is a turnkey enterprise voice experience with mature call handling, PolyAI is a strong contender and will win on time to value for a contact-centre replacement.

Where Dilr Voice earns its place is the regulated UK deployment: calls that cross the still-analogue PSTN, where attestation is not yet available, and where the enterprise needs the trust state of every number recorded, governed and auditable rather than left implicit. If your voice programme is a single US outbound campaign, one of the self-serve tools may be the faster route, and we will say so. If it spans regulated inbound, brand-number protection and a defensible audit trail, that is the ground we are built for. See how we place AI inside enterprise systems and what DATS covers for the full picture.

Does caller ID prove who is calling?

No. Caller ID identifies a number, not a person, and the number itself can be spoofed. It is a useful routing and fraud-scoring signal, but it cannot authorise access to an account or a payment on its own. Proving who is calling is a separate discipline, handled through knowledge and possession factors and risk-based step-up, which we cover in our guide to voice AI caller identity verification.

How does this connect to outbound campaign design?

Number trust runs through outbound as well as inbound. Presenting a valid, recognised CLI keeps legitimate campaigns from being blocked, and rotating or disguising numbers to boost pick-up rates now risks both PECR enforcement and network filtering. Outbound pacing, voicemail handling and calling-hours rules are covered in our guide to voice AI outbound batch timing; caller ID authentication is the trust layer that sits underneath all of it.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read more across the voice AI blog.

Caller ID authentication will not stay a niche concern. As the AI adoption gap widens, with around 88% of enterprises using AI but only about 6% capturing material value per McKinsey's State of AI, the programmes that pull ahead are the ones that treat trust, provenance and audit as first-class engineering. For voice, that starts with refusing to believe the number on the screen until you have earned the right to.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.