AI

AI Granny

Virgin Media O2 built Daisy, an AI granny chatbot that keeps phone scammers trapped for 40 minutes. Here is how she works and what she costs them.
An AI granny chatbot named Daisy wasting a phone scammer's time on a vintage telephone with speech waveforms around her head.

Introduction

The AI granny is a software persona that answers phone scam calls, keeps the fraudster on the line, and refuses to hand over anything useful. Virgin Media O2 built the most famous version, named Daisy, and the first wave of AI chatbot commentary mistook her for a joke. The underlying numbers are serious because the FBI logged $893,346,472 in AI-related scam losses from 22,364 complaints in 2025, and the volume is still climbing. Hiya reports that one in four Americans received a deepfake voice call in the last 12 months, which is the the attacker side of fraud of the same market. An AI granny does not stop these calls, yet she can tie up a fraudster for forty minutes and make the next victim slightly safer during that window. This article unpacks what Daisy actually does, why she exists, the economics behind the trick, and whether consumers can build their own.

Quick Answers on AI Scambaiting Bots

What is the AI granny chatbot?

An AI granny chatbot is a voice-based conversational agent that answers scam calls, impersonates an elderly target, and keeps the scammer talking so they cannot defraud real victims during that window.

Who built the Daisy AI granny?

Virgin Media O2 built Daisy with its in-house AI agency Faith and the London creative agency VCCP, launching her in November 2024 as part of the Swerve the Scammers campaign.

How long does the AI granny keep scammers on the phone?

Reported Daisy calls run from a few minutes to over forty minutes, with Virgin Media O2 saying the record session has repeatedly crossed the forty-minute mark through 2025 and into 2026.

Key Takeaways on the AI Granny Playbook

  • Daisy pairs a speech-to-text pipeline details layer, a large language model, and a text-to-speech voice to run an elderly persona in real time against live scammers.
  • The business goal is not to recover money but to drain the scammer’s working minutes, since every call tied to a bot is a call not aimed at a real target.
  • The playbook is already being copied by banks, consumer apps, and regulators building voice honeypots to study scammer tradecraft.
  • Legal risk sits in recording consent, impersonation, and automated-dialer rules, and these rules differ between the UK, the EU, and the United States.

What Is the AI Granny Chatbot

An AI granny chatbot is a voice agent that answers suspected phone scam calls and impersonates an elderly target with rambling stories, keeping the fraudster talking so real users are shielded during the call window, with Daisy being the best-known example from Virgin Media O2.

Scambaiter ROI Explorer

Pick a center size, average call length, and bot cost. See how much attacker revenue a Daisy-style bot destroys per shift.

5 operators

120

20 min

560

$100

Low intentHigh intent

Attacker minutes consumed

600

per shift of five operators pulled into Daisy

Attacker revenue denied

$1000

per shift based on operator-hours lost

Return per $1 of Daisy

100x

against roughly $0.60 per Daisy conversation hour

Model inputs calibrated against FBI IC3 2025 reporting and Hiya State of the Call 2026. Figures are directional, not audited.

The Virgin Media O2 Daisy Project in Plain Terms

Virgin Media O2 unveiled Daisy in November 2024 as the newest member of its fraud prevention case studies team. The carrier worked with its in-house AI agency Faith and the London shop VCCP, and partnered with the YouTube scambaiter Jim Browning to tune her persona. Daisy sits behind a dedicated phone number that Virgin Media O2 seeded into databases scammers scrape, so her calls arrive without human bait. When a fraudster rings, she answers in character and launches into stories about her cat, her grandchildren, and her patchy memory for banking details. The carrier reports that she has kept individual scammers on the line for more than forty minutes in a single call. The point is not to arrest anyone; the point is to burn the attacker’s working hours so the next real target never gets dialed.

The campaign framing was deliberately playful because Daisy exists inside a larger message about carrier responsibility for voice fraud. Virgin Media O2 pointed to UK fraud losses of $2.9 billion reported across 2023 and said that one in five British citizens hits a scam attempt every week. The carrier’s own survey found that 71% of Britons want to retaliate against scammers but will not sacrifice their own time to do it. Daisy is designed as the retaliation Britons said they wanted without the time cost those same Britons refused to pay. The emotional premise maps cleanly onto the technical premise, and that alignment is why the project resonated outside of trade press. It also gave Virgin Media O2 a story to tell regulators and competitors about proactive fraud prevention.

Industry observers quickly placed Daisy inside the “campaign as product” tradition the WARC and Cannes awards jury likes to reward. The WARC Cannes case study on Daisy describes a three-month build and launch cycle with press pickup in more than forty countries. Decision Marketing, PRWeek, Ads of the World, and the Guardian covered the launch, and The Guardian’s November 2024 feature framed Daisy as a cultural rebuttal to the scam economy. The free press alone exceeded the campaign’s media budget, which is the metric agencies live and die by. For a telecommunications brand that fights churn against three domestic rivals, the goodwill reservoir was its own ROI. Daisy is now a case taught in UK marketing graduate programs for exactly that reason.

Source: YouTube

How Daisy Works: Implementation Pipeline Behind the Scenes

Daisy runs on a three-stage pipeline that is now standard for any voice agent, including modern conversational AI products. The first stage is automatic speech recognition technology overview, often a Whisper variant. It turns the scammer’s audio into text in near real time. That text streams into a large language model with a persona prompt. The model also carries a long rolling context of the conversation so far. The LLM returns a candidate reply, which is then sent to a text-to-speech voice tuned to sound like an elderly woman with regional UK inflection. The round trip has to beat a human conversational pause of roughly 600 milliseconds, or the scammer senses the lag and hangs up. Virgin Media O2 has not disclosed the specific vendor stack, but industry teardowns suggest a mix of commercial ASR, a hosted LLM, and a custom voice.

Moving on to the behavior layer, the persona prompt is where most of the design work happens. Daisy’s prompt instructs the model to drift from direct account questions, volunteer irrelevant stories, and ask the scammer to repeat themselves often. The behavior layer carries stop phrases that end the call politely when the attacker senses the trick. This matters because an angry scammer who realizes they were trapped sometimes tries to retaliate by escalating to abuse or by returning with a different number. The prompt engineering is iterative, and Virgin Media O2 has said it adjusts her repertoire monthly based on logs from the previous period. Prompt changes are the cheapest way to improve the bot, which is important because fine tuning a voice model is comparatively slow.

The infrastructure sits behind a dedicated inbound number that routes scam calls into the Daisy conversation stack. Virgin Media O2 seeded that number into scammer databases through what it calls “scambaiter seeding” with its research partners. Jim Browning’s own research network provided the first several thousand suspected numbers used to tune the system. Daisy only answers calls that reach that number, so she never intercepts legitimate Virgin Media O2 subscribers by accident. The carrier bears every per-minute cost and the compute cost, which is modest at the current call volume but will scale with any wider rollout. A technical white paper has not been published, and Virgin Media O2 has declined to confirm the exact per-call cost in its press briefings. Analysts estimate the all-in cost at under one British pound per minute, which is roughly the price of a London bus fare.

A real limitation is that Daisy cannot be used to defend an individual subscriber’s own number without re-engineering the stack. The current product routes a different inbound number to the bot and does not sit between the scammer and a protected subscriber. That architectural choice avoids the thorny two-party consent problem that would arise if Daisy picked up a real subscriber’s calls. It also means Daisy is a research and PR instrument rather than a per-line scam shield at this phase of the project. Carrier-level call screening products from the same market, such as those built on top of Hiya’s voice fraud intelligence, operate differently. A future Daisy that intercepts live scams on behalf of a specific line is possible but demands a separate consent and regulatory framework.

The Scambaiting Tradition That Trained the Approach

Shifting from the technical plumbing to the lineage, scambaiting is a two-decade-old subculture that predates generative AI. YouTubers such as Jim Browning, Kitboga, and Pierogi built audiences in the millions by hopping on the line with Indian call center fraudsters and wasting their time. The subculture built a shared library of personas, delaying tactics, and reversal moves that eventually became training data for AI equivalents. Virgin Media O2 did not invent the technique, it productized a hobby. That lineage matters because the ethical and legal arguments against AI scambaiting are the same arguments once raised against human scambaiting, and most of them lost in court. The genre also gave Daisy her tone, since the best scambaiters perform warmth, confusion, and distraction in sequence.

A separate strand of academic work at UC Berkeley and other universities studied scambaiting as a form of security research into social engineering. Those studies found that scammers follow predictable scripts and that interruption at specific moments breaks the whole flow for the rest of the call. Daisy weaponizes that research by hitting the predictable moments with digressions and misunderstandings tuned for maximum derailment. The result is a product whose roots run through academic security research, three million hours of scambaiter YouTube, and a modern voice stack. The genealogy also explains why the launch resonated so widely across both technical and lay audiences in late 2024.

Phone Fraud Numbers That Explain Why Daisy Exists

Beyond culture, the fraud economy that Daisy addresses is enormous and still growing year over year across every major market. The FBI IC3 report for 2025 logged $893,346,472 in AI-related losses from 22,364 complaints, up substantially on 2024. Hiya’s State of the Call 2026 survey found Americans receive 9.9 unwanted calls per week, growing roughly 16% year over year since 2023. The UK carrier industry and Ofcom put annual fraud losses in the billions. The Commons Library briefing confirms 2.9 billion pounds of reported UK fraud in 2023. The underlying trend is consistent in every market where carriers publish numbers, from Canada to France to Germany. The problem is accelerating faster than enforcement because an AI-assisted scammer can run far more calls per hour than a human one.

Breaking the number down, the FBI’s 2024 Elder Fraud Report put losses from Americans over 60 at $4.885 billion, a share far larger than the 60-plus population itself. Hiya reports that seniors aged 55 and older lose an average of $1,298 per phone scam incident, roughly triple the loss suffered by younger victims each cycle. The scammers know this, which is why the overwhelming majority of AI-generated voice attacks are aimed at parents and grandparents. The emotional lift of a grandchild in trouble is the most effective hook. This is the segment Daisy was explicitly designed to protect by consuming the attacker’s time before the attacker reaches a real target. It is also the segment that regulators cite when calling for carrier-side defenses rather than consumer-side education campaigns.

A less obvious statistic concerns the labor supply behind scam call centers in India, Bangladesh, and West Africa. Hiya and the Microsoft Digital Defense Report estimate the global scam call center workforce at 300,000 to 500,000 workers worldwide. Those workers have monthly call quotas, so every minute lost to Daisy is a minute subtracted from a quota the worker must still hit. Daisy therefore reshapes the labor economics of scam centers at the margin, which is the only lever a carrier can actually pull. Shutting down a call center is a law enforcement question in a foreign jurisdiction, and that pathway has been slow and uneven. A productized time sink changes the unit economics of the scam while enforcement catches up.

The Economics of a Wasted Scammer Hour

Beyond the headline figures, a human scammer runs about 10 to 20 target calls per hour on an auto-dialer. Trade reporting from Webpronews on scam call professionalization suggests a successful scam yields between $500 and $5,000 in extracted value per hit. A typical center converts fewer than 1% of called numbers into a payment, so the per-hour expected value sits somewhere between $25 and $200 per operator. If Daisy pulls that operator onto a 40-minute call, she consumes two thirds of the attacker’s working hour. Across a shift of five operators, that is three full hours of productive attacker time destroyed per Daisy interaction. The arithmetic is why reporters keep asking why every carrier in the world has not deployed one yet.

Shifting to the carrier side of the ledger, running Daisy at Virgin Media O2 scale costs on the order of pennies per minute of attacker time. Compute, voice synthesis, and telecoms minutes add up to perhaps 60 pence per hour of Daisy conversation at current cloud pricing. Against an attacker loss of $25 to $200 per hour, the ratio for the carrier is between 50x and 400x in favor of running the bot. Daisy also doubles as a telemetry source because every call records the scam script, number of origin, and conversational patterns for future detection. That telemetry feeds into carrier-level screening products that are themselves sold as fraud defenses to enterprise customers. Virgin Media O2 therefore gets a PR win, a telemetry stream, and a direct attack on scammer unit economics for the same money.

A quieter dimension is reputational insurance against the day a major deepfake reporting technology voice fraud lands on a vulnerable Virgin Media O2 subscriber. The OECD incidents database already lists deepfake voice calls targeting one in four Americans, and the UK trend is tracking that curve. When the inevitable headline hits, a carrier that already runs an AI granny can point to a visible defense and a public campaign. A carrier without one looks flat footed, which is a market risk in a sector that trades on consumer trust. That reputational layer, like other AI carrier security considerations, is often omitted from the public ROI case. Yet it may be the largest line item of all. Insurance that pays out in brand equity rather than cash is still insurance.

How Scammers Use AI to Attack the Same People

Looking at the attacker side, every tool that lets Daisy stall a scammer also lets a scammer impersonate a stranger’s grandchild. The same voice pipeline powers attackers who clone a voice from a social media clip under 20 seconds long. Hiya’s 2026 data shows 24% of Americans unsure they could tell AI voice from real. The attacker only needs to clear that low bar of recognizability to succeed with the clone. The upside for the attacker is scale, because one operator can run dozens of parallel cloned-voice calls when wired to a model. The voice signal that elderly victims trust most is the one easiest for attackers to counterfeit. This is the arms race the industry refers to when it says the attack surface is expanding faster than the defense.

A particularly cruel pattern involves scraping a victim’s grandchild’s voice from TikTok or YouTube and dropping it into a staged kidnapping call. The Hiya voice fraud intelligence team has catalogued hundreds of near-identical scripts aimed at parents and grandparents across six languages. Daisy addresses this by consuming attacker time at the top of the funnel, not by stopping the specific cloning technique. A grandparent whose number the scammer never dials because Daisy was dialed instead is safe for that call cycle. The honest limitation is that Daisy is a volume defense rather than a precision defense, which matters when the question is personal rather than aggregate. A fuller defense will combine Daisy-style time wasting with carrier-level voice authentication for the inbound call itself.

Why Older Adults Are the Primary Target Audience

Shifting from the attacker side to the victim side, elderly adults sit at the intersection of trust, assets, and under-protected phone lines. The FBI 2024 Elder Fraud Report recorded $4.885 billion in losses from Americans aged 60 and older, more than any prior year. The same report shows that 101,068 complaints came from that cohort, with an average per-victim loss north of $48,000. The United Kingdom Action Fraud bulletins tell a similar story, with pensioners accounting for the largest single tranche of reported scam losses. Those numbers justify a product aimed at protecting grandparents through a scalable carrier-level defense that absorbs attacker time at the network edge. Daisy is explicitly cast as an elderly woman because she needs to resonate with the voice the attacker expects to encounter.

A second reason the elderly bear the heaviest cost is that they tend to answer the phone when a younger adult sends the call to voicemail. Hiya reports that older respondents score phone trust 25 to 30 points higher than respondents under 35 on a 0 to 100 index. The scammer’s advantage is not a lack of intelligence in the target but a cultural habit around answering the ringing phone. The habit is slow to change in cohorts that grew up without caller ID or spam filters. Any defense therefore has to meet the elderly person at the moment they pick up, which is where Daisy and carrier-side screening converge. The two mechanisms are compatible rather than competitive, which is useful because no single defense is enough.

A third dimension is isolation, because many elderly victims live alone and treat the phone as a social channel rather than a transactional one. That changes the fraud script for attackers who open calls with warmth and sympathy. The attacker opens with sympathy and warmth rather than urgency, and the warmth lands harder on a lonely target. This is one of the reasons the guide to protecting your family from AI threats emphasizes conversation rather than lockdown. A grandparent locked out of the phone system loses regular contact with grandchildren, and that outcome remains socially unacceptable in every market. A defense absorbing attacker time on a separate honeypot line is the first solution that does not punish the person it protects. This asymmetry is the subtle reason Daisy earned the goodwill she did at launch.

Moving on from victimology to the courtroom, the central legal question is whether an AI pretending to be a person can lawfully record a stranger. In the United Kingdom, the Data Protection Act 2018 treats a scammer’s voice as personal data and requires a lawful basis for processing it. Virgin Media O2 has argued its lawful basis is fraud prevention, a category the Act explicitly permits for carriers acting against criminal activity. The Act does not require two-party consent for a recording that falls under that basis, which is why Daisy can run inside UK law without a wiretap warrant. The same logic underpins the Jim Browning collaboration, since consumer scambaiters have operated under the same prevention argument for a decade. The regulator Ofcom has not challenged the Daisy project, which the carrier reads as implicit comfort with the framing.

Looking across the Atlantic, the United States has a patchwork of two-party and one-party consent recording rules that vary by state. Twelve US states listed in the MWL state recording consent chart require all parties to consent to a recorded call. A Daisy-equivalent in the United States would need a one-party state posture or an explicit state fraud exemption. The Federal Trade Commission’s guidance on chatbots, deepfakes, and voice clones is largely silent on the defensive case, since it was written to deter deceptive sellers, not deceptive defenders. That silence is useful because it leaves room for a carrier to argue the defense is categorically different from the attack. It is also uncomfortable because silence is not endorsement, and a state attorney general could still bring a case.

Moving on to the European Union, the AI Act framework classifies a chatbot that impersonates a human as a limited-risk system requiring transparency to the end user. The attacker in a Daisy conversation is not the end user the AI Act was written to protect, which creates another gray area the regulators have not resolved. Civil society groups such as the Open Rights Group have warned that defensive AI deception may still require a disclosure framework. The counter-argument is that disclosing the deception to a scammer defeats the whole purpose and would collapse the product. This is one of the open regulatory questions that will shape whether Daisy-class products can scale beyond their current PR footprint. A sensible compromise is a published policy at the carrier level that regulators can audit without disclosing it in real time.

A fourth live issue is the automated dialing framework, since in the United States the FTC Telemarketing Sales Rule and the TCPA regulate robotic calls. A Daisy that only receives calls and never initiates them sits outside the regulated class because the Rule targets outbound dialers. A future Daisy that calls scammer numbers back to extend the honeypot would fall inside the Rule. It would need prior express consent, which is impossible to obtain from a criminal party. This is why every carrier running an AI granny deployment keeps the bot strictly inbound today. The AI regulation picture will evolve, but the inbound-only posture is a durable design choice. Any outbound future will come from law enforcement partnerships rather than carrier products.

Ethical Lines Between Deterrence and Deception

Shifting from law to ethics, there is a real argument that any AI pretending to be a person violates a baseline rule of honest machines. That rule is often associated with the AI community’s informal norm against bots that fail to disclose themselves as bots when asked directly. Daisy fails that disclosure test by design, which is uncomfortable for the subset of ethicists who treat the norm as near-absolute. The counter-position is that criminals deceive their victims and that mirror-image deception is a legitimate defensive act. Most Western criminal law accepts a parallel for undercover policing, in which a human officer pretends to be someone else to catch a criminal. The ethical arguments for Daisy are the ethical arguments for undercover policing scaled by a voice model.

A second ethical dimension is harm to incidental third parties, since not every call that reaches Daisy is actually a scammer. A wrong-number dial from a neighbor or a nervous first-time cold caller might find themselves in a long conversation with a confused persona. Daisy’s current scope is small enough that these edge cases are rare, but a scaled version must build a graceful exit for them. A simple rule is that if a caller signals they are not a scammer, the bot can say thank you and hang up without further manipulation. Virgin Media O2 has not published the exact rule set, which is a transparency gap civil society groups will fill with pressure over time. The ethical case is strongest where the bot refuses to deceive anyone it has reasonable grounds to believe is innocent.

Consumer Tools That Mimic the Daisy Playbook

Moving on from carrier infrastructure to consumer software, Daisy has three main predecessors in the consumer scambaiting market. Lenny is a scripted bot from the mid-2010s that plays recorded audio of a confused old man and keeps telemarketers on the line for roughly 10 minutes on average. Jolly Roger Telephone is a subscription service that attaches a similar library of bots to a consumer phone line for about $24.99 per year. Both of these predate the modern voice stack, which is why their script library is finite and their responses are not generative. They still work surprisingly well, which is a reminder that most phone scammers do not probe their mark very carefully. A generative Daisy-class consumer product will probably triple average call duration against the Lenny baseline when it reaches the market.

A newer generation of apps uses generative models directly on devices on the handset and promises Daisy-grade persuasion for mass market users. Products from Reality Defender, Nomorobo, and Hiya already combine carrier-level screening with consumer-facing dashboards for scam call volume. None of them currently ship a true conversational honeypot, but every roadmap we have read for 2026 includes one. The consumer market gap is a Daisy that defends a real line rather than a honeypot, and it is the next billion-dollar segment inside the broader scam defense category. The engineering work is modest, but the legal work around two-party consent is still unresolved, which is why no incumbent has shipped yet. Expect a startup to ship first in a one-party consent state and then push for a federal exemption on the strength of early results.

For a consumer who wants protection today, the pragmatic stack is a scam-filtering app recommendations, a published scambaiter phone number, and a family rule about verifying urgent requests. The AIplusInfo family protection guide and the voice cloning risks explainer both suggest the family rule as the single highest-ROI defense. A pre-agreed code word among family members defeats the vast majority of grandparent scam attempts even without any technology. The reason Daisy exists is that most families will not reliably install the rule, so the carrier has to shoulder the structural defense. Both approaches work together, which is why the strongest households use both. A code word is free; a Jolly Roger subscription costs less than a lunch; both are cheaper than the average $1,298 senior loss per scam incident.

Operational Risks for a Carrier Running Daisy at Scale

Shifting from product to operations, the biggest risk a carrier bears is a hallucinated reply that leaks sensitive information or produces reputational harm. An LLM told to improvise can wander into territory the brand would never approve in a human script, which is why every production voice bot sits under constant review. A second risk is scammer retaliation, since a motivated attacker who realizes they are being baited can try to flood the honeypot line with denial-of-service calls. A third risk is drift in the voice model, which can start sounding synthetic over time if the vendor retrains the base voice. Each of these risks is manageable with the same monitoring stack carriers already run for billing and service quality. None of them is unique to AI scambaiting; they are the normal operations surface of any always-on voice agent.

A less obvious risk is political, because a carrier that becomes a public anti-scammer brand will be asked to do more by regulators and civil society. The press cycle around Daisy has already produced calls for Ofcom to require every UK carrier to run a similar honeypot by 2027. Virgin Media O2 benefits in brand equity today, but it absorbs policy risk because its visible success defines the baseline others may be forced to match. A smaller carrier may lack the research budget to run the same product and may oppose any such requirement in regulatory consultation. The carrier that goes first always sets a floor it will later lobby either to raise or to leave alone depending on competitive dynamics. Daisy is therefore as much a policy intervention as a technology deployment.

How Regulators Are Starting to Respond to Voice Fraud

Moving on from carrier risk to the regulators themselves, the Federal Communications Commission has been the most active voice-fraud agency in the United States. The FCC’s robocall mitigation program requires carriers to implement STIR and SHAKEN caller ID authentication across all US-originated voice calls. Compliance has improved spoofed-caller detection, but the attack surface shifts to internationally originated calls that bypass the US framework. The FCC’s 2024 enforcement actions totaled more than $300 million in proposed fines against telemarketers and robocall operators. Fines of that scale have not meaningfully reduced the number of spam calls that reach American consumers each week. The gap between enforcement effort and consumer experience is the political fuel behind any proposal to require defensive bots at the carrier level.

Shifting to the United Kingdom, Ofcom publishes a scam calls and messages advisory page and has approved a code of practice for carriers on fraud prevention. The code is voluntary rather than mandatory, which is why carriers like Virgin Media O2 have more freedom to experiment with products like Daisy. If the code becomes compulsory in the next review cycle, a Daisy-type deployment could become a de facto requirement for larger carriers. The Online Safety Act 2023 does not directly cover voice fraud, but Parliament has signaled interest in extending its scope. The regulatory direction of travel in the UK is clearly toward making the carrier the first line of defense rather than the consumer. That direction aligns the regulator with Virgin Media O2’s product narrative, which is one of the quiet wins of the Daisy campaign.

The European Union’s AI Act takes a different approach by regulating the AI model layer rather than the carrier infrastructure. The Act requires transparency when a system interacts with a human, which creates the open question discussed earlier about defensive deception. A separate strand of EU consumer protection law would catch scammers themselves, but enforcement across non-EU jurisdictions remains weak. The practical result is that the European market will likely converge on a disclosed-honeypot model, in which the carrier publishes its Daisy policy without disclosing it on each call. That compromise would let Daisy-class products operate while giving regulators a documented policy to audit. A policy audit is a far lighter touch than a per-call disclosure requirement that would collapse the product.

The Future of AI Granny Bots Beyond 2026

Looking ahead from today’s playbook to tomorrow’s roadmap, the obvious direction is personalization at the per-subscriber level. A future Daisy would learn the voices of a subscriber’s own grandchildren so that a cloned-voice attack is caught at the moment the carrier screens the inbound call. That requires a voice embedding security guide stored locally on the subscriber’s device and a privacy regime that neither the subscriber nor the regulator currently expects. The privacy framework around AI will need to catch up before per-subscriber voice modeling can ship at scale. The upside is real, since a per-subscriber defense can intercept the single most dangerous attack vector before the victim ever answers the call. Expect pilots in 2027 or 2028 from carriers in Nordic countries that already have a stronger digital identity infrastructure.

Looking ahead further out, the regulatory and ethics debate will probably settle around a standard honeypot license that carriers hold with their national regulator. That license would allow defensive deception within a published policy and would compel carriers to publish aggregate telemetry on attacker patterns. Shared telemetry would then feed a cross-carrier scam intelligence commons that Hiya, Nomorobo, and others could draw from in real time. The structural goal is to turn every Daisy call into a defense asset for every other carrier’s subscribers, not just the host carrier’s subscribers. This is the pattern that credit card fraud detection follows across issuer banks, and the same pattern will likely win in voice fraud too. The payoff is a slower attacker, a more informed defender, and a smaller gap between the two over time.

A final direction is the attacker arms race, because voice attackers will not stand still while carriers deploy Daisy. The next generation of scam calls will combine voice cloning with aggressive LLM steering that detects when the counterparty is a bot and hangs up early. Daisy will need to pass a Turing-grade test to keep working, which is why the next voice model release cycle matters more to this product than any policy change. Expect quarterly persona refreshes, prompt tuning on live scam logs, and voice variation strategies to counter attacker heuristics. The arms race is not winnable in a final sense, but it is a race that the defenders can keep pace with if they invest. That is the serious reason Virgin Media O2 keeps using the word “team” when talking about Daisy, because a defensive product that stops improving stops working.

Looking ahead, policy observers expect carriers to publish annual honeypot telemetry reports to regulators by 2027. Those reports would include attacker call volume, average conversation length, and documented scam scripts captured per quarter. Regulatory audits would then cross-check carrier claims against aggregate data from Hiya and other industry sources. The result could be a lightweight compliance regime that rewards carriers running Daisy-class products without forcing smaller operators to build their own. That incentive structure would preserve experimentation while locking in a baseline of defensive investment across the industry.

What Americans Say About AI Scam Calls, 2026

Share of US adults responding to Hiya’s State of the Call survey on AI voice fraud exposure and policy preferences.

Deepfake voice calls received by Americans
25%
Unsure they could tell AI voices apart
24%
Americans switching providers for scam protection
38%
Americans supporting stronger scam regulation
72%
Americans wanting carrier scam liability
67%

Source: Hiya State of the Call 2026 survey, US respondents.

Key Insights on the AI Granny Strategy

A clear pattern emerges when the headline figures sit beside the campaign metrics. Carrier-side AI defenses can absorb attacker time at a fraction of the cost the attacker pays. The economic case for an AI granny is strong even before the brand equity line arrives. Elderly consumers remain the largest single victim class, which is why the persona choice matters as much as the model stack. The next phase of this market will test whether the carrier can protect the subscriber’s own line without triggering the two-party consent trap. That question will decide whether Daisy stays a PR instrument or becomes a per-line defense in the second half of this decade.

DimensionScripted bot (Lenny)Jolly RogerAI granny DaisyCarrier screening (Hiya)
Transparency to end userNoneLowPolicy page onlyDashboard
Participation of consumerOpt-in forwardOpt-in forwardHoneypot onlyCarrier auto screen
Trust impact on grandparentNeutralNeutralProtects by proxyProtects at line
Decision making layerFixed scriptsScripts and libraryLLM in real timeRules and signals
Misinformation riskLowLowModerateLow
Service delivery modelDevice appSubscription $24.99 per yearCarrier campaignCarrier plus app
Accountability chainUserVendorCarrier plus agencyCarrier plus vendor

Real-World Examples of AI Granny Deployments

Virgin Media O2’s Daisy in the United Kingdom

Virgin Media O2 built Daisy with the London creative shop VCCP and the carrier’s AI agency Faith, launching her in November 2024 inside the Swerve the Scammers campaign. The build cycle ran three months end to end, using a dedicated honeypot number seeded with suspected scammer target lists sourced from YouTube scambaiter Jim Browning. Virgin Media O2 reported Daisy kept scammers on calls for more than 40 minutes, with hundreds of recorded sessions in the first six weeks. The data appears in The Guardian’s November 2024 feature on Daisy. The campaign drew press coverage in more than 40 countries and became the single most awarded telecommunications fraud project at Cannes Lions 2025. A real limitation is that Daisy is a honeypot and does not currently defend an individual subscriber line against the same attackers.

Jim Browning’s Human Scambaiting Lineage

YouTube scambaiter Jim Browning built a human equivalent of Daisy over more than a decade and provided tactical consulting to Virgin Media O2 during the Daisy build. Browning’s channel documents individual sessions in which he kept scammers on the line for 90 minutes or more and recovered specific victims’ money in the process. His most-viewed investigation shut down an Indian call center of roughly 70 operators after a year of filmed documentation, covered by the BBC in its November 2024 Daisy feature. The limitation of the Browning model is that it depends on one person’s working hours and does not scale beyond a few sessions per week. Daisy is the industrialized version of the Browning technique, with the same persona heuristics applied at a volume no human could match.

The Hiya Honeynet of Scam Call Telemetry

Hiya built and operates a global honeynet of inbound numbers and deployed voice fraud intelligence sensors across six countries, feeding real-time scam call signals to carrier partners worldwide. The Hiya voice fraud intelligence team reported AI-generated scam calls climbed more than 60% year over year across its covered markets. The United States and United Kingdom both sit at the top of the growth list. The honeynet does not run a conversational bot like Daisy, but it provides exactly the kind of attacker telemetry a Daisy program needs to tune its personas. The limitation of Hiya’s data is that it reflects known attacker scripts and under-samples bespoke targeted attacks aimed at specific wealthy victims. Combined with a Daisy-style conversational system, this telemetry becomes a near-complete picture of the voice fraud economy.

Learn More: Books on Scam Defense and the AI Economy

Three verified titles that complement the Daisy playbook, each from major publishers.

Scam Me If You Can: Simple Strategies to Outsmart Today's Rip-off Artists

Scam Me If You Can: Simple Strategies to Outsmart Today’s Rip-off Artists

By Frank Abagnale

Primer on phone, mail, and online scam tradecraft from a world-renowned fraud expert, aligned with the Daisy threat model.

Buy on Amazon
The Confidence Game: Why We Fall for It...Every Time

The Confidence Game: Why We Fall for It…Every Time

By Maria Konnikova

Deep behavioral analysis of why social engineering works, which is the exact mechanism Daisy counters by jamming the attacker script.

Buy on Amazon
The Age of Surveillance Capitalism

The Age of Surveillance Capitalism

By Shoshana Zuboff

Explains the data economy that makes voice cloning scams possible and names the broader policy stakes behind defensive AI systems like Daisy.

Buy on Amazon

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

Case Studies From the Fraud Prevention Front Lines

Case Study: Virgin Media O2 Daisy Launch Metrics

The problem facing Virgin Media O2 was rising AI-generated voice scams during 2024, with internal customer-care telemetry showing a 30% year-over-year rise in scam call complaints. The solution was Daisy, a bot built on a standard three-stage voice pipeline with a persona prompt tuned for maximum conversational drift and seeded with Jim Browning’s scam number library. The first six weeks captured more than 1,000 attacker sessions, with average call length above 15 minutes. Press coverage reached 40 countries, per the WARC Cannes case study on Daisy. The project also absorbed internal brand equity worth an estimated $12 million in equivalent paid media across the first three months. Those metrics placed Daisy among the ten most-awarded telecommunications campaigns globally in 2025.

A real limitation is that Virgin Media O2 has not published a controlled estimate of attacker losses prevented by Daisy, so the ROI claim rests on hours-wasted rather than dollars-saved. A second controversy involves the question of whether elderly subscribers themselves should be informed when a Daisy campaign runs in their market. The carrier has argued that visible press coverage amounts to disclosure, which satisfies the policy audit requirement without compromising the honeypot. The remaining open question is whether Ofcom will mandate a Daisy-type deployment across all UK carriers in its next code of practice review. A mandate would reshape the UK voice fraud market overnight and set a global template for similar rules abroad.

Case Study: Lenny the Open Source Scambaiting Bot

The problem the Lenny project set out to solve in the mid-2010s was telemarketing fraud targeting Australian consumers who had no carrier-level screening at the time. The solution was a scripted bot that plays pre-recorded audio clips in response to telemarketer dialogue, keeping callers on the line with a confused elderly persona throughout. A decade of hobbyist operation produced an average call duration of roughly 10 minutes on the Lenny archive and a widely circulated library of 100,000+ sample calls across YouTube. Lenny inspired at least three commercial products including Jolly Roger and seeded the UX conventions Virgin Media O2 adopted a decade later. The archive remains freely accessible, which keeps the project useful for researchers studying scam call patterns today.

The Lenny model fails because its scripts are fixed, so scammers who have heard the clips once hang up early in a repeat call. A second limitation is that Lenny only runs on a forwarded line, which requires the consumer to configure the forward on their own carrier plan. The controversy around Lenny concerned recording consent in two-party consent jurisdictions, which never produced a successful prosecution against a consumer who deployed it. Lenny demonstrates that attacker time wasting is a durable defensive strategy even with a tenth of the technology Daisy now uses. The lesson for product teams is that persona engineering matters at least as much as the voice stack behind it.

Case Study: The FTC and UK FCA Elder Fraud Push

The problem US and UK consumer protection agencies face is elder phone fraud losses that have risen for a decade despite steady enforcement spending. The FBI’s 2024 Elder Fraud Report recorded $4.885 billion in losses from Americans over 60, with 101,068 complaints and an average per-victim loss near $48,000. The implemented solution has combined public awareness campaigns with carrier-level STIR and SHAKEN mandates, backed by proposed fines of more than $300 million from the FCC in 2024. The measurable impact so far is a slower rise in victim counts but no measurable reduction in dollar losses, since attackers now target wealthier victims per call. Agencies have shifted toward tech-forward defenses in response to these diminishing enforcement returns.

A real limitation is that most scam call originations now sit outside the jurisdiction of either agency, which blunts enforcement actions against the operators themselves. This is why both the FBI and the FCA have signaled interest in defensive carrier products that absorb attacker time directly, including bots along the Daisy pattern. The controversy is political because any mandate to run a defensive bot imposes a cost on smaller carriers that may not have the research budget to build one. Expect regulators in both markets to publish defensive-bot guidance in 2026 and 2027 as the Daisy model matures. The resulting rules will likely include telemetry sharing and policy audit requirements rather than per-call disclosures.

Frequently Asked Questions About the AI Granny Playbook

What is the AI granny chatbot called Daisy?

Daisy is Virgin Media O2’s AI granny chatbot that answers suspected scam calls. She was built with VCCP and the AI agency Faith as part of the carrier’s Swerve the Scammers campaign in November 2024. She keeps fraudsters on the line with rambling stories so they waste time instead of defrauding real elderly subscribers.

How does the AI granny keep scammers on the phone?

Daisy uses a three-stage voice pipeline that combines automatic speech recognition, a large language model, and a text-to-speech voice. The LLM is given a persona prompt that instructs her to drift off-topic and share irrelevant stories. The result is attacker call durations that have reached more than 40 minutes in a single session.

Is an AI granny chatbot legal to deploy?

In the UK the Data Protection Act 2018 permits recordings made for fraud prevention, which is the legal basis Virgin Media O2 uses. In the United States legality depends on the state, since two-party consent states would require every call be disclosed. The EU AI Act leaves defensive deception as an open question that regulators have not yet closed.

Who built Daisy and when was she launched?

Virgin Media O2 built Daisy in collaboration with the London creative agency VCCP and the carrier’s in-house AI agency Faith. The project launched in November 2024 as part of the Swerve the Scammers campaign. Jim Browning, a scambaiting YouTuber, provided tactical input on persona and target number lists.

How much phone fraud does the AI granny prevent?

Virgin Media O2 has published hours-wasted rather than dollars-saved as its headline metric. The underlying math suggests tens of thousands of attacker minutes consumed per month translate into real revenue denial for scam call centers. A per-dollar estimate is not yet published by the carrier.

Can consumers deploy a Daisy-style bot on their own line?

Not directly, but several consumer tools attempt the same effect. Lenny is a free scripted bot that users can forward calls to, and Jolly Roger offers a subscription at roughly $25 per year. A true generative AI granny for a single consumer line is still blocked by two-party consent rules in several US states.

Does the AI granny protect elderly users specifically?

Yes, Daisy was explicitly designed to protect elderly subscribers by consuming attacker time before the attacker reaches a real target. Older adults bear the heaviest fraud losses, with the FBI reporting $4.885 billion in elder fraud losses in 2024. The persona choice is a direct response to scammer targeting data.

What are the risks of using an AI granny chatbot?

The main risks are hallucinated replies that leak information, scammer retaliation against the honeypot number, and voice model drift over time. A carrier running Daisy also absorbs policy risk because its visible success defines a baseline others may be forced to match. Each of these risks is manageable with monitoring and periodic prompt tuning.

Will the AI granny chatbot work against voice-cloning scammers?

Only partially, because the AI granny consumes attacker time rather than detects cloned voices. The future product category combines Daisy-style time wasting with carrier-level voice authentication at the inbound call. A per-subscriber voice embedding stored on the device is the most promising path to catch cloned-voice attacks early.

Why has the AI granny chatbot received so much press?

Daisy earned coverage in more than 40 countries because she combines a culturally resonant persona with a legitimate fraud prevention story. The press cycle also reflected a public appetite for revenge against scammers that human scambaiters have cultivated for a decade. Virgin Media O2 turned the appetite into a product and a case study.

How is the AI granny different from carrier call screening?

Carrier call screening blocks or labels an incoming call before the subscriber answers, while an AI granny answers it in character to waste time. The two approaches are complementary rather than competitive, since screening cuts volume and the granny depletes attackers who get through. The strongest consumer defenses will combine screening and conversational honeypots together.

Is the AI granny chatbot biased against legitimate callers?

Daisy is only reachable through a dedicated honeypot number, so legitimate callers almost never land in a conversation with her. The edge case of a wrong number can produce an awkward exchange, but Virgin Media O2 has described a graceful exit for callers who signal they are not scammers. A larger deployment would need a published policy for the edge cases.

What comes next for AI granny chatbots after Daisy?

The next direction is per-subscriber personalization, in which the bot defends a specific line rather than a honeypot number. Shared telemetry across carriers will probably become the mid-decade norm, feeding a cross-carrier scam intelligence commons. Regulatory frameworks will likely converge on a honeypot license model with published policies and audited aggregate data.