Is AI Cold Calling Legal in the UK?
Ask the pages that rank for this question and you get two opposite answers. One says AI cold calling is "generally illegal" in the UK. Another says AI voice agents "are not prohibited in the UK". Both are written by companies selling calling technology, and neither quotes the law it is summarising.
The law is the Privacy and Electronic Communications Regulations 2003, and the whole question compresses into six words in regulation 19: "sounds which are not live speech." Whether an AI cold call is banned outright or merely heavily restricted depends on which side of those words a real-time generated voice falls, and as of today no regulator, court or tribunal has decided.
Before the walkthrough, our interest declared: DeskCaller answers phones and makes no outbound calls, a position our FAQ states plainly. We are not selling you outbound AI, so we have no reading of this statute to defend. What follows is what the instrument, the ICO's guidance and the enforcement record actually say, quoted, with links, and with the unsettled part marked unsettled.
The short answer
For a cold call, the practical answer is that the legal route is close to a null set, whichever way the open question resolves.
If an AI call is caught by regulation 19, it needs the recipient's prior consent, and the ICO's bar is specific: "General consent for marketing, or even consent for live calls, is not enough – it must specifically cover automated calls." A cold call is by definition a call to someone who has given no consent at all. Caught by reg 19, AI cold calling is simply prohibited.
If an AI call escapes reg 19 as "live speech", it is still a live marketing call under regulation 21: no calls to any number on the TPS or CTPS registers, caller line identification presented, identity disclosed, your own do-not-call list honoured. That is the best case, and it is the regime that produced a £160,000 fine for one energy broker this year.
And the ceiling for getting it wrong moved. Since 5 February 2026, breaches of the marketing regulations carry penalties up to £17.5m or 4% of worldwide turnover, with personal liability for directors where a breach happened with their "consent or connivance" or through their neglect.
The six words, in context
Regulation 19(1), current text as at 9 September 2026:
"A person shall neither transmit, nor instigate the transmission of, communications comprising recorded matter for direct marketing purposes by means of an automated calling or communication system except in the circumstances referred to in paragraph (2)."
Two load-bearing phrases. The ban attaches to "recorded matter". And regulation 19(4) defines the system it covers:
"an automated calling system is a system which is capable of— (a) automatically initiating a sequence of calls to more than one destination in accordance with instructions stored in that system; and (b) transmitting sounds which are not live speech for reception by persons at some or all of the destinations so called."
That definition was written in 2003, when the only things a machine could say down a phone were recordings. The ICO's guidance runs on the same binary. Its live-calls guidance, on the live page as of 9 September 2026: "A live call is therefore a telephone call where a live person is speaking to the person they are calling", as against an automated call, "made by an automated dialling system that plays a recorded message".
A real-time LLM voice agent is neither of those things. It plays no recording; there is no "recorded matter". It is also not a live person speaking. It generates speech that did not exist until the moment of the call, which is a category the instrument does not contemplate.
Two further wrinkles, for completeness. Reg 19(1) was amended in 2011 to cover an "automated calling or communication system", but only "automated calling system" is defined anywhere in the instrument; we pulled the interpretation regulation and the word "automated" does not appear in it. And the EU directive this regulation implemented scopes the rule to systems operating "without human intervention", words the UK regulation omits. An LLM agent has no human intervention, so even the narrower European wording offers it no shelter.
The most direct statement we found that the question is open comes from the data protection consultancy Handley Gill, writing in October 2025, and we quote it as analysis rather than law: "it remains to be seen... whether the use of more sophisticated AI software which is trained on content but does not involve pre-recorded information could be considered to constitute live speech and therefore to fall outside the scope of the restriction under PECR."
The safe reading is to treat an AI voice call as a reg 19 automated call and obtain specific consent. That is a risk position, not settled law, and anyone who tells you it is settled is selling something.
What the "AI calling" fines actually punished
In September 2025 the ICO announced £550,000 of penalties against two energy-sector firms in what the SERP now routinely cites as AI-calling enforcement. Read the penalty notices and something more interesting appears: there was no AI in either case.
Both firms used avatar software. In the ICO's words, recipients thought they were talking to "'Jo, Helen or Ian' from the UK", but the voices "were in fact scripted lines recorded by voice actors and played by call agents abroad". A human in an overseas call centre pressed buttons to fire pre-recorded clips. Home Improvement Marketing was fined £300,000 for 2,449,380 such calls. Green Spark Energy was fined £250,000 for 9,587,050 of them, generating 497 complaints, including one recording in which the recipient asks whether the caller is a robot.
Two details in those notices matter for the AI question. First, the calls were interactive: one complainant described how "the recorded software responds based on a recipient's answers". Interactivity did not make them live calls. The audio was recorded, so reg 19 applied, and that was the end of it. Second, both fines were issued near the ceiling of the old £500,000 penalty regime, because the conduct predated the new one. The same conduct today faces a ceiling thirty-five times higher.
Neither notice addresses generated speech, because neither case involved any. The enforcement record on actual AI voice is empty. That is the gap the vendors fill with confidence.
The rest of the 2024 to 2026 run-rate shows the live-call regime working as designed: £240,000 and £100,000 against two lead-generation firms in April 2024, £90,000 against AFK Letters in 2025, £160,000 against EPDL in May 2026 for over 700,000 calls to TPS and CTPS numbers, one of which opened "No, this is not a sales and marketing call, I just wanted to share our electricity prices with you", and £190,000 against Elderly Aids in August 2026 for 758,053 calls to TPS-registered numbers, cold-selling call blockers to elderly people.
What the vendors tell you, and where it came from
Researching this piece we traced one sentence across at least three vendor compliance pages, near-verbatim: that "ICO informal guidance" suggests two-way conversational AI "may be treated more like a live call". No such ICO guidance exists. We searched for it directly and adversarially, on the live ICO site and in archives. One vendor asserts the exact opposite, that the ICO "classifies AI-generated conversational voice systems as automated calls", with equally no source. Both certainties are invented, and both are copied between sites that cite each other.
The ICO's live pages, refreshed on 28 April 2026 and carrying a banner that the guidance is under review following the Data (Use and Access) Act, still describe exactly two categories: a live person speaking, or a recorded message. No AI category exists anywhere in them.
Three other patterns worth knowing before you trust a compliance page:
The September 2025 avatar fines are widely cited as AI enforcement, usually with the wrong total. Several pages still state the pre-2026 £500,000 penalty ceiling as current. And a striking amount of "UK compliance" content is American law with the labels left on: one vendor's compliance API literally namespaces its guardrails tcpa:, after the US statute; another sells the TCPA's "8am to 9pm" calling window as universal compliance; a London-addressed AI calling vendor offering to ring "thousands of leads per day" describes its compliance posture entirely in terms of GDPR, without once mentioning PECR or the TPS, which are the rules that actually govern the calls.
Even lawful outbound burns the number
Suppose the open question resolves in favour of AI callers, consent is somehow obtained, both registers are screened, and every reg 21 duty is met. The extra-legal costs remain, and they attach to the phone number itself.
The UK has no STIR/SHAKEN and no remediation playbook. Ofcom shelved CLI authentication, so the American advice that dominates search results for "number flagged as spam" does not transfer. What the UK has instead is analytics gatekeepers such as Hiya inside BT and EE's Digital Voice, and now handset-level AI screening: Pixel Call Screen has run in the UK since late 2024, and iOS 26 answers unknown callers with a screening bot by default. One outbound vendor now publishes scripts for talking its AI callers past Apple's bot, which is a sentence that describes the arms race better than we could.
Disclosure makes it worse in one direction only. On inbound calls, an agent announcing it is an AI reduces hang-ups in the one large analysis we found, an American vendor's study of roughly 450,000 calls. On outbound, a survey found 22% of people hang up immediately on hearing it. Inbound and outbound are different products wearing the same technology.
And the recipients are learning. MoneySavingExpert threads this year document AI-voiced "energy advisor" cold calls, "Gabby" and "Emma Kirby", warming leads for human closers. A cold call from an AI agent to a Hacker News user became a public incident for the company that sent it when the bot hung up on being asked for a human. A developer who received an AI cold call prompt-injected it, extracted its system prompt, and booked hundreds of fictitious pirate-themed parties into the deploying business's calendar, at a reported cost to himself of about £0.08.
The commercial record is not better. The flagship of the category, Air AI, was sued by the US FTC in 2025 and settled in March 2026 with an $18m judgment, mostly suspended, and lifetime business-opportunity bans for its principals; its domain now hosts an unrelated defence company. Those are American proceedings under American law, but the product they concern is the one whose UK legality this page is about. We could not find a single named UK business publicly operating a successful outbound AI voice campaign. The only named UK deployments the record shows are the two firms the ICO fined.
What is genuinely unsettled
Three things this post will not claim, because nobody can.
Whether a real-time generative voice call is an "automated call" under reg 19 is undecided in both directions. No ICO document, enforcement notice or court decision resolves it. Treating it as one is the safe reading, not the settled law.
The ICO has issued no guidance, formal or informal, classifying conversational AI calls. Every vendor sentence claiming otherwise is untraceable. What does exist is a deadline for change: since May 2026 the ICO is under a statutory duty to produce a code of practice on AI and automated decision-making, which is the likeliest vehicle for an official answer.
And UK law does not currently require a caller to disclose it is an AI. The pressure is indirect: reg 24's identity duties, UK GDPR fairness, and the avatar penalty notices treating the impersonation of humans as an aggravating feature. The EU AI Act's disclosure duty is EU law, not UK. The wider disclosure picture is in our disclosure guide.
The line the law actually draws
The precise thing UK law restricts is outbound marketing. A purely administrative call, in the ICO's words, does not count as direct marketing if used "purely for administrative purposes", and the qualifier is the whole rule: "if you include promotional content within your service call, then the call counts as direct marketing." One upsell sentence converts the call. The same knife-edge exists for texts, which is why the missed-call text-back business is harder than it looks, and there is no soft opt-in for calls at all: that concession exists for email only.
TPS registration beats customer status, a point that surprises almost everyone: the ICO's guidance says you must not call a TPS-registered number "even if they are an existing customer (unless the call is in relation to pension schemes and you meet a strict criteria, see below)". Businesses get the same protection, which surprises the rest: B2B callers must screen the CTPS as well, and sole traders count as individuals.
So the honest map is this. Cold calling by AI: prohibited on one reading of six words from 2003, restricted to the point of impracticality on the other, with a 35x-higher penalty ceiling either way and the regulator's AI code still to come. Service calls: lawful, on a line thin enough that one sentence crosses it. Inbound: none of the above applies, because the caller rang you.
We built on the third of those. Not because outbound is unbuildable, but because a small business should not have to hold an unresolved question of statutory interpretation between itself and a £17.5m ceiling. The screening rules for what reaches your line are in the spam screening ruleset, and what our agent does instead of ringing anyone is in the FAQ.