The choice takes less than a minute. A dropdown, six to ten voices with first names attached, a play button beside each one. You listen to three, pick the one that sounds least strange, and move on to something that feels more important.
That is how almost every company selects the voice its customers will hear several thousand times a year.
The claim I want to defend here is that this choice is ceasing to be a setting and becoming a purchase – something with an origin, a contract and, before long, a technical obligation attached to it. Not because voices are getting more expensive, but because three questions that have been left open until now are closing at roughly the same time.
Part one: the voice is already an asset to the people selling it
The first signal comes from the supplier side, and it should be read with the scepticism it deserves.
Voices, a marketplace that represents voice actors and sells licensing of human voices for AI use, published its Amplified 2026 report in late January, based on a Censuswide survey of 700 business leaders and consumers. In it, 79 percent of leaders say inauthentic AI voices damage brand perception, the same share say they prefer voices from real, attributed actors over purely synthetic ones, and 77 percent say exclusive, brand-specific licensing is critical to standing out. Among consumers, 76 percent say they expect transparency about how AI voices are created and licensed.
The company's business depends on that answer being the right one. The sample is small, it mixes leaders with consumers, and questions about what people "expect" tend to produce high numbers regardless of what you ask. The figures prove nothing.
They are worth reading as a signal about where a market is heading. When an entire supplier category starts selling provenance – where the voice came from, who consented, what the licence covers – it is because somebody has started asking. And the first person to ask is rarely the caller. It is the procurement lead at a larger company who has to put their name on a contract.
This is where the Nordics differ, and it is a structural observation rather than a measured one: the market for Norwegian, Swedish and Danish voices is thinner than the English one. The consequence is that two competitors in the same sector stand an unusually good chance of ending up with the same voice from the same list, and that "which voice is this" has considerably fewer possible answers here than it does in English. That makes the question easier to answer. It does not make it less worth asking.
Part two: from December, the audio has a legal property
The second part has a date on it, and the date is close.
Article 50 of the AI Act was left untouched when the EU adopted its simplification package this summer. Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, was signed on 8 July 2026 and entered into force on 27 July. It pushed the high-risk deadlines far out – the Annex III obligations now apply from 2 December 2027 – but left the transparency rules exactly where they were.
Two of them reach voice agents, and they reach two different parties.
| Obligation | Who it binds | From when |
|---|---|---|
| Article 50(1): disclose that this is AI | The deployer – that is, the business that owns the phone number | August 2026 |
| Article 50(2): machine-readable marking of synthetic audio | The provider of the generative system | 2 December 2026 for systems already on the market on 2 August 2026. No transition period for systems placed on the market after that date |
Most people recognise the first one by now: the disclosure duty that has to be met in the opening line. The second is far less well known, and the most important thing about it is that it is not your obligation.
Which is precisely why it is worth understanding. Article 50(2) requires the provider of the generative system to mark synthetic audio in a machine-readable format that allows it to be detected as artificially generated. That is the supplier's responsibility, not yours. But it is your audio, on your number, and a supplier who has not solved this by their own deadline has a problem that becomes yours the moment somebody asks.
The Commission's code of practice on marking AI-generated content is not finished. A second draft appeared in March 2026, and the final version is expected later this year. The standard for what the marking should actually look like is still moving. That is an argument for asking the question now, while the answer costs an email, rather than in December, when it may cost a migration.
If you operate in Norway, Iceland or Liechtenstein, one further caveat applies: the regulation is EEA-relevant, national implementation is still under way, and the date on which each individual requirement binds is not necessarily the same as inside the EU. About the direction there is no doubt.
Part three: a switching cost you build without noticing
The third part has no date attached. That is why it is easy to miss.
Try describing the voice on your own service number right now, without looking it up. Most people cannot. The customer who calls every other week can do it without thinking – not by describing it, but by recognising it in half a second. That is the kind of recognition that is worth something, and it is also the only kind that is recorded nowhere.
It is also built on something you probably do not own. The voice in the dropdown belongs to the model provider. It can be changed, renamed or removed in a version update, in exactly the same way that the model underneath the agent can be swapped out without anyone calling you first. The difference is where the change surfaces. A model that gets swapped shows up in the numbers. A voice that gets swapped shows up with the customers, and it shows up first with the ones who call most often.
It is an unusual kind of debt. It grows the better things go.
The counterargument, put as strongly as I can put it
The objection is a good one, and it deserves to be stated properly.
For the large majority of Nordic businesses the voice is a dropdown, and it should be. A plumber taking four hundred calls a month gets nothing back from an exclusive, licensed brand voice. He gets everything back from the phone being answered. The value is in somebody picking up at seven in the morning, not in who picks up.
The argument reaches further than the small business. This industry has a well-established habit of selling complexity into decisions that ought to be simple, and "voice provenance" has every hallmark of the next round of it: a real problem for a handful of large players, marketed as a universal risk by the people who happen to sell the remedy. The report I quoted above is, after all, published by a marketplace for voice actors.
And the marking obligation in Article 50(2) genuinely does not bind the buyer. It binds the supplier. A company that spends internal resources building expertise in machine-readable audio marking has spent time on something a contract should have settled in two sentences.
All told, that is a better argument than the claim it attacks – if the claim were that everyone should license their own voice. It is not.
Why the claim holds anyway
The distinction is between owning something and knowing something.
Nobody should buy a bespoke voice because a supplier report says 79 percent of leaders think it matters. But everyone should know what they picked, and right now almost nobody does. Not because it is difficult, but because the question has not been asked loudly enough for anyone to have prepared an answer.
That is a cheap state to be in, and it stays cheap right up until something happens. It stops being cheap the day the voice disappears in an update, or the day a procurement lead at a group asks for documentation of where the audio on a number they are about to outsource came from.
Three questions cover it. They are put to the supplier, and they are not solved by you.
Where does the voice come from? A model name and an identifier is a sufficient answer. "The Nordic female voice" is not. You are not asking for a chain of copyright title, you are asking for a string you can look up again in a year and recognise.
What happens if it changes? There are three usable answers: it can be pinned to a version, you get notice in advance, or you can switch back. If none of them exist, the voice is rented by the day – which is fine, as long as everyone knows it.
Who marks the audio, and by when? This is the supplier's obligation and your question. An answer without a date in it is not an answer. The same principle applies to the rest of the supplier agreement.
Three questions, one meeting, no cost. By comparison, changing the voice after two years in production takes a decision at management level, because somebody has to explain why the phone sounds different.
Finally
It is worth noticing what none of the three questions is about.
None of them is about whether the voice sounds human. That debate takes up the most space in the industry and matters least to the person calling – she has been told she is talking to a machine before she had time to form an opinion about it either way. The opening line has already said so.
What is left once that question is out of the way is more prosaic and more durable: that the audio on your number has an origin, that the origin has a name, and that somebody at your company knows what that name is.
Threll.ai builds voice agents in Norwegian, Swedish and Danish. Choosing a voice takes under a minute. Knowing which one you chose takes about as long – but only if you do it now.
