Two errors are available. They do not cost the same.
A fabricated statement accepted costs you the claim value. A genuine statement wrongly rejected costs a person who has just had a fire, a crash or a bereavement, and then costs you the complaint, the ombudsman file and the story. Set your threshold with that asymmetry in front of you.
What has changed in the claims file
Recorded material has quietly become ordinary evidence. A telephone statement taken by a handler. A voicemail from a witness. A voice note the policyholder sends because typing it out was too much. Dashcam and doorbell audio. Third-party recordings forwarded by a broker. Almost none of it arrives as an original file, and almost all of it has passed through at least one messaging platform.
The new question is whether a voice on the file was ever spoken. It is asked in both directions. An organised claim can now include a fabricated witness. Equally, a claimant whose account is inconvenient can be dismissed on the suggestion that their recording is synthetic, and a handler with a plausible-sounding tool has a new way to reach that conclusion quickly.
The asymmetry, stated properly
| Error | Who bears it | What it costs |
|---|---|---|
| False negative A fabricated statement is read as genuine | The insurer, and ultimately the pool | The value of the claim, plus whatever the same operation takes next. Recoverable in principle; detectable later through pattern analysis. |
| False positive A genuine statement is read as fabricated | A single claimant, at the worst moment of their year | A refused or delayed claim, an accusation of fraud they cannot rebut, and a process where the burden has quietly inverted. Not recoverable by the person it lands on. |
Both errors are real and both have non-zero rates. But they are not symmetric in kind. The first is a financial loss to an institution built to absorb financial losses. The second is an accusation against an individual who cannot produce evidence that a recording of their own voice is authentic — because no such evidence exists.
That is why we would rather you set a conservative threshold and accept more unclear results than run a sensitive one and generate flags on ordinary bad audio. A claims operation that tolerates a slightly higher false negative rate in exchange for far fewer false positives is making the right trade, and it is the trade a regulator or ombudsman will expect you to have made deliberately.
What this cannot do for you
It cannot tell you whether a claim is fraudulent. It analyses audio. Fraud is established by the claim narrative, the physical evidence, the financial pattern and the interview — not by a score on a voice note.
It cannot tell you who is speaking. No speaker identification, no matching against a known sample. Whether the voice is your policyholder is a different question we do not answer.
It cannot verify a statement is genuine. Likely human means no known signature was found. It is the weaker verdict and must not be written into a file as authentication of a witness.
It cannot cope well with the audio you actually hold. Telephone statements are narrowband and compressed. Voice notes have been re-encoded by a messaging app. Doorbell audio is noisy and heavily processed. Expect unclear often, and treat it as the honest answer rather than a defect.
It cannot decline a claim. A named handler makes that decision on the whole file and must be able to explain it without relying on a model output. Where a decision has a significant effect on a person, several regimes restrict outcomes reached solely by automated means, and this is precisely the situation those provisions were written for.
Where it earns its place
Not at first notification of loss, and not across the book. Three narrower uses hold up:
- Claims already referred for investigation. Where an SIU is engaged for reasons independent of the audio, an analysis is one more documented input, run on a specific recording for a stated purpose.
- Linked-claim analysis. Attribution to the same generator family across several unrelated claims is a linking signal that no single-claim review would surface. It does not name anyone; it tells you where to look.
- Defending the genuine claimant. The use that gets forgotten. Where a third party alleges a policyholder’s recording is fabricated, an analysis and a documented method is a better answer than an assertion — provided you present it with its limits attached.
Rules on fraud investigation, claimant notification, automated decision-making and the handling of special category data vary by jurisdiction and line of business. [VERIFY: verify the position in each market and with your own compliance function] Nothing here is legal or regulatory advice.
Questions from claims and SIU
Do we have to tell a claimant we ran an analysis?
Assume yes where it forms part of a decision about them. A claimant who learns after the fact that a machine result contributed to a decline has a strong complaint even if the decline was otherwise sound, and transparency costs you very little at the point of use.
What confidence threshold should we act on?
We will not set it for you, because the cost of each error is yours to weigh. What we would say is that the threshold at which you investigate further and the threshold at which you act adversely should not be the same number, and the second should be considerably higher.
Can this be part of our fraud scoring model?
As a feature among many, with a documented weight, yes. As a feature that alone moves a claim into a decline path, no. Whatever weight you give it, keep the audio quality ceiling as a companion field — a strong verdict on poor audio is not a strong signal.
Will you support a decline in litigation?
A verified report is a dated record of an analysis and can be produced as such. It is not an expert opinion and there is no examiner to be cross-examined. Where the audio is central, instruct a qualified forensic examiner. More on that here.
Decide the threshold before the first case
Write down what you will do at each confidence band, and who signs it off, while nothing is at stake.
Reviewed