Destill.ai · Proof of Agent Work
For reinsurers and AI-liability underwriters

Underwrite what AI agents actually did, not what anyone says they did.

Proof of Agent Work turns every run of an AI agent into a record that is signed, stamped by an independent time authority, and checkable by your own people in seconds. Change one step afterwards, and the proof breaks.

Claims fraud · the leak below the waterline

What has been paid is what nobody sees.

Fraud is counted where it is caught. A doctored photo that is not caught is paid, and appears in no statistic. The ship only sits a little lower in the water.

> €6 bn a year

Germany alone

The German insurers' association GDV estimates the damage from insurance fraud at more than six billion euros a year, up from four to five. Roughly one claim in ten in property and casualty is suspicious enough to investigate; half of those are motor.

Estimate, not count

The blind spot

Every fraud figure is an estimate, because the fraud that is paid without suspicion leaves no trace. Today a claim photo is judged by eye or by a detector that guesses. Neither tells you how much you missed.

Check, not guess

An open model, a pilot threshold

A photo sealed at live capture answers "edited, someone else's, or old?" with yes or no, before payment. Where that bites, where it does not, and why we set at least half of the addressable fraud as the pilot threshold: the open audit.

Read the fraud audit →

Live proof 2 · the claim photo

Seal a claim photo, enlarge the damage, and watch the check catch it.

Runs in this browser tab with the Fortress SDK; nothing is uploaded. The photo is synthetic, drawn from numbers. The seal here is set by this page and is not signed: it shows the check. In production ProofCam sets the seal at live capture, and the badge then says "Captured live & unaltered", nothing more.

Live since 4 Nov 2026: ProofCam for Android exports a seal file signed by a hardware key in the phone and attested by Google: the genuine ProofCam app on a genuine phone sealed exactly these bytes. A browser seal says only "unaltered since sealing". Not available: a defence against photographing a screen, and an iPhone app with attestation. No seal says what was in front of the lens. How a claim photo travels → · Check your own file (German)

Part 2 · POAW and Fortress for insurers

The claim photo and the agent decision: estimate, or check?

6:45 · Deutsche Fassung · chapters and the six role clips: by role.

Deepfake detectors estimate

A detector looks at a finished picture and returns a probability. It can be wrong in both directions, and every new generator model makes its job harder.

Fortress SEAL checks

A photo taken live in our camera app is sealed at the moment of capture. The check compares the photo with that seal and answers yes or no: changed since capture, and if so, where. Existing or uploaded photos cannot be sealed, and the badge says "Captured live & unaltered", nothing more. See the heatmap for claim photos →

POAW: a receipt per agent step

Every step of an AI agent gets a receipt: signed with ML-DSA-65, time-stamped per RFC 3161, checkable offline with OpenSSL. It supports the evidence for the AI Act logging duties (Art. 12, 19 and 26(6)); it does not make anyone compliant.

Planned, not built: one evidence chain from the sealed photo to the decision. Pilot partners wanted.

After a loss, the agent's log is the only witness. Today that witness lives in a database the insured controls — and nobody can prove it was not rewritten.

For claims departments · 42 seconds

Who decided what? The short version.

Voice: neural text-to-speech · the claim in the film is example data · Deutsche Fassung

Start with why · seven minutes, hand-drawn

Why a decision made by a machine needs a receipt — and where that earns its keep first.

The whole case in one film: why, what a receipt is, what the EU AI Act and the Product Liability Directive ask for, and three sectors with a worked example each. Every claim is sourced; one scene says what the receipt does not do.

Voice: neural text-to-speech · music CC BY 4.0 (Kevin MacLeod) · sources in each scene's caption and at the end card.

Two minutes, two films

The case for insurers, and the case for reinsurers.

For insurers

Underwrite what the agent actually did.

Claim photos · under five minutes

A claim photo that proves itself: sealed at capture, one chain to the decision.

In 40 seconds

Any edit after capture breaks the seal and shows where; every AI step of the claim carries a signed receipt.

The full film

Under five minutes, with a walkthrough of the demo chain. The seal proves "unchanged since capture", not that a scene is true. ProofCam for Android signs its seal with an attested hardware key; iPhones and photographed screens are not covered.

Sources of the numbers in the film: survey of March 2026, 1,000 US consumers and 300 claims professionals (Verisk) · UK van case, June 2026 (Openkoda) · seal: FORTRESS lab benchmark, small sample.

Break the demo chain Seal at the sensor

Why now

Liability is moving into software, and evidence is becoming the deciding factor.

01 · 9 Dec 2026

Software becomes a product

The new EU Product Liability Directive explicitly covers software, AI included. It applies to products placed on the market after 9 December 2026.

Directive (EU) 2024/2853, Art. 4(1), Art. 2
02 · Disclosure

No evidence, presumed defective

Courts can order a defendant to disclose relevant evidence. If it is not disclosed, the product can be presumed defective. Whoever holds a credible record holds the case.

Directive (EU) 2024/2853, Art. 9 and Art. 10(2)(a)
03 · Logging duty

Logs are now mandatory

High-risk AI systems must log events automatically, and deployers must keep those logs for at least six months. The law asks for logs. It does not ask whether anyone can trust them.

EU AI Act, Art. 12 and Art. 26(6)
What it is for

Three places where a checkable record changes an insurer's numbers.

The effect in your portfolio is not measured yet; that is what a pilot is for. The problems are documented.

Claims fraud

Edited claim photos

Allianz UK reported a rise of 300 % in claims backed by manipulated images, videos and documents from 2021/22 to 2022/23. A sealed photo turns "does this look edited?" into a yes/no check with the edited area marked, before payment.

Source: The Guardian, 2 May 2024, reporting Allianz UK figures
Silent AI

AI losses in policies not written for AI

AI losses can sit unpriced in E&O, cyber and D&O policies. Signed agent runs show which AI ran in an insured process and which controls actually ran, so the exposure can be seen and priced.

Source: Armilla AI, on AI sublimits in existing policies (armilla.ai)
Product liability · 9 Dec 2026

Disclosure, or a presumption

Under the new Product Liability Directive a court can order disclosure; without it, a defect can be presumed. A structured, checkable record helps the insured, and you in recourse, to answer that order. It does not exclude liability.

Directive (EU) 2024/2853, Art. 9 and 10(2)(a)

Want the whole chain in one case file, from the claim photo to the AI steps to the decision? Check a case file (German) →

Your supervisor's frame

Where a signed record helps you evidence an obligation you already have.

A record helps you show what ran and when. It does not fulfil any of these obligations for you. Not legal advice.

DORA

ICT incidents, and us as a provider

Insurers and reinsurers are financial entities under DORA. A time-stamped trail helps you evidence the timeline of an ICT incident (Art. 17 and 19). If you use our hosted service, we are your ICT third-party provider (Art. 28 ff.); the record service can also run in your own data centre.

Regulation (EU) 2022/2554, Art. 2(1), 17, 19, 28 · eur-lex.europa.eu
Solvency II · ORSA

Controls you can show to the risk function

Where AI agents run underwriting or claims steps, their operational risk belongs in your system of governance and your own risk and solvency assessment. Signed runs show the risk function and internal audit which controls actually ran, not only which were designed.

Directive 2009/138/EC, Art. 41, 44 and 45 · eur-lex.europa.eu
EIOPA

Documentation and record keeping for AI

EIOPA's Opinion on AI governance and risk management names documentation and record keeping as one of six areas, next to human oversight and explainability, and ties it to the existing duty to keep orderly records.

EIOPA-BoS-25-360, 6 August 2025 · eiopa.europa.eu
ISO/IEC 42001

Records for an AI management system

If you run or plan an AI management system under ISO/IEC 42001, signed runs can serve as its operational records. We are not certified under ISO/IEC 42001, and a record does not certify you.

ISO/IEC 42001:2023
EU AI Act

Logs, and marking of AI content

High-risk systems must log automatically (Art. 12); providers and deployers keep the logs for at least six months (Art. 19, 26(6)). Art. 50(2) asks for machine-readable marking of AI-generated content. A record helps you evidence the log; whether its scope is enough for your purpose is your decision.

Regulation (EU) 2024/1689 · eur-lex.europa.eu
GDPR

Accountability, privacy by design

Sensitive values sit in a record only as hashes, which supports accountability and privacy by design (Art. 5(2), 25, 32) and answers to an automated decision (Art. 22(3)). It supports these duties; it does not replace your own assessment.

Regulation (EU) 2016/679 · eur-lex.europa.eu · DPIA template
What a record changes

From questionnaires and reconstructions to evidence.

Underwriting

Controls you can see working

Instead of a questionnaire answer like “a human approves every payment”, you see signed runs that show whether the approval step was actually there.

Claims

The decision path in minutes

You get the exact sequence of steps the agent took: what it received, what it estimated, who approved it. And you can be certain it is the record written at the time.

Recovery

A record nobody can disown

The operator signs the record with its own key, and an independent authority dates it. Neither the vendor nor the insured can later claim the log was something else.

Don't trust us

Three checks. None of them needs our software.

The demo above runs these checks in your browser. Your engineers can repeat them with Python and standard OpenSSL, without a single line of our code. We ran exactly that on 21 October 2026.

1 · Integrity

Every step, in order

Each step is fingerprinted with SHA-256 into one Merkle root. Change, remove or reorder any step, and the root no longer matches.

2 · Authorship

Signed by a known key

The root is signed with ML-DSA-65, the NIST post-quantum signature standard (FIPS 204). The public key is published on the signer's own domain.

3 · Time

Dated by a third party

An independent RFC 3161 time-stamping authority (here: DigiCert) stamps the signature. Nobody can backdate the record afterwards, us included.

Run the checks yourself →

Read this before your actuaries do

What a proof tells you, and what it does not.

✓ A valid proof establishes

  • Unchanged: the steps are exactly those recorded when the record was signed.
  • Who: the record was signed with the key the operator published.
  • When: the record existed no later than the time of the independent stamp.

× It does not establish

  • Correctness: that the agent decided well. The proof shows what happened, not whether it was right.
  • Completeness: that every action was recorded. It covers what the runtime logged, so the logging point is part of the pilot design.
  • Key safety: keys are held in software today. Hardware key custody (HSM) is on the roadmap and not yet in place.
  • Qualified status: we are not a qualified trust service provider under eIDAS; the time stamp is a standard RFC 3161 stamp, not a qualified one.
Proof of concept

One agent workflow. Your verifier. Your team tries to break it.

A typical pilot covers one agent workflow, either yours or one of an insured's, and takes two to four weeks.

Step 1

Connect

We connect one agent workflow, for example claims intake. Every step gets recorded, signed and timestamped as it runs.

Step 2

Receive

Your team receives the records, the open verifier and the commands needed to check them without our software.

Step 3

Attack

Your security and claims people try to forge, backdate or edit a record. The decision then rests on that result, not on a slide.

Success criteria

Agreed in writing at kick-off

  • Every agent run in the pilot workflow produces a record your team can check.
  • Your team checks a record on its own, without our software, and sees an edited one fail.
  • The extra effort per run is measured; the target is set with you at kick-off.
Source: our pilot offer; what we commit to measure, not results from an earlier pilot.
What the pilot measures

Your numbers, before and with records

  • Disputed cases: how many cases in the workflow lead to a query, recourse or complaint, and in how many of them the record was used.
  • Effort per disputed case: hours to clarify a case, measured by your team, before the pilot and with records.
  • Cost per case: those hours times your internal rate.
  • Claim frequency: four weeks are too short to show a change. The pilot records the baseline; a change can only be read over a longer period.
We claim no effect before your pilot has measured it. Our calculator (German) takes exactly these numbers as your own inputs. Full pilot measurement plan →
Terms

Price logic and operation

The pilot has a fixed price, agreed after a first call about scope. After a successful pilot: an annual licence, scaled by volume. Amounts only in the written offer.

Hosted by us in the EU, or the record service in your own data centre (Docker, with your own signing keys).

Contracting party: destill.ai Ltd. (Cyprus, in formation). Until it is registered, Lebensfluss e.V. (Austria, see footer) signs the pilot.

Set the evidence standard for AI cover, before claims set it for you.

DESTILL licenses Proof of Agent Work to insurers and reinsurers. Talk to the inventor directly.