Life pricing is high-risk AI. Keep the proof from the first quote.
From 2 December 2027, the EU AI Act treats AI used for risk assessment and pricing in life and health insurance as high-risk. HONESTAS keeps the evidence those rules ask for on every application and every claim: the inputs, the model versions, the person who signed and reasons a customer can read. And it holds one boundary fixed: emotion and psychometric inference never reach an insurance decision.
Application LIF-2291, accepted
- Decision
- Accept with a +50% loading on life cover
- Chair
- Niamh Ward, chief underwriter
- Evidence
- GP report, disclosures, blood pressure readings
- Not used
- Call audio, voice tone, social media, any personality score
- Checked
- Two independent models agree
- Signed
- Passkey at her desk, 10:12:27 UTC
- Rules
- Underwriting manual 2026.2, AI Act Annex III 5(c)
- Scope
- It does not judge the applicant's honesty or character
Six decisions a supervisor will ask about first.
Each one is argued, checked by a second model, signed by a named person and sealed. Open one to see what goes in, what comes out and who signs.
Underwriting a new application
The application, disclosures and medical evidence are read and filed with their sources. A rating is proposed with every input it used, and an underwriter decides.
- Goes in
- The application, disclosures, GP report and tests
- Comes out
- A rating or decision with its inputs and reasons
- Signs
- Underwriter, within their authority
Pricing and rating models
Each pricing model is kept with its owner, versions, tests and the inputs it may use. A change shows which customers it moves before anyone approves it.
- Goes in
- Model versions, rating factors and test results
- Comes out
- An approved model pinned to its tests, ready for the AI Act file
- Signs
- Pricing actuary, then the model risk owner
The input boundary
Every input is traced back to its source. Anything derived from emotion or psychometric inference is refused as an input to underwriting, pricing or claims, and the refusal is recorded.
- Goes in
- Input lists and their sources
- Comes out
- An input list with every source, and a record of what was refused
- Signs
- Nobody. This check cannot be switched off
Declines, loadings and human review
A decline, a loading or an exclusion carries reasons a customer can read. The customer can ask for a person to review it, and the review is recorded too.
- Goes in
- The sealed decision and the customer's request
- Comes out
- A plain explanation and a recorded human review
- Signs
- A senior underwriter who did not make the first decision
Claims
Income protection, serious illness and health claims are read against the policy and the medical evidence, with sources. A decline goes to a decision room, never out from a model alone.
- Goes in
- The claim form, medical reports and the policy
- Comes out
- A decision with reasons, and payments with their checks
- Signs
- Claims assessor, with the chief medical officer where needed
Fundamental rights impact assessment
Before you use a high-risk pricing system, the AI Act asks you to assess its effect on people. HONESTAS drafts it from your records: who it affects, the oversight and what happens when it is wrong.
- Goes in
- The system's purpose, its records and its oversight
- Comes out
- A drafted impact assessment, linked to the records behind it
- Signs
- Compliance lead
Emotion and psychometric inference never reach an insurance decision.
Not as an input, not as a tie-breaker and not as a fraud signal. The contact centre understands what a caller says. It never scores how they sound, and nothing it hears about their mood or character is passed to underwriting, pricing or claims.
| Input | Comes from | Allowed |
|---|---|---|
| Age, smoker, sum assured | Application form | Yes |
| Blood pressure readings | GP report, p. 2 | Yes |
| Medication | GP report, p. 3 | Yes |
| "Hesitation" score | Voice analysis of the sales call | Refused |
| Personality type | Marketing survey | Refused |
Marketing analytics asked to add the hesitation score on 3 September. Refused, and sealed with who asked.
On every model release, every input change and every night, in your EU cloud or your own data centre.
Refused at the source. The model cannot go live while a refused input is attached.
Checked at the source
Each input is traced back to where it came from. A path to an emotion or psychometric score fails the check, and the model cannot go live.
Refusals are kept
When someone tries to add such an input, the refusal is sealed with who asked and when. The boundary itself becomes evidence.
Nobody can switch it off
Not an administrator, not a model owner and not us. It holds the same way in our EU cloud and in your own data centre.
A life application with a blood pressure history.
Garrow Life, a fictional life insurer, and one applicant. Six steps from the application to a sealed decision with reasons he can read. Scroll, and the screen follows the step.
- 01 · Application
Daniel applies for life cover.
Daniel Byrne, 47, applies through his broker for €400,000 of life cover and serious illness cover. He discloses blood pressure treated since 2023. The underwriting rules ask for a GP report.
Aisling Doyle, new business underwriter, checks the file.
honestas · new business · LIF-2291Sample dataApplication 2291 · Term life and serious illnessLIF-2291NewApplicationEvidenceRatingRoomDecisionSealedThe application- Applicant
- Daniel Byrne, 47
- Cover
- €400,000 life, 20 years
- Also
- Serious illness, €100,000
- Through
- Corran Brokers
Disclosed- Smoker
- No
- Blood pressure
- Treated since 2023
- Family history
- None given
- Status
- Evidence due
Rule that firedTreated blood pressure at this sum assured needs a GP report. Requested from Dr Kelly on 6 October.
The rules asked for evidence. Nothing has been decided.
- 02 · Medical evidence
The GP report, read line by line.
The report is read into facts, each linked to its page. The application says "well controlled". The readings average says otherwise. Both are kept in their own words.
Aisling reads the facts. Nothing is decided yet.
honestas · new business · LIF-2291 · evidenceSample dataMedical evidence · GP reportLIF-2291Evidence filedApplicationEvidenceRatingRoomDecisionSealedWhat the report says4 sourcesTreated with one medicine since March 20231. The average of the last six readings is 148/942. The two most recent are 136/86 and 134/843. No damage to heart or kidneys recorded4.
1GP report, Dr Kelly, p. 3PDF2GP report, readings table, p. 2PDF3GP report, p. 2, last two rowsPDF4GP report, p. 4PDFWhere the papers differThe application says "well controlled". The six-reading average is above the manual's threshold. Both are kept, in their own words.
Health data is read only for this application and only by the people on it.
- 03 · Rating
A proposed rating, with every input shown.
The rating model proposes +50% on life cover. Every input it used is listed with its source. Nothing derived from emotion or psychometric inference was tried, and postcode is not an input.
The pricing actuary owns model LR-7.
honestas · pricing · LIF-2291 · ratingSample dataProposed rating · LIF-2291LIF-2291ProposedApplicationEvidenceRatingRoomDecisionSealedRating model LR-7, v3- Life
- +50% loading
- Serious illness
- +75%, or exclude
- Basis
- Six-reading average
- Tested
- Fairness tests, 2 Oct
Inputs it used- From the form
- Age, smoker, cover
- From the GP
- Readings, medication
- Refused
- None tried
- Postcode
- Not an input
Why +50%The manual rates the six-reading average at +50%. The two latest readings are lower, which the rating does not use yet.
A rating is a proposal. An underwriter decides.
- 04 · Decision room
+50%, or +25% on the latest readings?
One AI seat argues for the manual's +50%. One argues the treatment is working. The challenger checks the model is not counting age twice. Niamh rules +50% with a review after two years.
Niamh Ward, chief underwriter, chairs. The second model agrees.
honestas · decision rooms · LIF-2291Sample dataRating · Application 2291LIF-2291Chair to ruleApplicationEvidenceRatingRoomDecisionSealedChairNiamh WardChief underwriterAI · for+50%The manual's averageAI · against+25%Latest readings lowerChallengerFairnessProxy check on ageCase againstThe last two readings are in range. Rating on the old average overcharges a customer whose treatment is working.ChallengerThe loading rises with age faster than the manual says. Check that the model is not counting age twice.Niamh WardAccept at +50% on life, as the manual says. Offer a review after two years of readings in range. Age is counted once: checked.Second model: agrees with +50%. Serious illness goes to a separate decision.
- 05 · Decision
Reasons Daniel can read.
The decision letter explains the loading in plain words, says how to have it looked at again, and tells Daniel he can ask for a person to review it.
Niamh Ward signs with her passkey.
honestas · new business · LIF-2291 · decisionSample dataDecision · LIF-2291LIF-2291Ready to sealApplicationEvidenceRatingRoomDecisionSealedDecision- Life cover
- Accept, +50%
- Premium
- €61.40 a month
- Review
- After two years
- Signed by
- Niamh Ward
What Daniel readsWe can offer life cover at a higher price because of the blood pressure readings in your GP report. If your readings stay in range for two years, ask us to look again. You can also ask for a person to review this decision.
Seal and sendChangeSend backThe explanation is drawn from the record, not written afterwards.
- 06 · Sealed
Sealed, and ready for the AI Act file.
The decision, the inputs, the model version, the room and the signature are sealed into one record. It goes into the evidence the AI Act asks a deployer to keep for high-risk pricing.
The compliance lead can open it at any time.
honestas · sealed records · R-2291-01Sample dataSealed record R-2291-01LIF-2291SealedApplicationEvidenceRatingRoomDecisionSealeddecided Accept, +50% on life cover chair Niamh Ward, chief underwriter inputs 7 used · 0 refused · sources kept model LR-7 v3 · second model agrees rules Manual 2026.2 · AI Act Annex III 5(c) time 2026-10-21 10:12:30 UTCSealed · witnessed outside Garrow · fingerprint 4be0 a1c3 9d27 7c19Send decision letterShare passport linkAdd to the AI Act file
- Applicant
- Daniel Byrne, 47
- Cover
- €400,000 life, 20 years
- Also
- Serious illness, €100,000
- Through
- Corran Brokers
- Smoker
- No
- Blood pressure
- Treated since 2023
- Family history
- None given
- Status
- Evidence due
Treated blood pressure at this sum assured needs a GP report. Requested from Dr Kelly on 6 October.
The rules asked for evidence. Nothing has been decided.
Treated with one medicine since March 20231. The average of the last six readings is 148/942. The two most recent are 136/86 and 134/843. No damage to heart or kidneys recorded4.
The application says "well controlled". The six-reading average is above the manual's threshold. Both are kept, in their own words.
Health data is read only for this application and only by the people on it.
- Life
- +50% loading
- Serious illness
- +75%, or exclude
- Basis
- Six-reading average
- Tested
- Fairness tests, 2 Oct
- From the form
- Age, smoker, cover
- From the GP
- Readings, medication
- Refused
- None tried
- Postcode
- Not an input
The manual rates the six-reading average at +50%. The two latest readings are lower, which the rating does not use yet.
A rating is a proposal. An underwriter decides.
Second model: agrees with +50%. Serious illness goes to a separate decision.
- Life cover
- Accept, +50%
- Premium
- €61.40 a month
- Review
- After two years
- Signed by
- Niamh Ward
We can offer life cover at a higher price because of the blood pressure readings in your GP report. If your readings stay in range for two years, ask us to look again. You can also ask for a person to review this decision.
The explanation is drawn from the record, not written afterwards.
Mapped to the AI Act and the rules around it.
Each sealed record is linked to the articles it helps answer, so the technical file and the impact assessment start from evidence, not from memory.
- EU AI Act, high-riskRegulation (EU) 2024/1689, Annex III 5(c)Risk assessment and pricing for people in life and health insurance is high-risk AI from 2 December 2027.
- Deployer dutiesAI Act Art 26 and 27Trained people to oversee the system, logs kept, and a fundamental rights impact assessment before first use.
- Right to an explanationAI Act Art 86A person affected by a decision based on a high-risk system can ask for a clear explanation of the AI's part in it.
- Emotion recognitionAI Act Art 5(1)(f)Banned in workplaces and schools since 2 February 2025. HONESTAS keeps it out of every insurance decision too.
- GDPRRegulation (EU) 2016/679, Art 9 and 22Extra protection for health data, and the right not to be subject to a decision based only on automated processing.
- CJEU, C-203/22Dun & Bradstreet Austria, 27 February 2025A person must be told, in terms they can follow, how an automated decision about them was reached.
- Insurance DistributionDirective (EU) 2016/97 (IDD), Art 17 and 25Act in the customer's best interests, with product oversight from design to sale.
- EIOPA opinion on AIOpinion on AI governance and risk management, August 2025Supervisors' expectations for fairness, explanation and human oversight of AI in insurance.
AI Act dates as amended by Regulation (EU) 2026/1744, the Digital Omnibus, published in the Official Journal on 24 July 2026. Not legal advice.
Go deeper
Start with the decisions a supervisor asks about first.
Bring one product, such as term life or income protection. We will show an application argued, checked, signed and sealed, with the AI Act file it feeds.