Riaan Kleynhans shows you:
The step-by-step process an insurer uses to get AI out of pilots and into the daily work
Get the Pilot-to-Production Blueprint
All thirteen steps, as a PDF
Case · insurance
One insurer, from a stalled pilot to AI in the daily work
Muster Versicherungsgruppe is a composite insurer in Germany, Austria and Switzerland. This page takes it through every step, the way I would take you.
Muster is fictional: every name and figure is invented. The method is not.
The problem, in the COO’s words
“We started a claims document pilot nine months ago. It works in the demo, and it has never touched a real claim outside that team. Our CISO saw it for the first time last month. Finance asks what it saves and nobody can say. I don’t need more pilots. I need one that goes live, and to tell the board what we run.”
- Pilot theatreThe claims pilot works in one team for nine months and goes nowhere.
- AI nobody approvedA meeting summariser in daily use, with no owner and no oversight.
- Security and legal say no lateThe CISO sees the pilot in month eight.
- Value nobody signedFinance asks what it saves; nobody agreed a figure before it started.
- The AI Act, unansweredA health underwriting score is planned: high-risk under Annex III 5(c), and nothing is classified.
Where Muster stands
Muster Versicherungsgruppe is a composite insurer in Germany, Austria and Switzerland: 6,800 people, four business units, about 9,000 property and casualty claims a month. Like most insurers, it has spent two years experimenting with AI.
Five AI systems run or are planned. A claims document pilot works in one team. A chat assistant answers customers’ policy questions in a pilot. Fraud scoring runs in production. The actuaries want a health underwriting score. Staff use a meeting summariser nobody bought.
The board has asked for a register of every AI system and its AI Act position before the next budget round. Muster is hiring a Head of AI Governance. Nobody owns the answer yet.
What Muster does not comply with today
A preliminary reading of Muster’s position in October 2026. Not legal advice: counsel decides.
The chat assistant talks to customers without saying it is AI
The transparency duty has applied since August 2, 2026: people must be told they are interacting with an AI system. Muster’s pilot does not tell them.
Regulation (EU) 2024/1689, Article 50(1)
Staff use AI without a rule or training
Since February 2, 2025, a company using AI must ensure its staff have sufficient AI literacy. The meeting summariser has no owner, no rule and no training behind it.
Regulation (EU) 2024/1689, Article 4
Health data planned for AI without a documented basis
The health underwriting idea would process health data. No legal basis and no impact assessment are recorded for it.
Regulation (EU) 2016/679 (GDPR), Articles 9 and 35
No register the board or an auditor could open
Nobody can show which AI systems run, who owns them and how they are overseen. Every question from the board, an auditor or a supervisor starts from zero.
The board’s own request; every AI Act duty starts with knowing what runs
What is coming
From December 2, 2027, the obligations for high-risk systems in Annex III apply. They cover risk assessment and pricing of people in life and health insurance: the very use case the actuaries want. The date was August 2, 2026 until the Digital Omnibus moved it, so a plan written before July 2026 rests on a date that no longer exists.
Dates from the engine’s AI Act calendar: Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 — Digital Omnibus on AI.
What is at stake
- For breaches of the high-risk and transparency obligations, the AI Act sets a ceiling of €15 million or 3 % of worldwide annual turnover, whichever is higher. On Muster’s fictional turnover of €2,400,000,000, that ceiling is €72,000,000.
- The fine is not the largest risk. An authority can also order a system withdrawn, and a high-risk use case built without its obligations has to be rebuilt.
- Meanwhile the pilot costs money every month, the board’s question stays unanswered, and the deadline comes closer whether or not anyone owns it.
A statutory ceiling, not a prediction and not legal advice. Authorities calibrate the actual penalty.
What Muster decides
The COO does not commission another pilot. The COO books a Scan: three weeks to find out what runs, what it is worth, where Muster stands under the AI Act, and which one pilot to take live, signed by the people who can say no. The thirteen steps below show how.
Thirteen steps, one chapter each
Each chapter shows one step on the real engine: what happens, what it solves and who signs it. The videos follow as they are recorded, in English and in German.
0 of 13 chapters recorded
Before Muster commits
1The problem
Nine months, one team, no real claim touched. The five problems, each found in one paragraph.
Solves: Naming the stall before anyone buys anything.
Signed by: nobody yet
See this step on the siteVideo coming- 2
First touch
The COO runs the self-assessment: readiness 46, value the weakest. Then the intake shows an insurer like Muster.
Solves: The buyer arrives with their own answer, not a pitch.
Signed by: nobody yet
See this step on the siteVideo coming - 3
The free call
Five questions: which pilot, who can scale it, who can stop it, by when, and will they stop a pilot that does not earn its keep.
Solves: The risk is visible before a contract.
Signed by: the fit is recorded
See this step on the siteVideo coming - 4
The quote
Ten drivers from Muster’s facts set the price: €20,900 net, inside the published range, signed online.
Solves: A price a regulated buyer can check line by line.
Signed by: the COO
See this step on the siteVideo coming The Scan · three weeks
5Align
The charter with three outcomes. The CISO’s and the DPO’s conditions are recorded as requirements on day one.
Solves: Security and legal saying no late.
Signed by: COO, CFO, claims operations lead
Shown in the engine, on a call.
Video coming- 6
Discover
Five processes, five AI systems. Two without stated oversight, one with no owner at all.
Solves: AI nobody approved: the board gets its register.
Signed by: head of claims, data steward, CIO
See this step on the siteVideo coming - 7
Diagnose
Ten dimensions with reasons: 46, value the weakest, five gates failing, each gap with a closure test.
Solves: A readiness score with reasons, not a maturity slide.
Signed by: model risk, change lead
See this step on the siteVideo coming - 8
Prioritize
Six use cases ranked by rule. Claims document intake is the only pilot. Two life and health cases are high-risk and wait. Fraud scoring waits for counsel.
Solves: The AI Act, unanswered: every use case has a class and an owner.
Signed by: COO, head of claims, model risk, DPO, compliance
See this step on the siteVideo coming - 9
The readout
A business case with a value floor and a named owner. The CFO and the COO sign go.
Solves: Value nobody signed.
Signed by: COO, CFO
Shown in the engine, on a call.
Video coming The Last Mile · 90 days
10Design
Controls with a technical control and an owner each: every decline and payment stays human; the model runs on Muster’s own endpoint and key.
Solves: Governance built in, not bolted on.
Signed by: model risk, CISO, DPO, compliance, data, CIO
See this step on the siteVideo coming- 11
Prove
The evaluation gate on 200 past claims, an agreed bar, the newest run counts. Value against the floor, cost per claim. Two endings: scale, or stop.
Solves: Pilot theatre: a demo is not proof.
Signed by: COO, head of claims, CISO, CFO
Shown in the engine, on a call.
Video coming - 12
Scale
The evidence pack, the operating rhythm, the handover with a dated confirmation that Muster runs it alone.
Solves: Dependence on the consultant.
Signed by: COO, model risk, CIO, claims operations lead
Shown in the engine, on a call.
Video coming Assurance · monthly
13Assurance
Seven health checks, a reassessment against Muster’s own baseline, the register kept current as the AI Act moves.
Solves: Knowing it still works next quarter.
Signed by: reviewed monthly
Shown in the engine, on a call.
Video coming
Want this for your own pilots?
Start with your own answer. The self-assessment takes ten minutes and nothing leaves your browser.