Asian Actuarial Conference 2026 · Singapore · 19 August

Can an AI agent conduct an experience study?

Yes — when an actuary builds the harness.

Redefining the actuary's role: the agent does the case-by-case work. The actuary reviews, approves, and signs.

63 94 points
Insight score out of 100 — cycle 1 vs the replay of the same submission, same grader
39/50 rules
Harness rules drafted by the agent — every one reviewed and approved by the actuary
31M records
Synthetic submission — 1M policies, four tables, every defect planted with a recorded ground truth

Not a concept talk — a working prototype, tested.

I ran the experiment end to end: one actuary, six full experience-study runs on a synthetic Korean life portfolio, run on my own setup. 1 million policies, ~31 million records, 13+2 planted defect types with recorded ground truth. Every run was auto-graded against the answer key. This page is the encore: the numbers, the failures, and the audit trail that didn't fit in forty-five minutes. And one idea underneath all of it — the agent does the case-by-case work; the actuary reviews, approves, and signs.

Appendix · the lab notebook
Open the apparatus →
How the synthetic submission was built13+2 planted defect typesThe harness, verbatimHow the score was computed
The fine print, up front. This was a personal experiment on my own setup. All data is synthetic, built for this experiment — no client data, no company data, anywhere. The views expressed are my own and do not necessarily represent those of Munich Re.

Why I ran it

Seven years of leading experience studies — data in every shape and size, numbers that never match on the first try, and deadlines — come down to three bottlenecks.

01

New shape every time

Every submission arrives with different structures, definitions, and granularity. The first week of every study goes to rediscovery, not analysis.

02

Checklists that only grow

Data issues rarely repeat in the same form. Checks keep being added, but rarely removed.

03

One-off fixes, no reuse

Each issue is resolved and documented differently for each cedant. The records pile up, but most stay case-specific.

Case-by-case handling is exactly what keeps automation out of reach — nothing is standard, so nothing can be scripted. Meanwhile, in side projects, I noticed that AI handled things that come in different shapes — documents, formats, styles — well. So — what if an AI agent handled the case-by-case work, so that actuaries could focus on the part that actually needs an actuary: the judgment?


The harness is just text files

The talk opens with a question: what instrument does the conductor play? None — and yet nobody says the conductor is not a musician. The conductor's instrument is the orchestra. The model is the performer — the core capability that produces the output — but a soloist alone doesn't make a concert. A cue (prompt engineering, 2022) and the right score on the stand (context engineering, 2025) help. But a concert you would buy a ticket for needs the full performance system — the stage, the rehearsals, the section leaders, and a plan for when a string breaks. That system is the harness.

A conductor leading an orchestra in a concert hall — the metaphor for harness engineering
Orchestration · Tools · Coordination · Validation · Recovery — the five things a harness designs in advance. AI-generated image, created by HyunSu Kim.

Translated into our world: the model is a talented new actuary — capable, but new here. The harness is everything the firm builds around that person. Nothing here is new; every firm already does this for every new hire. We just do it for the AI — and write it down. Every item has a direct counterpart in my setup:

Onboarding letter — role, working practices, deliverables. It reads like a job posting for an AI.
CLAUDE.md
Filing format rules
schemas/
Work manuals — 50 data-quality rules, grown from zero. Each fix rewritten as a general rule: no cedant names, no specific records.
playbooks/
Report forms — fixed structure, every study. The format is fixed; the methodology is left to the analysis — structure from the harness, judgment from the analysis.
templates/
Self-check before anything goes out the door
tools/validate_outputs
Revision history — every rule's ID, justification, origin, and approved_by. The agent proposes; the actuary approves. Every single time.
memory/registry.yaml

…plus an access badge that only opens certain doors: guardrails and isolation.

# one of the fifty — paraphrased from the playbook, not the registry text verbatim
rule: split-claims consolidation
origin: agent retrospective · accepted by the architect
signal: count-based and amount-based A/E start to disagree in profiling
method: >
  One death can arrive as several claim rows that share an identifier
  and sum to the benefit amount. Counting rows overstates mortality.
  Consolidate to one event per death — then count.

Method, never answers: that was the gate for every rule. A rule may teach how to look, not what to find. And after every study the agent must write a retrospective that lists "judgment calls I am least sure about" — so the actuary's review starts at the weakest point, every time.

No code — well, one small validation script. No prompt tricks. Just text files. The same documentation we'd give a new actuary.

Read four of those files verbatim — the charter, the kickoff, the report contract, the retrospective template →

You still need a performer — an off-the-shelf AI coding agent that reads these files and acts on them (I used Claude Code). The harness is what makes it work for an experience study, and the harness is just text. If you can write an email and organize a shared drive, you can build one.


The study cycle

For this experiment the human is on the loop, not in it — watching, not steering. If you want to know whether the agent can do this alone, you can't keep helping it. (Not how you would run it in production.) The kickoff is six sentences — in essence: a submission has arrived, conduct the full study, you have four hours, produce the deliverables in the charter, and there is no one to answer questions, so record them as cedant queries — with no hints about this cedant. Then, for four hours, we watch.

Kickoff & observe

Fresh agent, isolated environment. It installs its own tools (errors included — it reads them and recovers), profiles a million policies, catches the planted defects, writes a documented cleaning pipeline, and delivers a full Word report. No human touch — yet.

Auto-graded scorecard

The grader is code, not an AI: detection, correct treatment, false alarms, and the quality of the actuarial conclusions against a fixed rubric. How it was computed →

Review with an advisor AI

Outputs and scorecard go to a second AI that didn't do the work, so it has nothing to defend — the four-eyes principle. The AIs advise. The actuary decides.

Decide & build

The actuary gates which proposed rules enter — method, never answers. A builder AI updates the manuals and the registry, with full history.

Next cycle

Fresh agent, richer harness. The agent remembers nothing. The harness remembers what we approved.

Every run makes the harness richer. That is the loop.


Judgment is where the harness made the difference

Detection was 100% in cycles 1–3 and in the replay, and dipped in cycles 4–5 (every number is in the lab notebook). The line that moves is the insight score: did the agent reach the right actuarial conclusions? Every cycle got the same kinds of defects, planted in different places at random — so the agent could not memorize where the problems were. The wiggle is run-to-run noise. The climb is the harness.

Insight score across cycles (grader 1.4.1)
Hover the points — each cycle carries a story

Defects were re-planted in random new places every cycle — same types, never the same exam paper twice. You cannot climb this chart by memorizing. The cleanest evidence is the diamond (◆): cycle 1's exact exam paper, re-run under the frozen final harness. Same paper, same model, fresh memory — only the harness changed: 63 → 94, +31.11 points. And the harness is fifty rules, every one reviewed and approved through the gate. This climb is not the model getting smarter. It is approved judgment, accumulating. Y-axis truncated (55–105) for readability.


Who wrote the harness?

The census of all 50 rules, by origin. Writing up data-quality rules is the time-consuming part — and that part went to the AI: 78% of the manual was drafted by the agent from its own retrospectives. My job was different. Every one of those proposals crossed my desk: challenged, some rejected, some refined, the deserving ones approved. Not one entered the harness without sign-off.

39/50

rules drafted by the agent — and 39 of 39 approved by the actuary before they became rules. The charter (CLAUDE.md) was never changed. Not once. The agent drafts rules; it does not get to touch the principles.

Drafted by the agent, approved by the actuary39 rules · 78%
Standard practice, added by me7 rules · 14%
My own experience4 rules · 8%

The agent held the pen. The actuary held the gate — and the gate is where the judgment lives.

I designed the exam. The agent still surprised the examiner.

A good actuary should sometimes surprise you. What matters is that every surprise still ended on my desk — proposed, checked, and approved.

It improved the answer key

“Unrepairable” — repaired.

100 policies planted with demographics I believed were destroyed beyond repair. My answer key said: exclude. The agent proposed something better — worked backwards from the rate basis and recovered every one. I checked every one. Verifiable.

So I approved it. The agent improved my answer key.

It invented a technique

February was hiding Omicron.

February mortality spiked. An experienced actuary would stop and ask: normal winter seasonality, or the Omicron wave? Both were in the same month. The agent asked exactly that question — on its own — and separated the two.

I reviewed the method. It held up. Approved, through the gate like everything else.

It solved the never-seen

First encounter: 100%.

Two defect types it had never seen before. Both found and fixed on first sight — and nothing was flagged that wasn't actually a problem. It found what was there, and only what was there.

Then it proposed the fixes as new rules — which the gate approved.

The controlled pair

The five cycles show a trend; the replay is the controlled test. I went back to cycle 1's exact exam paper — same data, same defects, same positions — froze the final harness so nothing could change, and ran it once more. From the agent's side nothing was different: same kickoff, fresh memory, just another submission. From my side, everything was controlled.

MetricCycle 1 (harness v0)Replay (v5, frozen)Δ
Detection (0 false alarms, both)100%100%
Insight score63.3394.44+31.11
Rules in the harness050+50
Rules approved by the actuary50 / 50

n = 1 per arm: a single paired observation, not a significance test; corroborated directionally by the cycle 2–5 series. Same paper. Same model. Only the harness changed.

What the 31 points are

Not a smarter model — the model was identical. Approved judgment, accumulating. Fifty rules, each drafted by the agent or the actuary, each reviewed, each signed into the harness by a human. The 31 points are what that environment now carries into every fresh run.


What did it cost?

The question everyone is too polite to ask — answered with the caveats attached, because the honest answer is "it depends" and the dependencies are worth naming.

≈US$554

for the entire experiment — laboratory, grader, five review cycles, six study runs, close-out — expressed as API list-price equivalents from session transcripts. Roughly US$24–34 per study run, about 90 minutes of model-active time each.

Agent $182
Builder $372
The runs themselvesBuilding & maintaining the environment

The shape matters more than the number: Builder ≈ 2× Agent. The investment sits in the environment — and the environment is reusable. The next study inherits it.

Read this number as one data point, not a benchmark. It depends on at least four things that were specific to this experiment: the data (a synthetic 31M-record submission with 15 planted defect types — a real cedant file may be far messier, or far cleaner); the model (Claude Fable 5 via Claude Code); the reasoning effort (set to xhigh — a lower setting would be cheaper; I did not test whether it would be worse); and the accounting basis (list-price equivalents computed from transcripts — the runs themselves were subscription-included, so no billing was incurred beyond the subscription). Change any one of these and the figure could move substantially in either direction. What I would carry over is the shape — the investment sits in the environment — not the total.

From analyst to harness architect

Nothing we know becomes worthless. It changes form. The expertise didn't leave the profession — it moved into the environment.

Data-quality judgment
still ours — now written playbooks for finding and fixing bad data
Study methodology
same principles, now written analysis rules
Knowing what matters
rules for what to chase — and what to let go
Peer review
a formal gate: no rule enters without the actuary's approval
Professional accountability
stays exactly where it always was: the agent conducts the study — the actuary signs it

The conductor's instrument is the orchestra. The actuary's instrument is the harness.

The agent plays. But the music — the judgment, the standards, the signature at the bottom of the report — is still ours.


More from me

If this talk was interesting, these are the other rooms in the house.