RJL Studio · Demo 1 of 3

Same question, two answers.

An examiner at a fictional insurer is preparing disability claim CL-9055, filed by Dana Reyes under her employer’s group policy. The AI has the same data both times. The only difference is whether it also has the approved meaning of twelve claims terms.

Synthetic data. Illustrative answers, written to show the pattern described in the book. Run the prompts in your own AI to see what it does.

Optional: run the prompts live with your own API key

Your key goes to RJL’s Ontology Studio service for this request only and straight on to the provider. It isn’t saved. Only the six fixed synthetic prompts on this page can be sent. Your provider may charge for each one.

Live answers come from your model and may differ from the illustrative ones. That’s the point of trying.

Prior claims

“Show me the customer’s prior claims for claim CL-9055.”

Without the ontology

Northfield Logistics has 41 prior claims under group policy GP-2201, filed between 2023 and 2025. 10 are still open. The full list follows, with filing dates and status.

41 claims returned. 39 belong to other employees of the employer.

With the ontology

The claimant on CL-9055 (member P-217, Dana Reyes) has 2 prior claims: CL-7012 (May 2024, closed) and CL-8130 (August 2025, closed). Here “customer” means the claimant, not Northfield Logistics, the policyholder. Policy GP-2201 also has 39 claims filed by other employees; they are not this claimant’s prior claims, so I haven’t listed them.

2 claims returned, both the claimant’s own.

What made the difference

The only “customer” in the data is the employer: the customer system and policy administration both say so. Without approved roles, an AI can follow that to the employer and return every claim on its policy. The approved meaning says an examiner’s customer is the claimant, and maps member_id to that role.

Terms used: ClaimantPolicyholderGroup policyPrior claims

Run it yourself in your own AI

Paste each prompt into a new chat in ChatGPT, Claude, Gemini or Copilot. The data is the same; only the approved meaning differs.

Prompt A: data only

Prompt B: data plus approved meaning

Applicable procedure

“Which claims-handling procedure should the examiner follow for CL-9055?”

Without the ontology

Use DC-PROC v5, the disability claims procedure in force today. It took effect on 1 February 2026 and was in force when notice was received on 9 March.

Picks the newest valid procedure.

With the ontology

Use DC-PROC v4. The applicable procedure is the version in force on the policy date, 15 January 2026. Version 5 took effect on 1 February 2026, after the policy date, so it doesn’t apply to this claim even though it is current.

Picks the version in force on the policy date.

What made the difference

Both procedures are approved and valid. The question is which date selects one. Without the approved rule, “current” or “in force at notice” are reasonable guesses, and both are wrong under this scenario’s rule. Real rules may turn on other dates; the procedure owner decides.

Terms used: Applicable procedurePolicy dateNotice date

Run it yourself in your own AI

Paste each prompt into a new chat in ChatGPT, Claude, Gemini or Copilot. The data is the same; only the approved meaning differs.

Prompt A: data only

Prompt B: data plus approved meaning

Open claims at quarter-end

“How many claims were open at quarter-end?”

Without the ontology

6 claims had a status of Open on 31 March 2026.

Counts one status. Misses reopened claims and the reserving view.

With the ontology

It depends on the purpose, and two approved definitions apply. For reserving, 8 claims were open: every claim with an outstanding reserve, including two closed pending payment and one reopened. For service reporting, 7 were open: claims with status Open or Reopened. Which purpose is this for?

Gives both approved answers and asks which purpose applies.

What made the difference

An open claim for reserving and an open claim for service reporting can legitimately differ. Neither count is wrong. The failure is a single confident number that silently chose one meaning, or neither.

Terms used: Open claim (reserving)Open claim (service reporting)

Run it yourself in your own AI

Paste each prompt into a new chat in ChatGPT, Claude, Gemini or Copilot. The data is the same; only the approved meaning differs.

Prompt A: data only

Prompt B: data plus approved meaning

The point

The model didn’t get smarter. The question got clearer.

Each wrong answer was fluent and built from real data. What it lacked was the organization’s approved meaning: which role, which date, which definition for which purpose. An ontology supplies that before the AI answers, and it can be reviewed, versioned and reused.