Most carriers with a capable IT team ask whether they could build claims agents themselves on top of the core system they already run. Often they could. The useful question is what that build actually contains, beyond the part that looks like the demo.
What a claims workbench is made of
and writes a summary"] end subgraph H["What production needs"] H1["Policy as of the loss date"] H2["Rules people can read and test"] H3["A decision record"] H4["Grounding and citations"] H5["Autonomy levels, hard caps,
pull-back, feedback"] H6["Payment controls and authority"] H7["Traces, prompt versions,
PII redaction, budgets"] H8["Events to underwriting,
reinsurance, portfolio"] H9["A workspace people
actually use"] end V --> H style V fill:#f4f6ff,stroke:#a78bfa style H fill:#fff4e2,stroke:#f4a03a
The demo is a small fraction of the work. The rest is what makes a regulator, an auditor and a claims executive comfortable letting the demo near a real claim.
Three paths
| Path | Strengths | What it costs |
|---|---|---|
| Build in-house | Fits your process exactly; you own the roadmap | Every item in the right-hand box above, built and maintained; model-risk sign-off on your own controls; a team that stays |
| Extend the core system | One vendor; data already there | Agents bounded by what the vendor's configuration allows; rules and decisions often still opaque; connections to other products by batch |
| Adopt a workbench | The controls and record exist; agents work inside them on day one | Integration with your core and policy systems; adapting rules and templates; change management |
Questions to ask of any option
- Can a claims manager read, test and publish a rule without a developer?
- Can you replay why any claim was referred, held or denied, with the facts?
- What stops a model's invented number reaching a reserve or a letter?
- What can an agent do without a person, who decides that, and how is it reversed?
- Does a payment an agent requests face the same controls as a person's?
- How does underwriting learn about a large loss, and how fast?
- Do the tests ever call a real model?
InsightCW's answers to each are in this blog: Rules Studio, the decision record, grounded figures, the autonomy dial, payment controls, claims into underwriting and governance.
When building is right
If your claims process is genuinely unlike anyone else's, your team has built and run governed AI in production before, and you will fund the controls as well as the model, building can be the right call. Count the right-hand box before deciding.