Use caseClaims audit
Three months on, an auditor asks how a theft claim was denied. What did the AI suggest, on what, and who agreed? The answer is in mailboxes.
Every AI call keeps its question, answer, model and dateEach answer points to the document version it citedEvery human decision carries a name and a reason
The question as sent, the answer as shown, the model and the time. Rule 7 failed: no forced entry noted.
source · the AI call logThe rule itself was too strict · the fix is a new version, past cases reviewed
What did the AI suggest, on what, under which rule, and who decided? Internal audit, the risk function and sometimes the state regulator ask those four questions about every claim decision the AI touched. Claims teams now use AI to check and summarize files, and the answers do exist. They are just spread across mailboxes and memories.
This record is written while the work is done. Every AI call keeps its question, its answer, its model and its date. Every answer cites the version of the document it read. Every human action carries a date and a name. The auditor reads the facts of the day, and draws her own conclusions.
Audit, risk and claims teams all know these screens. Laura Marshall's audit holds because each one leads to the next: the sample comes from the repository, each AI call replays as it was, each source opens at the version that was cited, and each human decision has its name. The three acts below follow her audit until the finding is signed.
Laura Marshall is testing the third-quarter homeowners claims where the AI gave a verdict. One of them was denied, and the customer complained.
Laura asks the repository, in plain words, for every homeowners claim decided between Jul 1 and Sep 30 with an AI verdict. It returns a table of 4,812 claims and offers an Excel export. She draws her sample of 25 in her own audit tool, weighted toward denials and overrides, and adds claim CL-26-18840 on purpose.
Laura opens the history of CL-26-18840. It lists every event in order: documents received, each AI analysis, each human action, with a date and a name. Every action is timestamped and sealed with a chained hash, so a later correction appears as a new entry.
Laura replays what happened on the day: the call, the sources, the people, and then the rule itself.
Laura opens analysis 2. The console keeps the question as it was sent, the answer as Jason saw it, the model, the provider and the time. Ravi Patel, who runs the AI platform, confirms the model listed was the one in service that day.
The answer cites page 2 of the police report, version 1. Laura opens it in the viewer at that page: the officer wrote that no damage was found on the door or the windows. The AI cited what the report says.
On CL-26-18840, Jason followed the AI verdict. The history shows he opened the police report and the policy before signing, and wrote a reason in his own words. Across the sample, Laura lists the five overrides with the reasons given, and asks Carmen Diaz, the team lead, about one of them.
The homeowners claim rules are a document in the repository, with their versions. In August, the claims administrator rewrote rule 7 after three complaints. Laura compares version 12 and version 13 side by side. Under version 13, CL-26-18840 would have been set to Attention required and looked at again.
The facts are clear. Laura shares them with the risk committee, and keeps personal data out of the pack.
Laura exports the cited pages of CL-26-18840. The AI detects the personal data and proposes masks: name, street address, phone and the bank account number on the loss notice. Amy Nguyen, the privacy officer, reviews the proposal and keeps the city. The masked copy is produced for distribution; the claim file itself is untouched.
For the committee's file, Laura exports the history of the claim and the log of its two AI calls. The log exports as it is: each entry dated, attributed and chained to the one before.
Laura asks the assistant to draft a summary of the audit case. Each statement links to the piece it comes from. She checks them, adds her recommendation in her own words, signs and sends it to the risk committee.
Each claim in Laura's sample comes out of the replay in one of three states, and each state asks the claims team for a different kind of work.
The call, the sources it cited and the human decision are on file, dated and attributed.
After that: the decision stands, and the audit says why.An override with a reason too short, an authority not cited. The audit names the step and the person.
After that: the team lead completes the file and the practice changes.The adjuster applied the rule of the day. The rule was the problem.
After that: a new version of the rule, and a review of the claims decided under the old one.It decided no claim. Across the nine steps on this page, every denial, every override and every finding is signed by a person, and the record shows who.
The AI checked, cited and summarized, and every call left its entry. The control is in what is kept, and in who may read it, so an audit months later reads the facts of the day.
Counted on Laura's quarter: one audit case, twenty-five claims, one denied theft claim. A bigger sample or a different line of business gives other totals.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
The auditor, the AI platform owner and the claims team lead all answer for these decisions, each to someone different.
Signing off on a decision is hard when nobody kept the reasoning behind it.
Each model added to the stack is one more you will be asked to explain.
An audit lands on decisions your team took months ago.
With one decision. Take a recent AI-assisted claim and the questions your auditors ask, and check that the record answers each one. One claim replayed end to end tells you quickly whether the record holds.
Your claims system, your repository and your audit tools stay as they are. The record is kept alongside them.
Ready to try
Bring a recent AI-assisted claim and your audit questions. We will show your audit and IT teams the record they would read.