Use caseClaims audit

Audit trail for AI decisions

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

Actions on this decision
What the record showsAI call, Jul 16, 15:41

The question as sent, the answer as shown, the model and the time. Rule 7 failed: no forced entry noted.

source · the AI call log
What the audit finds on a decision

The rule itself was too strict · the fix is a new version, past cases reviewed

01

Why it drags today

In two paragraphs

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.

Claim CL-26-18840 · HistoryCase management
July, in one line of dates. Documents, AI analyses and human decisions in order, with the logged call in front.
Audit AUD-26-Q3-07Case management
One claim, opened to its roots. The sample on its own audit case, read-only for the claims team, each piece filed under the claim it explains.

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.

02

The sample

Laura Marshall is testing the third-quarter homeowners claims where the AI gave a verdict. One of them was denied, and the customer complained.

Step 1Assistant

The population comes out of the repository

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.

Ask the repositoryAI console
The answer names the class of documents it searched. It sees only what the auditor is allowed to see.
The population is a query, so it can be run againThe auditor's sample starts from a list anyone can reproduce, and the 25 claims sit on an audit case the claims team can read but not change.
Step 2History

The decision record of the denied claim

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.

Claim CL-26-18840 · HistoryCase management
Each AI event links to its call. Each human event carries a name and a role.
The facts are already in orderNobody reconstructs the claim from mailboxes. The record was written as the work was done.
03

The replay

Laura replays what happened on the day: the call, the sources, the people, and then the rule itself.

Step 3AI console

The AI call replays as it was

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.

AI call, Jul 16, 15:41AI console
The audit reads the answer the adjuster read, word for word.
Suppose your auditor asks tomorrow what the AI told an adjuster in July. How long before you could show the answer, word for word?
A suggestion from July reads the same in OctoberThe check replays from what was kept on the day, so the audit judges the answer the adjuster actually had.
Step 4Viewer

The cited source opens at the version that was read

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.

PR-2026-0731.pdf · version 1Viewer
The citation points to the document, the page and the version.
Each check can be verified at its sourceThe auditor does not have to trust the AI's reading. She reads the same page herself.
Step 5Case management

The human decisions and the overrides

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.

Overrides in the sampleCase management
Overrides are normal. The audit checks that each has a name, a reason and the right authority.
A person stands behind each decisionAgreements and overrides alike carry a name, a date and a reason. The one that is too short is found.
Step 6Viewer

The rule is compared with its next version

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.

Homeowners claim rules, v12 and v13Viewer
Additions, deletions and changes show on one screen.
The rule is audited along with the decisionThe adjuster applied the rule of the day correctly. The finding is about the rule, and its fix is already dated.
04

The evidence and the finding

The facts are clear. Laura shares them with the risk committee, and keeps personal data out of the pack.

Step 7Viewer

The evidence pack, with personal data masked

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.

Evidence pack, CL-26-18840Viewer
The full internal version stays. The mask travels with the export.
The evidence leaves with only what the reader needsA person approves every mask before the copy goes out, and the export itself is recorded.
Step 8History

The trail is exported as it was sealed

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.

Export · CL-26-18840Case management
The export adds its own entry: who exported what, and when.
The committee reads the record itselfNo one retypes the trail into a slide. What the committee reads is what was written in July.
Step 9Assistant

The finding is drafted, then written by the auditor

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.

25 decisions replayed5 overrides, all named1 reason too short1 rule too strict, fixed in v13
Audit AUD-26-Q3-07AI console
Figures are sample values. The finding is the auditor's text.
The audit ends with a decision people can act onThe committee gets facts it can check and a recommendation signed by the auditor.
05

The possible results

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.

Traced

Every step replays

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.
Gap

A reason is missing

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.
Rule issue

The rule itself was too strict

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.
The limits we set

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.

06

What it saves

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.

25 decisionsreplayed with their AI call, their sources and the person who decided.
0 emailssearched to rebuild a decision. The record was written while the work was done.
1 exportof the trail as it was sealed, with personal data masked and approved.

Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.

Why these numbers hold

Kept on the dayTraceabilityEvery AI call keeps its question, answer, model and date, and replays months later.
Sources with their versionEvidenceEach answer points to the piece, the page and the version it read.
Decision supportA person decidesSuggestions are recorded apart from decisions. Each decision carries a name, a role and a reason.
07

For your teams

The auditor, the AI platform owner and the claims team lead all answer for these decisions, each to someone different.

Risk and compliance

Internal auditors and compliance

Signing off on a decision is hard when nobody kept the reasoning behind it.

  • Every AI call replays with its question and answer.
  • Each source opens at the version that was cited.
  • Overrides carry a name, a reason and an authority.
  • Evidence leaves with personal data masked by a person.
IT and architecture

CIOs and AI platform owners

Each model added to the stack is one more you will be asked to explain.

  • Every call is logged with provider, model and date.
  • The AI reads only what the user is allowed to read.
  • The model can change per use, with no redeploy.
  • Your servers or your cloud: the logs stay where you host them.
Business operations

Claims team leads

An audit lands on decisions your team took months ago.

  • The history shows what each adjuster opened and decided.
  • A short reason is found by the audit, before a complaint.
  • Rule changes are versioned and approved.
  • The audit case is read-only for your team.

Where to start

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

Replay one of your own decisions

Bring a recent AI-assisted claim and your audit questions. We will show your audit and IT teams the record they would read.

The screens on this page are drawn to show one step at a time, with example data.

Uxopian Software helps organizations put AI to work across the systems they already run, without giving up control of their data, their infrastructure or their critical processes.