Revenue Leakage Auditor
Demo · Fieldline is a made-up company · The AI agents are real. Their inputs are mock data, but the outputs are actual

500 accounts. $18.0M ARR. Somewhere in here, money is leaking.

An AI auditor for a software company's billing. It finds revenue the company is failing to collect, prices every leak to the dollar, and drafts the fix. Nothing changes until a human approves it.

Replay of a real recorded agent run · synthetic company, real agents.

Recorded run

Replay of a real recorded agent run · synthetic company, real agents.

Leakage found
$0/yr
Revenue billed short or never collected, per year
At-risk ARR (separate, not leakage)(not leakage)
$0
Not lost yet, but likely to be

Each square is one customer account.Click any coloured square to see the evidence, the math, and the drafted fix.

A0Waiting to start

Five ways money leaks

Run the audit to fill this in

    Code does the searching and the money math. AI does the judgment calls code can't make. A human makes the final decision.

    Four stepsrecorded 0:00

      code deterministic Python/DuckDB · AI Claude Code subagent · you human approval. Times are recorded wall-clock time, including subagent startup.

      Nothing to approve yet. Run the audit (or skip to results) and the fixes the AI drafts will land here for your approval.

      The AI drafted a fix for every confirmed leak: a customer email, a payment retry, or a billing change. You decide. Approve or reject each one and watch the recovered revenue add up. It's a demo, so nothing is actually sent.

      Recovered by your approvals$0 of $0
      Forgone $0At-risk ARR protected $0 of $0
      Friction = how much the fix could annoy the customer. Approving only moves a proposal to the send queue.

        How do we know it works? The fake company was generated with planted leaks and decoys: cases that look exactly like leaks but are legitimate, like a discount a VP approved in a CRM note. The answers were sealed in a key the AI never saw, then the results were graded against it.

        No grade yet. Run the audit (or skip to results), then see how its findings compare with the hidden answer key.

        Run 1 vs Run 2

        What changed between runs

          Misses, and what the agents caught

            Why the model never touches the money math

            SQL finds every candidate and Python computes every dollar, with a formula you can read in each drawer. The model only does what code can't: read a CRM note, judge the risk of a fix, and write to the customer. Any draft containing a dollar figure the code didn't produce is rejected.

            Assumptions in the dollar math

            Labeled as assumptions in every drawer, too.