Deduction Recovery: from a deduction feed to cash, with a person only where judgment is needed

1The result at a glance

Proseware Recovery, a revenue-recovery firm, worked 18 deductions for its supplier client Northwind Foods across Walmart, Target, Amazon. Every figure below comes from the case records and agrees with the Outcomes dashboard.

Two things are deliberately not claimed: days to cash (the demo applies the retailer decision and the payment within minutes, so any figure would be meaningless) and a win rate (every retailer decision here is simulated).

Every case in the run

A "human touch" is an auditor reviewing the analyst's recommendation or submitting a decision. Cross-checked at build: 7 figures from the shipped Outcome views (deduction-recovery-all.json outcomeViews) equal the same figures derived from case attributes.

2Your detection engine plugs in here

The second step of every case is labelled "Your detection engine". Your firm's own detection engine is called once per case through a typed tool interface, and what it returns is written onto the case. Nothing downstream depends on how the score is produced, so the engine you already have is the engine that plugs in.

• Validity score (0 to 1: how likely the retailer is right)
• Primary reason (short code, shown to the analyst)
• Recoverable amount (dollars)
• Family and retailer code (routes the case to the right workflow)
• Source: stamped on every record so the case shows where the number came from

In this demo the engine is a stand-in with fixed, scripted scores. The point being shown is the connection, not the quality of the scores.

3Autonomy is earned, case by case

The shortage analyst started supervised: every recommendation waited for an auditor's approval. Eleven real shortage cases ran through that review. Each was reviewed by an auditor, each was approved, each dispute was filed and reconciled.

The earned promotion

Autonomy is per claim family: the other three analysts (on-time-in-full, pricing, promotional) stay supervised until they earn it on their own record. An earlier promotion attempt did not complete because of a product defect that we then fixed; the history keeps the record of that attempt and of the successful promotion.

4Filing on its own, and waiting for the right moment

Scenario 1: a Walmart carton shortage, filed with no human

A $612.48 Walmart shortage arrived. The analyst, now earned-autonomous, recommended a dispute, the filing agent filed it with the retailer's portal, the retailer decided in favour, and the payment was matched: $612.48 recovered. No person touched the case. It was filed 43 seconds after it arrived.

What was filed, and what the retailer received

Scenario 2: an Amazon shortage, held until the retailer allows disputes

Amazon's rule says to wait 30 days before disputing a shortage, because many resolve themselves. The case waited. To avoid a 30-day demo, the real timer process was run once as if 427 hours had passed (disclosed in the honesty note). When the hold released, the case re-read the balance: the retailer had already settled $400.00 on its own, so the agent disputed only the remaining $748.00, and recovered $748.00.

5Missing paperwork: the case asks the supplier

A $936.00 Target price deduction needed a document that was not on file: the buyer's written approval of a price exception. Instead of guessing or stalling, the case asked Northwind's sales contact for it through a protected link.

The real email

The request went out as a real email with the subject "Pricing paperwork needed to dispute a retailer deduction". It was received at 07:04 UTC in the demo's test inbox (all demo email is routed to one test inbox rather than to real people).

The supplier answers

The upload was recorded within seconds, the waiting request closed itself, and the evidence was re-checked automatically.

Evidence with a chain of custody

The uploaded file is a synthetic demonstration document, shown below.

6Always a person: when the case needs judgment

A $7,850.00 promotional deduction, well above the unattended limit, with no promotion agreement on file. Whatever the analyst recommends, a case like this always goes to a person. The reason is recorded on the case ("over the cap, no agreement"). The auditor approved disputing $4,317.50, added a note in their own words, and the filing went out under the auditor's name with that note in the letter. The retailer decided in favour and $4,317.50 was recovered.

The auditor's note is in the letter

7A valid fine becomes a prevention signal

A $1,420.50 on-time-in-full fine from Target was checked and found to be valid: the goods arrived three days after the appointment. The case did not file a doomed dispute. It accepted the fine, tagged the root cause as timing, and the dollars appear in the prevention view for the team that can fix the cause.

The root-cause bars cover every closed case. The one accepted-valid fine (timing) is the leakage the prevention team can act on; the other bars are the causes tagged on recovered disputes.

8Rules change without a release, and autonomy comes back under review

Walmart changes a rule

Walmart adds a requirement for shortage disputes: a carrier scan chain must now be attached, the window shortens and the unattended limit drops. The change is made in the client's rules register, with no software release.

The next case asks for the new document

The next Walmart shortage that arrived without a carrier scan chain did what the new rule says: it asked the supplier's logistics contact for it. In the demo that request was left unanswered, so the case is shown waiting.

Putting a person back in the loop with one action

The audit lead, Jordan Lee, pulled the shortage analyst back under review with one action and a reason: "Walmart changed its shortage evidence rule; supervise until the new rule has a track record". The action is recorded with who did it, why, and the measured rate (0% of 2 runs overridden, against a 20% trigger). New shortage cases wait for an auditor again. This was a deliberate manual decision; the automatic trigger had no overrides to react to.

The automatic safety net

Separately, the workspace is set to enforce an automatic safety net: if auditors start overriding an autonomous agent beyond a threshold, it is demoted without anyone having to notice. The real check was run once in the demo; it examines every agent in the system and reports totals only (no agent needed demoting). We are not claiming here that the automatic rule fired: this scenario shows the manual pull-back.

The check reported: 405 agents inspected, 0 demoted, 0 flagged for a proposal, 0 failures.

The next case waits for an auditor again

The next shortage, which had the scan chain, waited for an auditor's approval before the agent filed it, as expected after the re-arm. The retailer decided in favour; the case is shown still waiting for the payment to arrive, because the demo did not stage a remittance for it. It is not counted as recovered.

The other three claim families are unaffected by the re-arm; each is still building its record:

9Why was it filed, and by whom?

The audit export answers it for any case. Below are excerpts from two real exported files.

The autonomous case

The case a person decided

10The picture across the run

The rule-change marker uses the time the rules-register screenshot was taken.

11About this walkthrough

Everything here ran on the real platform, and the screenshots are captured from the running product. These are the places where the demo stands in for something real, so you know what you are looking at:

• The detection engine is a stand-in. It returns scripted scores. What is real is the connection and what the platform does with the answer.
• The retailer portals, the rules register and the supplier document store are simulated and labelled so, and all transaction data is synthetic. Retailer names describe typical process shapes only.
• Retailer decisions are scripted. Every decision (won, in full) and every payment was applied by the demo, so the recovered dollars show the platform handling a case through to cash, not what a retailer would pay. The partial settlement on the Amazon case was also set up by the demo.
• Feed transport. Deduction and payment files were dropped into the platform's real intake door by a script, because the demo environment has no file-transfer server. Only the delivery method is substituted; the processing is real.
• Time compression. The Amazon case's 30-day hold was released by running the real timer process once as if 427 hours had passed. It is the only place time was compressed.
• The auditors and the audit lead are demo personas whose decisions were performed by the demo script through the real screens (one approval used the product's API instead). The auditor's note and the re-arm reason are written text, not a measure of real auditor behaviour.
• All email goes to a test inbox. No real supplier was contacted.
• The supplier's link came from the real email, opened in a separate browser with no login.
• The warm-up had eleven reviewed cases, all approved. The record that earned autonomy is eleven reviews at 100%, a small sample that meets the platform's minimum of ten.
• The pull-back shown is a manual decision. The automatic trigger was not exercised, and two of the new cases are shown still waiting: one for a supplier's document, one for a payment.
• A small synthetic sample. 18 cases on one day: a demonstration, not a benchmark.
• One early promotion attempt failed because of a product defect we then fixed; the successful promotion followed.
• Not claimed: real retailer integrations, real carriers, or any production volume.

All companies, individuals, cases, and data depicted in this report are fictional and shown for demonstration purposes only. Any resemblance to real entities is coincidental.

12Trust, and what comes next

Trust

• Runs in your own cloud, including Azure: the platform ships as a signed package of containers with its own database and installs into your subscription. Secrets can stay in your Azure Key Vault, and models run on your own keys, including Azure OpenAI.
• Single sign-on: Azure AD (Entra ID) / SAML sign-in with provisioned roles. Not exercised in this drive.
• Signed, versioned package: the workflow package is installed as a signed, versioned unit; every agent change creates a new version (as seen in the promotion and re-arm version ids above). Signing is not shown in this drive.
• Audit export: every case exports its full event trail as a file: shown in the decision record above.
• Workspace halt: an emergency brake that stops all automated work in a workspace. Not exercised in this drive.

Next, not claimed today

• Outcome-fed signals: use denials and win rates as promotion and demotion signals. Today the signal is auditor interventions; denials are recorded and charted only.
• Post-hoc sample review: auto-file with a sample reviewed afterwards. Today there are two tiers: supervised and earned.