Case studies
Four programs, shown before and after.
Each study below is a composite: the operating patterns, failure modes and measured outcomes of several recovery programs, assembled so no lender, forwarder or borrower is identifiable. The plates show the operating posture before and after control; the numbers show what changed.
How to read these
Composite programs, real operating patterns
Composited on purpose
Figures are drawn from engagement patterns and aggregated, not lifted from one client file. Treat them as representative, not audited.
Before and after are postures
The paired plates describe how the program operated, not photographs of a site. The change that matters is in the controls beneath them.
Refusals are included
Each study carries one refusal the platform actually records in that shape — the negative action is part of the outcome.
CS-01 · Regional auto lender · 41 states
A dispatch queue that could not explain itself
- Assignment to first field attempt
- 6.4 days→2.1 days
- Disputes needing manual reconstruction
- 88%→4%
- Actions with recorded authority
- unmeasured→100%
SituationAssignments moved between six inboxes and two spreadsheets. Nobody could say who authorised a repossession attempt, and disputes were answered from memory.
ControlLifecycle chapters replaced the inboxes. Every transition ran through a guard that checked authority, vendor compliance and document state before the side effect fired.
OutcomeDispute responses became a query instead of an investigation, and the same guard that blocked bad dispatches produced the evidence used to answer regulators.
dispatch.create refused — vendor insurance certificate expired 3 days earlier; assignment held, compliance queue notified.
See the lifecycle chaptersCS-02 · Credit union portfolio · 9 states
Cancellations that arrived after the tow truck
- Attempts after cancellation
- 17 per quarter→0 per quarter
- Stop-work acknowledgement
- not captured→median 4 min
- Timeline reconstruction
- hours→immediate
SituationStop-work instructions travelled by phone call. A cancelled assignment could still be accepted by a vendor hours later, and the paper trail disagreed with itself.
ControlCancellation became a state, not a message. Once a lender cancels, the assignment is non-actionable: stale dispatch and acceptance attempts fail closed and the vendor receives an acknowledged stop-work.
OutcomeWrongful-attempt exposure dropped to the cases where the vendor was already on scene, and every one of those has a reconstructable minute-by-minute timeline.
assignment.accept refused — assignment cancelled at 14:02:11; acceptance token issued against a superseded version.
Inspect the audit chainCS-03 · Captive finance program · national
Twelve AI roles, none of them trusted with state
- Routine work handled without review
- 12%→71%
- Model writes to lifecycle state
- unrestricted→none
- Recommendations attributable to a role
- partial→100%
SituationAn earlier automation attempt let models write directly to records. When output was wrong, the wrongness was already durable and nobody could see which model did it.
ControlAI roles were confined to reading, drafting and recommending. Writes stayed with application commands, and each role got a declared permission boundary and escalation path.
OutcomeRoutine administrative work is now handled without review, exceptions reach a named human, and every recommendation is attributable to a role and a version.
agent.write_state refused — role A-07 holds recommend authority only; escalated to assignment owner.
Read the role boundariesCS-04 · Servicer transition · mid-market book
Money that reconciled twice
- Duplicate invoices reaching a lender
- 2.3%→0%
- Days to close a settled assignment
- 19→6
- Closures with open financial items
- common→structurally impossible
SituationRecovery proceeds, vendor payables and lender receivables lived in three systems. Duplicate invoices were caught by whoever happened to notice.
ControlOne financial spine: recovery to payable to receivable to invoice to reconciliation, with duplicate billing and duplicate payment blocked at the command level rather than in review.
OutcomeClosure now requires financial completeness, so an assignment cannot be archived while a payable or receivable is still open.
invoice.issue refused — payable P-4471 already invoiced on line 3 of batch 219; duplicate suppressed.
Open the economics modelWant the same before-and-after on your own book?
Trial 1 scopes a single portfolio slice, instruments it, and reports the same measures shown above against your current baseline.