Methodology
How the AR Recovery Audit works
Primary intent: MSP AR recovery audit — a transparent pipeline from ledger exports to a client-owned next-action plan. No debtor contact. No overclaimed automation.
An MSP AR recovery audit is not “AI that chases invoices.” It is a controlled sequence: reconcile the books, classify what is blocking each past-due invoice, gate low-confidence calls for humans, then draft language your team may send.
1. Data request & reconciliation
You export a detail-level AR aging, open invoice list, ~24 months of payment history, and customer/AP contacts — typically from ConnectWise, Autotask, or Halo plus QuickBooks Online or Xero.
We require a named as-of date and a tie-out: invoice balances must reconcile to the aging within a tight tolerance before analysis starts. Mixed dates and summary-only agings are rejected.
2. Normalize (deterministic)
Exports map into a fixed invoice schema. Age is days past due (from due date), net exposure is open balance minus applicable credits, and aging buckets are standardized.
No model invents dollars. Cents math stays in code.
3. Typed decisions per invoice
A decision model returns structured judgments: primary blocker, collectability tier, dispute signal, whether a promise looks real, contact usability, and whether escalation is warranted — plus a recommended next action.
These are classifications with probabilities, not freeform chat. The point is auditable automation with thresholds, not prose that sounds confident.
4. Confidence gates & hard stops
Rows below the confidence floor go to a human exception queue. Hard stops (for example consumer-debt signals or do-not-contact flags) override model suggestions in code.
Classifier confidence is not P(cash). We never treat model certainty as a recovery forecast.
5. Drafts from decided fields
A language model writes outreach and report narrative only from already-decided inputs: blocker, next action, evidence refs, account context.
Drafts are for your signature, in your voice. DSO Zero does not send them.
6. Human review → your approval
Nothing ships unreviewed. You receive a ranked action queue (often sequenced Week 1 / Week 2 / Weeks 3–4), exceptions, and a draft appendix.
Your staff owns send, timing, discounts, credits, write-offs, and counsel escalation.
Example next-action vocabulary
Actions are operational, not theatrical. Each queue row names a concrete next step and a human owner on your side.
- send_reminder
- resend_invoice
- request_po
- complete_onboarding
- assemble_docs
- hold_promise
- followup_broken_promise
- apply_credits
- verify_contact
- escalate_owner
- writeoff_review
See sample deliverables for report structure (illustrative, not a live client file).
Start with a data-fit conversation
Tell us your PSA and accounting stack, headcount band, and roughly how much AR sits past 60 days. We’ll reply with the intake checklist — no fake guarantees, no debtor-contact pitch.