Skip to the four decisions
Ghil · Labs Open a model

Back-office decision models · 04

Back-office work is
either reviewed
or rubber-stamped.

Four working models. Not slides. Each one takes a moment where a lean back-office team gets buried — a contract queue, a month-end close, a vendor list, an expense report — and opens it up: every assumption on screen, and exactly which decisions stay with a person, stated before you see a single result.

Start with the first case What it won't do for you

// no install · no login · runs in your browser

Case 02 · Finance Illustrative
Invoices / month4,680
Reviewed by a person18%
AP reviewers8
Branches23
Processing cost$14.80 / invoice

// finance blames AP, AP blames the branches,
// the branches blame the vendor's PDF.
// the queue doesn't care whose fault it is.

01  Starting point

Begin with the queue that's actually stuck, not the software category.

02  Assumptions

Every threshold and policy rule is visible on screen and editable.

03  Boundary

What stays with a person is stated before the model runs, not after.

Index of decisions

Four queues,
one person short

All four companies are invented. The situations are ones any finance, legal, procurement, or audit team would recognize, the figures are illustrative, and the model behind each is fully interactive. None of it is presented as a client result.

Northline Benefits Technology · fictional benefits-tech company

“Sales closed it in March. Why is it still sitting in legal in May?”

Northline is a fictional 720-person company with $8.4M in annual recurring revenue waiting on a signature. One general counsel and one contract manager read the same liability, IP, privacy and termination language every week — not because each agreement is unique, but because familiar exceptions keep showing up in unfamiliar paper.

The model checks every clause against a written playbook and drafts the fallback language counsel would approve anyway. It does not decide what the company will accept — it decides which nine agreements out of ninety-two actually need a lawyer's time this week.

  • Contract review
  • Clause playbook
  • 92 agreements / mo

Open the modelPlain-language guide

Lakebridge Home & Hospitality · fictional multi-location distributor

“Four thousand invoices a month. How many of them actually needed a person?”

Lakebridge is a fictional $410M distributor closing the books across 23 branches. Every invoice starts as an untrusted document — vendor names vary, PO numbers go missing, one PDF can carry freight, equipment and tax lines that each post differently — and stays untrusted until someone reconstructs the purchase by hand.

The model captures, matches and codes the invoices that agree with themselves, then routes only the mismatches to the eight reviewers who are already behind. Month-end close doesn't get faster because people work harder; it gets faster because normal invoices stop needing them.

  • AP automation
  • 3-way match
  • 4,680 invoices / mo

Open the modelPlain-language guide

Veridian BioSystems · fictional medical-device manufacturer

“Nobody missed the renewal date. They missed the one that mattered ninety days earlier.”

Veridian is a fictional $640M manufacturer sourcing regulated components, lab services, software and facilities from 100 vendors tracked across six spreadsheets. A renewal date sits on a calendar; the non-renewal notice window that actually decides your leverage sits on page 17 of a master agreement nobody re-reads until it's too late.

The model reads renewal, pricing, insurance and termination terms out of the contract itself, checks new quotes against the signed baseline, and finds the last date a counteroffer would still work. Five buyers can't hold a hundred clocks in their heads — the model can, and it still doesn't send anything without a buyer's approval.

  • Vendor risk
  • Renewal windows
  • 100 active vendors

Open the modelPlain-language guide

Rutherford Advisory Group · fictional professional-services firm

“We sampled five percent and found nothing. Is that good news?”

Rutherford is a fictional $285M firm generating 18,600 card and expense transactions a month across a remote sales and delivery team. A four-person audit staff can inspect roughly one in twenty of them after the quarter closes — enough to catch a bad receipt, not enough to notice three sub-$1,000 charges made eleven minutes apart.

The model screens every transaction against policy, timing and merchant history, links the ones that only look separate, and holds the high-confidence exceptions before the card settles. It flags a pattern worth a human's attention — it does not decide that anyone did something wrong.

  • Continuous audit
  • Pattern linking
  • 18,600 transactions / mo

Open the modelPlain-language guide

Method

What this
won't do for you.

A tool that hides where the automation stops is a liability wearing a dashboard. Every one of these four opens with exactly what the software will extract, score, and draft — and exactly what stays with a named person — before it shows you a single result.

Payment release, signature authority, supplier outreach, and any determination that a person did something wrong: none of that moves without someone accountable on the other end. That boundary isn't a caveat at the bottom of the page. It's the actual product.

Use these to see where your own queue would land, structure a pilot, or take a sharper question into a systems review. Not to replace the person who signs.

  1. 01

    Inputs are visible

    Thresholds, playbooks and policy rules sit on screen. Defaults exist so the page loads, not so you trust them.

  2. 02

    The boundary comes first

    Each case states what the software prepares and what a person retains before the workbench opens.

  3. 03

    Evidence, not verdicts

    Every flag carries its source document. A confidence score is a reason to look, not a finding.

  4. 04

    A person still signs

    Payment, contract language, supplier commitments and employee action need an accountable owner. The model narrows the queue; it does not clear it.

Across all four cases

What the software touches.
What stays with a person.

The four workbenches share nothing but this table — their volumes, industries and costs stay separate on purpose, because adding a fraud score to a contract deadline would mean nothing.

Legal · agent prepares
Clause varianceredline drafted
Legal · person retains
Signaturefinal language
Finance · agent prepares
3-way matchexception packet
Finance · person retains
Payment releaseclose certification
Procurement — agent preparescounteroffer draft
Procurement — person retainsoutreach, amendment
Audit — agent preparespattern, hold flag
Audit — person retainsfraud call, recovery

Four ways in

Which queue is
already stuck?

Pick the case closest to yours. Two minutes to understand the situation, then the model is right there with nothing to install.

Back to the four decisions