Preparing supplier invoices for approval

Multi-site distribution company

Decision studied

Reduce invoice preparation time without touching payments.

The company, which has 80 employees, processed 2,400 supplier invoices a month. Documents arrived in several email inboxes, and searching for purchase orders caused delays and reliance on outsourced administrative work.

Approach

  1. 01

    Frame

    The finance director aimed to cut preparation time by 30%. Payments remained approved by authorized staff. The initial scope covered ten regular suppliers.

  2. 02

    Investigate

    I reviewed 500 invoices, the hours billed by the contractor and the penalties actually paid, to determine whether the errors came from reading the documents, from accounting codes or from missing purchase orders.

  3. 03

    Explain

    Two explanations were compared: manual data entry, or the lack of shared documents between teams. The diagnosis led to centralizing documents before choosing an extraction tool.

  4. 04

    Build

    Four options were priced: the status quo, a standard form, rules-based automation and AI-assisted preparation. Structured invoices remain handled by rules where that is enough.

  5. 05

    Test

    Across 400 invoices, the tool first ran alongside the usual process. The thresholds were at least 99% accuracy on critical fields and 30% time saved, review included. 398 invoices were processed without error, and time dropped from 10 to 6 minutes.

  6. 06

    Decide

    The director approved a budget of $21,000, including $6,000 for strategy. The extraction is used to prepare validation and triggers no payment. Unusual invoices go back to manual processing.

  7. 07

    Steer

    The monthly review compares the expected savings with the contractor’s invoices and penalties. Any new document type is tested before it is added to the scope.

Test results

Invoices processed without error on critical fields398 invoices without errorThreshold: 99%0%50%100%2 errors out of 400 (0.5%)Time per invoice: 10 min to 6 min (40% drop, threshold: 30%)
Accuracy on critical fields: 398 invoices out of 400 processed without error, against a 99% threshold. Time per invoice goes from 10 to 6 minutes, review included.

What AI contributed

AI helped me identify exceptions and build the test sets. In day-to-day use, it reads unstructured documents and suggests the matching; a manager checks the amounts and references.

Business model (projection)

The projection includes 600 hours a year of subcontracted work eliminated at $55 an hour, or $33,000, and $6,000 in avoided penalties. Internal hours freed up are excluded. In the prudent scenario, both gains are halved.

Annual projection for this case
Initial investment$21,000 CAD
Annual gains$39,000 CAD
Costs to subtract$9,000 CAD
Net annual gain$30,000 CAD
Simple payback8.4 months (at steady state)
Prudent scenario$10,500 CAD net, payback in 24 months

Annual projection built from the results of the engagement. It is not a guarantee of performance.

A comparable challenge?

Describe the decision to make and the data available. I will suggest an approach to prepare it and the evidence to gather before deciding.

Let's talk strategy