case study / anonymised

A real business runs its back office on an AI ops system.

A Western Australian wholesale food distributor runs day-to-day operations on a custom AI system I built and supervise. Not a chatbot bolted onto a website. An operational brain on top of their existing ERP and Microsoft 365.

the shape of it

Two halves on top of the systems they already had.

One half is a library of 34 packaged workflows that do real recurring jobs: apply supplier cost changes across staggered price tiers, process emailed orders into the ERP, reconcile every supplier invoice line-by-line against its purchase order, draft the monthly rebate claims, forecast demand, build per-store reports.

The other half is a 49-page operations dashboard the whole team logs into, with five-tier role-based access, two-factor auth, and an AI assistant on every screen. It is not a reporting skin. The team approves price batches, activates promotions, works the retention list, and runs the warehouse pick queue from it.

before

The usual SMB pain, at volume

  • Pricing maintained by hand across wholesale tiers and RRPs, drifting in spreadsheets.
  • Orders arriving as forwarded emails, PDFs, and spreadsheets, each retyped into the ERP.
  • Supplier invoices eyeballed against purchase orders, or skipped when time ran out.
  • Weekly and monthly reporting assembled by hand.
  • Rebate claims filed late or not at all. Unfiled claims are margin left on the table.

after

The grind runs unattended

  • 44 scheduled jobs fire on their own: price lists, cost changes, reports, statements, claims.
  • Over a representative 14-day window: 3,361 successful runs against 10 failures, roughly 99.7% reliability.
  • Every invoice is now diffed line-by-line. Every claim is drafted on cadence.
  • Churn risk, stock problems, and cost drift surface early instead of late.

the real headline

The biggest wins were leakage, not speed.

The most valuable jobs were the ones quietly being skipped because they were too tedious to do consistently: checking every supplier invoice, filing every rebate claim, keeping every related product on the same margin logic. Automating those recovers money, not just hours.

On top of that sits capability that did not exist before at all: per-store analysis, churn scoring, demand forecasting, and competitor intelligence that nobody had the hours to do, now running on a cadence. The roughly 40 hours a week of admin recovered is real, and it is still the smaller half of the story.

scale

Measured off the live system.

Live and in daily use~4.5 months
Scheduled unattended jobs44, at ~99.7% success over 14 days
Packaged AI workflows34
Operations dashboard~49 pages, 10 sections, 5 access roles
Admin recovered~40 hours a week (operator estimate)

why it is safe

The hard part was never the AI.

The AI never writes to the live database directly. It works through four purpose-built capability servers that expose narrow, read-only-by-default operations. Every write to live data goes through a reviewed script that runs dry-run first, shows its output, and executes only on explicit human approval, with everything logged to an audit trail. Trust comes from the audit trail, not from hoping the model behaved.

The genuinely difficult engineering was this boundary layer: safe ERP access, reversible writes, and five roles locked down fail-closed. The AI doing the work was the easy half.

honest limitations

What it does not do.

the reusable pattern

The domain changes. The shape does not.

  1. Wrap the systems you already have. No rip-and-replace.
  2. Find the recurring admin and put it on a schedule.
  3. Hunt the leakage, not just the time.
  4. Gate every write, log everything, dry-run first.
  5. Package it so a non-technical team can use it.

If your business runs on an ERP, has a mailbox full of orders and invoices, and a person spending afternoons on admin that follows the same steps every time, this pattern applies to you.

Start a conversation Back to work page