Case study 07 · commodity should-cost platform

07 / 10

Prove the price.

A supplier says 'prices up 8%, blame steel and fuel' and most companies just pay, with no way to check. This breaks a product down into what it's actually made of and tests each piece against real market prices, so you can see exactly how much of a hike is justified and how much is padding. Built and owned end to end.

should-cost · mining client~$401M
copper ▲ +18%diesel ▲ +6%hrc steel ▼ -4%freight idx ▲ +2%
every gap tied to the driver that movedthe gap · ~$19.3M
every price traced to market datamillions in unjustified hikes caughtbuilt and owned end to end

the problem

A supplier writes: 'prices up 8%, blame steel and freight.' Procurement teams had no defensible way to answer. Did steel actually move that much, on that lag, for that share of the product's cost? The analysis existed only as one-off spreadsheet models that died with the project that built them.

Worse, nobody could see commodity exposure across a whole spend base (which categories ride on diesel, which on copper), so every market shock became a fire drill instead of a forecast.

what I built

A tool that tells whether a supplier's price increase is fair, or whether you're being overcharged. A supplier emails: 'we're raising prices 8%, blame steel and fuel.' Is that real, or are they padding their margin? Most companies had no way to check, so they just paid. This breaks a product down into what it's actually made of (so much steel, so much labor, so much fuel) and checks each piece against real market prices, month by month and supplier by supplier, so you can see exactly how much of a price increase is genuinely justified and how much is the supplier helping themselves. The gap becomes a negotiation pack you can defend line by line, and stops increases that were never justified in the first place. I built and owned how it came together, end to end.

how it works

01

Canonicalize the spend

Raw PO and invoice files normalize into a standard shape, and repeated commercial relationships (same item, same contract, month after month) are grouped so prices can be tracked through time.

02

Decompose into drivers

Each category splits into cost components (materials, labor, energy, freight, margin), each with a driver, a pass-through share, and a time lag. AI suggests structures from the data; humans approve every one before it counts.

03

Map drivers to markets

Every driver matches against a large index catalog: exact where possible, proxy where necessary, and explicitly 'held flat' where no honest match exists. Coverage and confidence stay visible instead of buried.

04

Compare expected to actual

The engine rolls indices through the cost structure to compute what each month should have cost, compares it with what was paid, and attributes every gap to a driver, a supplier, and a month, traceable down to the source PO.

design decisions

Traceability over cleverness

Every variance drills to its source transaction. In a negotiation, a number you can't trace is a number you concede.

Cost models governed outside the codebase

Client cost structures live in stakeholder-editable templates that regenerate the client model through validation. Domain experts own the model; engineering owns the rails.

AI proposes, never asserts

Every AI-suggested structure or mapping carries provenance and confidence, with deterministic fallbacks when the model is unavailable. The numbers clients see are mechanical rollups, not generations.

Honest gaps beat false precision

Unmapped spend, proxy matches, and held-flat drivers are surfaced as coverage audits. Saying 'we don't know this part' is what makes the rest believable.

what it changed

traceable

every price checked against real market data, down to the source

millions

in unjustified increases caught on a single client

owned

built and owned, end to end