Back to case studies
Manufacturing — Supply Chain

Procurement AI for PET Raw Material Risk

PET procurement cost reduced by 4.8% in year one; one forward purchase avoided a 12% spot price spike.

12-week deployment: data structuring and signal integration in weeks 1–5, model training and back-testing against 3 years of historical purchase data in weeks 6–9, dashboard build and procurement team onboarding in weeks 10–12. First live purchase recommendation at week 13.

The challenge

Raw material cost variance was treated as weather — unpredictable and unmanageable — when in practice a significant portion of it was forecastable with the right signals and a consistent decision framework.

The structural problem was that PET price movements have identifiable leading signals — crude oil futures, PTA/MEG spot markets in Asia, seasonal demand patterns from the membrane manufacturing sector, and currency dynamics for USD-denominated purchases by TRY-cost producers. These signals are publicly available. They were not being systematically used. Procurement decisions were made by experienced buyers using market intuition and supplier conversations — which is not nothing, but which is also not consistent, not documentable, and not scalable across different buyers or market conditions. The secondary problem was inventory: without a forward price model, the business defaulted to conservative inventory positions to avoid over-committing at high prices — which periodically created supply constraints when spot prices moved against them and forward purchasing would have been the rational choice.

What we did

The approach

A PET procurement AI model was built on a multi-signal forecasting architecture that ingests crude oil futures (Brent, WTI), PTA and MEG spot price indices from Asian markets, EUR/USD and TRY/USD exchange rates, and historical PET price data from the company's own purchase records. The model generates 4-week and 8-week forward price probability distributions — not point forecasts, but confidence ranges that enable the procurement team to make buy/hold/forward decisions against quantified risk rather than intuition. A procurement decision dashboard surfaces the model's output alongside current inventory position, committed production volume, and an optimised purchase volume recommendation calibrated to the company's working capital constraints.

MARKET SIGNALSBrent/WTI futuresPTA/MEG indicesFX ratesPurchase archiveInventory andproductionFORECASTING MODEL4-week distribution modelProbability bands, not pointforecasts8-week forward modelBack-tested on purchase historyWorking capital constraintsRecommendations within liquiditylimitsOUTPUTProcurementdashboardBuy/hold/forward signalsSignal rationaleAudit trailCost variance reduction

Key findings & actions

01

Multi-signal data ingestion layer

crude oil futures, PTA/MEG spot indices, EUR/USD and TRY/USD rates, historical purchase price archive

02

4-week and 8-week forward price probability distribution model (outputs confidence ranges, not point forecasts)

03

Inventory position integration — real-time stock level and committed production volume inputs

04

Working capital constraint parameterisation — purchase volume recommendations bounded by liquidity limits

05

Procurement decision dashboard with buy/hold/forward recommendation and supporting signal rationale

06

Post-purchase outcome logging for continuous model calibration

How we worked

01

Scope

Purchase price data archive structuring, signal source identification and data feed integration, model architecture and initial calibration, dashboard design and procurement workflow integration, buyer training on interpreting probability distributions, post-deployment calibration review at 90 days.

02

Timeline

12-week deployment: data structuring and signal integration in weeks 1–5, model training and back-testing against 3 years of historical purchase data in weeks 6–9, dashboard build and procurement team onboarding in weeks 10–12. First live purchase recommendation at week 13.

03

Operating model

Procurement decisions remain with the buyer — the model surfaces a recommendation with supporting signal rationale, not a mandate. Weekly model output review embedded into existing procurement meeting cadence. Quarterly recalibration using recent purchase outcomes. Working capital parameters reviewed and updated by finance team at each quarter boundary.

Outcomes

What changed

PET procurement cost reduced by 4.8% in year one; one forward purchase avoided a 12% spot price spike.

01

PET procurement cost reduced by an average of 4.8% in year one

equivalent to a direct operating margin improvement on a cost base representing 50–65% of total manufacturing cost

02

Raw material cost variance reduced by 31% quarter-on-quarter, enabling more reliable margin forecasting

03

One forward purchase decision in Q3 (model-recommended, buyer-approved) avoided a 12% spot price spike that materially impacted competitors purchasing on standard monthly cycles

04

Inventory optimisation

model-guided purchasing reduced average PET stock holding by 18% while eliminating the supply constraint events that had periodically disrupted production scheduling

05

Procurement decision rationale now documented for every major purchase

creating an audit trail and enabling performance review against model recommendations

Governance

Trust, collaboration & governance

01

Probability distributions rather than point forecasts — buyers see the uncertainty, not just a number, which prevents over-reliance on model output

02

Model back-tested against 3 years of actual purchase data before live deployment — performance characteristics known before any real money was committed

03

Human approval retained on all purchase decisions above a defined threshold — the system recommends, the buyer decides

04

Post-purchase outcome logging creates a feedback loop: the model's calibration improves with every purchase made against its recommendations

Reframe

The model quantifies uncertainty well enough to consistently outperform gut feel — that's a repeatable process, not a crystal ball.

Across every engagement, the goal is the same: engineer a system that makes better decisions — faster, more consistently, and at scale — than the process it replaces.

Start a discovery

Most engagements begin with a conversation about context.

We do not send a proposal before we understand the problem. Start by telling us about your decision context — we will identify the highest-leverage intervention areas before any scope is agreed.