From Zero Baselines to Measured ROI in One Quarter

The situation

A food & beverage enterprise ($500M–1B revenue, 2,000+ employees) had seven AI initiatives on the books. Five were customer-facing: a consumer chatbot, an AI social-content engine, personalization experiments. They demoed beautifully and returned nothing anyone could point to.

Meanwhile, the back office ran on effort. Demand planning lived in spreadsheets, with weekly SKU-level forecast error hovering around 30% — paid for in expedited freight and write-offs of short-dated product. Accounts payable keyed and matched roughly 2,400 invoices a month by hand.

This is the budget skew MIT documented: money follows the demo, returns hide in the back office. It came to a head when the CFO asked a simple question — what have the pilots returned? — and nobody could answer. Not because the answer was zero, but because nothing had been measured before anything was built. There was not one baseline in the entire portfolio.

The audit

The Readiness Audit ($9,500, two weeks, six stakeholder interviews) verified the suspicion. Self-reported score: 59. Verified: 46. Weakest domain: Domain A — pilots and ROI.

• Zero of seven initiatives had a documented pre-build baseline
• One had a named executive owner
• The two candidates with the clearest ROI math — demand forecasting and invoice capture — sat at the bottom of the priority list, because they demo like a spreadsheet

The audit also surfaced the engagement's best fact: the company's existing ERP and planning vendor included forecasting and invoice-capture modules the company already licensed and had never configured. The software was paid for. Nobody had asked it to work.

What we did

One senior consultant directing the client's own people and incumbent vendor — Delzey specs and verifies; the client's IT team and vendor build.

1. Baseline first. Before anything was configured, two baselines were frozen and signed off by finance: weekly forecast error of 31% MAPE on the top 200 SKUs, and roughly 360 hours per month of manual invoice matching in AP. No baseline, no ROI claim — so the baselines came first.
2. Kill the vanity pilots. The chatbot and the content engine were shut down, freeing roughly $240K in annualized spend. The surviving initiatives got named owners and kill criteria.
3. Directed sprint. A six-week pilot-to-production sprint ($35K, audit fee credited) put the two dormant modules to work. Delzey wrote the functional specs, the eval plans, and the acceptance gates; the vendor configured, and the client's IT team integrated. When the vendor pitched a platform upgrade mid-sprint, the gate did the refereeing: nothing in the acceptance criteria required it. Vendor-neutral cuts both ways — sometimes the right answer is the vendor you already have, and only what you already bought.

The results

Measured over the first full quarter against the frozen baselines:

• Weekly SKU-level forecast error fell from 31% to 25% MAPE on the top 200 SKUs — six points, worth real money in expedited freight and spoilage, and claimable only because the baseline existed first
• AP recovered roughly 90 hours per month, with about half of invoices now flowing straight through and every exception still human-reviewed
• Zero new software purchased; the entire rollout ran on modules already licensed

Six points of forecast error will never headline a keynote. It shows up somewhere better: the P&L.

"The chatbot got the applause for two years. The forecasting module we already owned paid for the whole engagement in a quarter." — Chief Financial Officer