Case Study · Retail · AI Agents

Margin slipped 180 bps. The CFO needed an answer before the board call.

We shipped a Claude-powered margin agent in 14 days. It found a fuel-surcharge anomaly two weeks before close, recovered $2.1M in run-rate procurement, and gave the procurement team the audit trail to take to the supplier the same day.

Industry: Retail · multi-format groceryGeo: APACSize: $500M+ annual revenue

Recovered

$2.1M

annualised procurement savings, signed off by Finance within 30 days

Margin

+2.4 pp

gross margin recovered between Q3 close and Q4 board call

Shipped

14 days

from kickoff to first agent flag the procurement team acted on

Context.

Challenge. Q3 gross margin slipped 180 basis points against plan. Finance suspected logistics; procurement suspected mix. With three weeks to the board call, nobody had a clean enough lineage to argue either side.

Constraint. No new tools. The existing data estate (Databricks medallion, Superset for analyst use, weekly Looker board for execs) had to stay. The CFO wanted an answer that survived an audit, not a dashboard.

Approach.

Week 1

Audit. Ten interviews across procurement, supply chain and finance. Catalogued every source of cost variance into the Gold layer and ranked the unexplained spend by absolute $ contribution. The top three buckets covered 78% of the slippage.

Week 2

Built the agent. A Claude Sonnet reasoning loop watching the daily ingestion of carrier invoices in the Silver layer, with Haiku running the high-volume anomaly classification ahead of it. Every signal carries lineage back to the raw invoice line, a confidence score, and a kill switch on the action handoff.

Week 3

Pilot in production. The agent flagged a fuel-surcharge step-change on one APAC carrier route that procurement had not contracted for. The procurement lead opened an RFP the same afternoon. The first month savings, signed off by Finance, were $174K.

Week 4–6

Hardened. Added the supplier-side audit trail (the same line-item lineage the agent uses), routed exceptions to a queue rather than the carrier directly, and trained two analysts on the prompt and override workflow.

Stack.

Anthropic Claude (Sonnet · Haiku)Databricks medallionSuperset (custom waterfall + whale chart)Slack action handoffPagerDuty for kill-switch escalations

"The agent did the easy part. The hard part was the lineage discipline. That's what Woodfrog built first, and it's why Finance signed off on the savings in the first month."

, Head of Data, APAC retail customer

Want to walk through this case?

30 minutes with the founder.

6-week engagement, fixed scope, fixed price.

Book a 30-min audit