Series: AI × Supply Chain build breakdown
Stack: Claude Code · Google Sheets · 9+ years of supply planning
The problem
At PepsiCo I always ended up rebuilding scenario logic in spreadsheets because the planning software couldn’t handle it. Doesn’t matter how expensive the ERP, every planner I know does the same thing. So I tried to hand it to Claude. Kinaxis who?
What I’m actually planning around:
- a forecast I don’t fully trust (data’s a year old, gaps, recipe changes)
- retail deals close to closing, timing still up in the air
- six-plus demand scenarios, each a real volume swing
- a packaging refresh on top
- hard limits: lead time and fixed monthly capacity
The tradeoff every planner lives in: don’t let stock be the reason growth stalls, don’t overproduce and watch shelf life die, and the whole time decide whether the cash is better in inventory or marketing.
Where it failed
Claude couldn’t live in the grey.
The question it choked on: “if we dip below minimum on one flavour for two weeks, am I ok with that, given what’s landing next month and the odds that retail deal closes?” That answer is intuition. I couldn’t get it to stop thinking in black and white.
So it’s not live. A 4, not because the build broke, but because the judgement isn’t there yet. Honest beats a demo.
What I took from it
Train the judgement before you trust the agent. I went straight for “build the agent” and the real win was realising the logic has to come first. Treat it like a new hire: teach it, watch it, hand over one decision at a time.
Next: wire it to the live run-rate and inventory numbers so it stops planning off a static sheet.
Leverage rating: 4 / 10
It didn’t work yet. But “train the judgement first” is the most useful thing I’ve learned on any build, and it’s why the supply planning agent I install is dead clear about where a human signs off.
Related: the Supply Planning Agent is the productised version of this. The prompt is forkable, email me.