Turning a report into a co-pilot

Sodexo relied on a legacy third-party tool, provided through iTradeNetwork, to confirm that distributor POS files were arriving correctly. That tool was being sunset and no longer worked reliably — leaving the client with no way to verify incoming files at all.

This wasn't a cosmetic gap. If a distributor's file doesn't reconcile with what's expected, the invoice built on it is wrong.

Engineering and PM's first instinct, working toward a minimal viable replacement, was to simply expose the underlying data as a report. It made sense — it was the fastest thing to ship. But that instinct undersold the actual problem.

A raw report only protects the business if someone reads it every day and correctly judges what's wrong from a table of numbers. For a non-technical reviewer, that's not a realistic expectation — and the first demo confirmed it. The client responded well to each incremental improvement we showed them, but kept surfacing the same hesitation: they needed to see that something was wrong, not infer it from raw figures. That gap — between "the data exists" and "someone trusts it enough to act on it" — is where the AI idea came from.

Before AI was anywhere near the conversation, I mapped what the job actually looked like for the people doing it today — ops and finance reviewers checking dozens of distributor feeds every morning, with nothing but a table between them and a billing mistake.

None of this is a willpower problem. It's that the system hands people a spreadsheet and expects judgment it never helps them form. That's the exact gap the AI layer needed to close.

I didn't want this comparison to stay theoretical, so I built both flows all the way to a clickable prototype — same data, same table, two very different experiences of using it.

There was a middle step too, between these two. Before AI, I tried color-coded flags and flexible period comparison — any period against any other, not just the prior one. It tested better, but the hesitation never fully went away: people could now see a flag, but still had to leave the screen to understand what to do about it. That's what pushed the pitch the rest of the way to an AI layer.

It would have been easy to let the AI scope creep — auto-resolving alerts, drafting distributor emails, predicting future delays. I deliberately minimized it to four capabilities: enough to close the trust gap without overreaching on a first release.

This scope only got set after internal alignment — Engineering, PM, and our Director all had to agree on the efficiency case before it went anywhere near the client. When we did show it, the client confirmed it was exactly what they needed.

Everything above is built, not just mocked up. A toggle in the running prototype switches between the plain-report version and the AI-enhanced version, so the comparison is something you can feel, not just read about.