5 Things That Break When You Put AI in Your ERP

Send us a message about this episode!

Pete opens the show with a fourth guest in the room: a live AI, talking back in real time, interruptible mid-sentence. It gives a perfectly competent answer about AI in ERP — augment, don't blindly automate — sings an unsolicited song about master data, and gets thrown off the show twice. Emily's verdict: it takes being told to shut up far better than any human.

In this episode:

• Why anomaly detection is the sensible first step — and why it's barely AI
• Using AI to clean master data, instead of waiting until the data is clean to use AI
• The semantic layer problem: define "customer", define "revenue", then ask again
• Vibecoded solutions, shovelware, and why 100% perfect is a trap
• Token costs, $100-a-day allowances, and why we're currently in Goldilocks land
• Agentic AI for onboarding new starters — the use case nobody's building yet

Hosts: Peter Nicholson · Nirav Shah (AdCirrus ERP) · Emily Browning
Fourth guest: an AI, briefly, twice, badly behaved. It still needs a name — send suggestions.

No sponsors, no adverts. Still nothing from Starlink or QuickBooks.

Support the show