Key Highlights

  • Epicor Director of AI Product Management Olaf Schweinsberg joined analyst Shelly Kramer on The Age of AI podcast to unpack what it actually takes to move from AI interest to AI readiness.
  • Epicor’s readiness roadmap starts with data: auditing and updating stale SOPs, cleaning outdated tables and dashboards, and using AI itself to extract context from documentation before deploying any agent.
  • Epicor is betting the real differentiator won’t be the underlying LLM — it will be the application layer. Epicor Prism is built on vertical ontologies for manufacturing, retail, and distribution, and enforces rule-based, policy-driven guardrails on every input and output.
  • Epicor does not retain or train models on customer data and requires the same commitment from its model providers.
  • AI-driven data mapping is compressing ERP migration timelines: customer Cornell Pump implemented Epicor Kinetic ERP in two weeks through the AI-powered Ascend program, and Epicor’s broader target is 90-day-or-less implementations.
  • Epicor’s Cognitive ERP vision rests on three pillars: Agent Foundry (customers building their own agents inside a “headless” ERP, rolling out to select customers next month), an Agent Marketplace (customers sharing agents built on Epicor), and Community Insights (aggregated, anonymized data, including automotive aftermarket parts data, used to surface competitive and macroeconomic signals).

Most enterprises have already answered the “should we do AI” question. The harder question, the I posed to Epicor’s Olaf Schweinsberg on a recent episode of the Age of AI,  is whether they’re actually ready. Schweinsberg, Director of AI Product Management at Epicor, brought a perspective shaped by an early career in supply chain data analysis and a stint building machine learning at Amazon before joining Epicor to lead what he calls the “workflow, execution, and guided workflow network” for ERP.

That grounded, ROI-first posture shows up in how Epicor frames its own AI investment. Big Tech, Schweinsberg noted, can fund a bet for years without demanding near-term return. Epicor doesn’t have that luxury, and doesn’t want it. Every AI investment is measured against a simple bar: does it save customers time or make them money. AI isn’t positioned as Epicor’s headline feature; it’s positioned as the thing that makes the core ERP better.

Data First, Always

When I asked Schweinsberg where customers should start, he didn’t hesitate: you start with context. That means understanding both how AI models work and, more importantly, what data will feed them, how sensitive that data is, who’s reading the data versus writing it, and which system that data lives in. A large share of that context sits in outdated SOPs and training documents (yep, you’re nodding about now, aren’t you?), and AI is particularly good at extracting and refreshing that material quickly. The unglamorous but essential second half of the equation: archiving what’s no longer needed, removing unused tables and dashboards, and mapping workflows before layering AI on top. Skip that step, both agreed, and results become unreliable — and token costs climb accordingly.

The Application Layer, Not the Model, Is the Differentiator

Schweinsberg argued foundation model performance is beginning to plateau as the industry exhausts publicly available training data, and that open-weight models are closing the cost gap with proprietary ones — and he’s right. This is a very real problem for the industry as a whole. That’s why open-weight models are attractive, and it’s certainly the case for Epicor Prism, the company’s application layer built on vertical ontologies for manufacturing, retail, and distribution. Prism is designed to make an underlying LLM safe, precise, and useful for a specific industry context, regardless of which model sits underneath it.

And I was glad to discover that for Epicor, security is inseparable from that pitch. Prism has been designed to intercept communication to and from the LLM and validate inputs and outputs against policies set by both Epicor and the customer. Schweinsberg reiterated that Epicor does not retain or train on customer data, and requires equivalent commitments from its model partners.

ERP Migrations, Reconsidered

Migrations have a well-earned reputation for being long, expensive, and painful, but AI-driven data mapping and translation are changing that math. As we discussed, this Schweinsberg pointed to customer Cornell Pump, which implemented Epicor’s Kinetic ERP in a fairly mind-blowing two weeks through Epicor’s AI-powered Ascendprogram. He was careful to note that timeline isn’t typical; most customers aren’t as prepared as Cornell Pump was. But Epicor’s stated goal — a 90-day-or-less implementation — signals how much the cost-benefit calculation has shifted for organizations still sitting on legacy systems. Two-weeks is wildly impressive, but to my mind, 90-days or less is pretty compelling as well.

What’s Next: Cognitive ERP

Looking ahead, Epicor is building toward what it calls Cognitive ERP across three components. Agent Foundry lets customers build and deploy their own agents directly inside the ERP — a “headless” approach that also supports third-party coding agents and harnesses — and is rolling out to select customers next month. An Agent Marketplace extends Epicor’s existing customer community by letting users share agents they’ve built with one another. And Community Insights aggregates anonymized data across Epicor’s customer base — Schweinsberg cited automotive aftermarket parts data as an example — to surface competitive benchmarks and broader economic signals, such as inferring new-vehicle sales trends from aftermarket parts demand.

The throughline across the conversation: AI readiness isn’t a single leap. It’s a series of achievable, sequential steps — get the data foundation right, experiment safely, modernize what’s overdue, and build toward agentic capability from there. The organizations that start now, Schweinsberg and Kramer agreed, will be the ones ready when the rest of the market catches up.

Watch the full Age of AI episode of my interview with Olaf Schweinsberg here:

Read more of my coverage:

DXC’s Customer Zero Journey: Inside the Amazon Quick Rollout Powering 115,000 Employees

Connectivity, Context, Control: Why the Data Layer—Not the Model—Decides Whether Enterprise AI Works