A lot of conversations about AI in manufacturing start from the wrong assumption: that the ERP is somehow the problem. It almost never is. SAP, Oracle, Epicor, Infor — these systems exist because they’re genuinely good at what they were built to do, and most manufacturers running one aren’t looking to rip it out.
The actual problem usually isn’t inside the ERP at all. It’s in the space around it — the manual work that exists specifically because the ERP doesn’t (and often shouldn’t) handle everything on its own.
Where the Real Gaps Live
Three gaps show up constantly, across almost every manufacturing operation we’ve looked at.
The first is communication. ERPs track orders and inventory, but they don’t read a supplier’s email and update themselves. Someone has to open that email, interpret it, and manually enter whatever changed.
The second is documents. Invoices, packing slips, certificates of compliance — these arrive as PDFs or scanned images, and getting that information into a structured system still usually means a person reading the document and typing the relevant fields in by hand.
The third is the in-between state of things in motion. A shipment that’s delayed, a partial delivery that doesn’t match the original PO, an order that’s been split across multiple trucks — ERPs are built around clean, structured records, and they tend to handle messy, in-progress situations poorly. Someone ends up tracking the messy reality in a spreadsheet on the side, because the ERP’s clean data model doesn’t have a natural place for it yet.
None of these are ERP failures. They’re just outside the scope of what an ERP is designed to do. That’s exactly the space where an AI agent is actually useful — not as a replacement, but as the thing that sits in that gap and absorbs the manual translation work.
What This Looks Like in Practice
This is the same logic behind the five agents in our SCM automation case study — a Communication Agent reading supplier emails, an AP Agent reading invoices and matching them against POs, a Logistics Agent tracking shipments in motion, a Procurement Agent following up automatically when something’s overdue, and a Reporting Agent pulling it all into one live view. Every one of those agents exists specifically because the ERP doesn’t do that particular job, and was never going to.
The pattern generalizes well beyond supply chain too. Recruitment screening, customer support triage, internal reporting — the same shape of problem shows up everywhere: a system that’s good at its core job, surrounded by manual work that exists purely because nothing automatically bridges the gap.
What AI Agents Shouldn’t Try to Do
It’s worth being equally direct about the limits here, because overclaiming is exactly what’s made a lot of manufacturers skeptical of “AI” pitches in the first place. An agent built to read supplier emails and extract order status shouldn’t also be the system of record for your inventory — that’s still the ERP’s job, and trying to make an agent replace core transactional infrastructure usually creates more fragility than it solves. The right framing is narrow and specific: agents handle the manual translation work at the edges, the ERP stays the source of truth at the center.
That’s also why these projects work best scoped narrowly and built incrementally — one workflow first, proven, then expanded — rather than as a sweeping platform overhaul. You can see how that plays out in more detail on our manufacturing automation page, including the specific FAQ on exactly this question of what gets replaced and what doesn’t.
If you’re trying to figure out which of your own ERP’s gaps are actually worth fixing first, that’s precisely what a free automation audit is for.