Your manufacturing team just invested in an agentic forecasting system. It’s supposed to autonomously predict demand swings, adjust production schedules, and flag supply chain risks in real time.
On day one, it asks for production data.
Your ERP delivers a 43-column spreadsheet you can run on Tuesday mornings.
That’s not a software mismatch. That’s a systems architecture problem. And it’s why companies serious about agentic AI in manufacturing quickly discover they need something their ERP can’t provide: operational software built for autonomous decision-making.
The Gap Between Agent Expectations and ERP Reality
Agentic AI systems operate on fundamentally different principles than the software they’re trying to automate.
Traditional manufacturing software - your ERP, MRP, WMS - was designed for human interaction. You run a report. You read it. You make a decision. You enter a transaction. The system records it. Next day, someone runs another report.
This works fine when humans are in the loop. Slower, sure. But deliberate.
Agentic AI breaks that cycle. Agents need to:
- Access data continuously, not batch-process reports
- Make decisions in minutes, not days
- Execute transactions automatically, without human approval gates
- Learn from outcomes immediately, feeding results back into planning
Your ERP was designed for the opposite: bounded data access, batched processing, approval workflows, and periodic audits.
When you try to bolt an agent onto that architecture, the agent becomes a frustrated translator. It’s asking for crisp, real-time production data. It’s getting weekly ETL extracts. It’s supposed to autonomously adjust schedules. It’s staring at a system that requires a supervisor to approve schedule changes through a form.
The vendor demo never shows you this part.
Why This Matters Right Now
Manufacturers are exploring agentic AI at scale. Some have already deployed pilots - forecasting agents, supply chain agents, quality monitoring agents. The early wins are real. Demand prediction improves. Lead times drop. Quality issues surface faster.
But the moment you move from pilot to production, you hit the architecture wall.
Your ERP can’t feed the agent in real time. Your approval workflows are human-centered, not agent-centered. Your data is structured for reporting, not for decision-making. Your system has guardrails designed to catch human mistakes, not to enable autonomous action.
This is especially acute in manufacturing because manufacturing systems are older. Most manufacturers run ERPs that are 8, 12, even 15 years old. Those systems were built when real-time decision-making wasn’t even a design consideration. They’re designed for data consistency, transaction control, and human approval chains - all excellent for what they were built for. But agents can’t work in that environment.
The constraint isn’t intelligence. It’s architecture.
What Agents Actually Require
Think about what a simple demand forecasting agent needs to do:
- Access current production schedules, inventory levels, and open orders - continuously
- Query historical demand patterns and seasonal trends
- Analyze supply chain lead times and supplier reliability
- Recommend adjustments to production plans or purchase orders
- Execute those adjustments (or trigger them for approval)
- Monitor outcomes and recalibrate
Now map that to your ERP. Where does it live? Which modules? How does it talk to your WMS, your supplier portal, your customer order system?
Your ERP doesn’t have a “real-time agent interface.” It has reports, APIs bolted on later, integrations duct-taped together, and batch processes.
So you build a workaround. You extract ERP data nightly. You load it into a data warehouse or AI orchestration platform. You let the agent work with that snapshot. You manually push the agent’s recommendations back into the ERP.
That’s not agentic AI. That’s a human acting as a bridge between a smart system and a dumb one.
And you’re paying for it: extra infrastructure, extra delays, extra points of failure.
The Real Constraint: Architecture, Not Intelligence
Here’s the uncomfortable truth: Most manufacturing software wasn’t designed for this. Your ERP predates the era when autonomous decision-making was possible or necessary.
Adding an agent layer on top doesn’t fix the foundation.
You need operational software that’s built from the ground up with autonomous workflows in mind. That means:
- Real-time data accessibility. Agents need live production data, not Tuesday’s export.
- Flexible approval workflows. Some transactions need no approval. Some need soft approval. Your ERP’s model doesn’t accommodate this.
- Autonomous transaction execution. Your system needs to let agents create purchase orders, adjust schedules, or flag exceptions without human gate-keepers for every action.
- Immediate feedback loops. The agent acts, the system records it, the agent sees the outcome - all in real time.
- Clean data for decision-making. Agents need normalized, accessible, trustworthy data. Your ERP’s data often lives in silos, in different formats, with inconsistent definitions.
Your ERP does most of those things - but not in a way agents can use.
Why This Leads to Custom Operational Software
When a manufacturing company gets serious about agents, they hit a choice:
Option 1: Build workarounds. Extract ERP data, build external AI platforms, manually sync results back. Slow. Fragile. Expensive.
Option 2: Overhaul the ERP. Rip and replace. Migrate all historical data. Retrain everyone. Shut down production during cutover. That’s a 12 - 18 month, multimillion-dollar project for a typical manufacturer. And the new ERP still wasn’t designed for agents.
Option 3: Build the software layer agents actually need. Custom operational software that sits alongside (or in front of) your ERP. It ingests ERP data and translates it into agent-friendly structures. It lets agents execute decisions into the ERP through controlled interfaces. It handles the workflows your ERP can’t.
That’s the software Jetpack Labs builds. Not rip-and-replace. Not bolted-on workarounds. Just the operational software that fills the gap between what your ERP does and what agents require.
It’s not free of tradeoffs either. You’re now maintaining another system, and someone on your team needs to own the interface between it and the ERP as both evolve. It’s a smaller, more contained obligation than options 1 or 2 - but it’s not zero.
Picture this manufacturing scenario: You’re running production across three plants. Your demand agent needs to consider current inventory at all three plants, in-flight production runs, current supplier lead times, machine downtime patterns, and quality hold status.
Your ERP has some of this. Some lives in Excel. Some lives in your operations manager’s head.
The agent can’t work with incomplete data and tribal knowledge. So what gets built?
Custom software that pulls real-time production status from your machines, normalizes inventory data across locations, integrates supplier data so lead times are live, flags quality holds automatically, and surfaces downtime patterns. Then it works the other direction: When the agent recommends a production adjustment, it pushes that intelligently into your ERP - creating the right transactions, respecting approval chains where they matter, executing autonomously where they don’t.
That’s the software that makes agentic AI actually work in manufacturing. And it doesn’t come from your ERP vendor.
The Takeaway
Agentic AI is real. The capabilities are proven. Early deployments show genuine ROI.
If you’re considering agentic systems for manufacturing, start by asking: does your operational foundation support autonomous workflows? If the answer is “we’d probably need custom software for that,” that’s not a strike against agentic AI. It’s just the honest answer.
That’s what we do at Jetpack Labs. We build the operational foundation that lets agents actually work in manufacturing, logistics, and industrial operations.
If you’re running agents against an ERP that was built for batch processing and human approval chains, let’s talk about what your system really needs.