AI Scheduling vs ERP: Why Job Shops Still Struggle

Why APS misses in Job Shops

Late orders persist. Constant re-scheduling is the norm despite decades of software.

In the 1960s, manufacturers embraced a new idea called Material Requirements Planning (MRP). Computers would tell us what to buy and when. It felt inevitable that order would arise from chaos.

Half a century later, most custom job shops are still waiting for that reduction in chaos. To understand why, we need to look at how today’s systems evolved and what foundational logic never changed.

 

AI Job Shop Scheduling vs ERP: what changed and what didn’t

A short history that sets the stage.

• MRP was built to explode a bill of materials from a master schedule – requirements planning. It assumed infinite capacity and fixed lead times. It told you what to buy and when. When reality changed, dates slipped, and buffers grew. And quickly … “order instability, or nervousness, is frequently cited as an obstacle to… MRP.”
• In the ’80s we got MRP II with more scope, financial ties, some what-ifs.  The foundational requirements planning engine didn’t change and implementation took much longer. Bolt-on capacity planning, shop floor data capture, standard costing, variance tracking was added. Cost accounting became the control system with focus on optimizing local efficiencies (and not velocity or productivity or due date performance).
• In the ’90s Enterprise Resource Planning (ERP) integrated everything. Cost accounting logic burrowed deeper into operations. ERPs integrated finance, purchasing, inventory, CRM, HR, etc. What didn’t change? The core planning engine still explodes requirements with fixed lead times and plans as if capacity were infinite. Full ERP installs often took two years.
• In the 2000s, bolt-on job shop scheduling software with finite scheduling try to band-aid the plan with more detail: Advanced Planning and Scheduling (APS), visual Gantts, optimizers. These tools read ERP data. They promised to fix scheduling. But the rules stayed the same. Despite fancier visuals and added complexity, the underlying system logic—requirements planning and cost accounting—still drives. The result: chaos, poor on-time delivery, and long lead times remain unchanged.
• In the 2010s ERPs moved to the cloud – SaaS model. Faster upgrades, slightly faster installs. The goal was detailed/finite scheduling (a lot more detail), dispatch lists, fancier Gantt charts, “visual boards,” drag and drop, pretty dashboards. What stayed the same? The same assumptions MRP started with. Cost-based logic is still ingrained. You’re optimizing a bad plan faster.

And that brings us to the current decade.

The labels changed. The rules did not.

At‑a‑glance: MRP → ERP → APS → “AI scheduling”

System	Core logic	Key assumptions	What changed	What didn’t	Evidence
MRP (1960s–)	Requirements planning from MPS/BOM	Infinite capacity, fixed lead times	Computerized purchase/work orders	Ignores real capacity and variability	IFM MRP overview
MRP II (1980s)	Adds finance, simulation, S&OP hooks	Same planning core	Wider scope, more data capture	High Nervousness and chaos. Low velocity, productivity & DDP.	Management Science
ERP (1990s–)	Integrates functions around MRP core	Same planning core	Enterprise integration	Cost accounting deeper in ops	CIO history
APS / Finite scheduling (2000s–)	Capacity‑constrained sequencing layer	Data is accurate and stable	Gantt/optimizers on ERP data	Base plan still unstable	APS review
Cloud ERP (2010s–)	SaaS delivery and faster upgrades	Same planning core	Lower infra friction	Faster, not different logic	CIO history
“AI scheduling” (2020s–)	Predict/optimize within existing rules	Data + old objectives	More detail, faster re‑plans	Rules and objectives unchanged	PP/DS; Oracle PS

The 2020s:  AI Job Shop Scheduling

In the 2020s the marketing has changed. Every vendor now claims AI job shop scheduling, machine learning, digital twin, predictive analytics, even “TOC-based.”

We have tons more detail and can analyze more, faster. What hasn’t changed? The focus is still cost accounting efficiency and requirements planning is the same.

AI sits on a poor, outdated foundation.

• Finite scheduling is mainstream, but it’s an add-on to the same planning core. This is true for embedded ERP scheduling modules and bolt-ons. Job shops face constant change and moving constraints. Requirements planning can’t solve that.
• Cost accounting is still entrenched by regulation and GAAP but also by inertia.
• Off-the-shelf TOC scheduling or Drum Buffer Rope software does not work in custom job shops, where the constraint shifts based on mix. Even so, many vendors still claim TOC, but print dispatch lists and try to predict each job at each machine.
• Schedule nervousness and instability remain an active problem. Most teams try to fix schedules by adding detail. They also update more often. AI makes that easier. But recent peer-reviewed work confirms “frequent plan updates can harm logistics performance”.
• AI models are trained on patterns and popular responses, most of which are rooted in conventional operations logic. When AI answers a question, it gives the most common, accepted response.  Ask “How should I schedule a job shop?” and you will get utilization targets, cost rates, finite Gantt charts, and dispatch lists. Ask about TOC and you still get an off‑the‑shelf answer that misses how constraints move in custom shops. Off-the-shelf AI will not help you to change your rules. AI does not spread new rules.  It cements old ones.

If these systems had fundamentally worked, custom manufacturing productivity should have soared. It hasn’t.

 

Evidence: productivity, costs, and ERP ROI

If these systems worked as promised, we’d see clear improvements in performance. But the data tells a different story.

• Manufacturing productivity slowdown: “From 2004–2016, manufacturing multifactor productivity declined by 0.3%/yr.”
• Recent performance: “Labor productivity decreased in 52 of 86 manufacturing industries in 2024.” The 2024 “winners” often won by shrinking hours faster than output. That is not a systems breakthrough. It’s headcount or hours control.  Job shop–intensive subsectors mostly lost ground despite decades of ERP/APS.
• ERP ROI disappointment: Only 40% of ERP projects finish at/under budget. “Almost a quarter… dissatisfied.” Many upgrades “simply automate existing inefficiencies.”
• IT project failures overall: Standish CHAOS report cites ~31% failure rate
• New evidence from MIT shows most firms get no ROI when they bolt AI onto legacy workflows; the outliers redesign processes and retrain teams.

–> Billions spent. Productivity stagnant. ROI elusive.

The current software pitch is simple: add AI and your schedule will self-heal.

The reality is simpler: if your rules foundation is requirements planning + cost accounting, you’ve automated the wrong rules.

1. Systems produce the results according to their core logic/rules.
2. MRP/ERP and Scheduling Software’s core logic presumes it can predict when every job will be at every machine or step in the process. It can’t. There’s too much change.
3. Cost accounting pushes you to improve local efficiencies and local use. This hurts flow based scheduling, due dates, and profit.
4. AI that reads and optimizes those same rules can only accelerate their consequences.

Therefore, AI sprinkled on top of old ERP logic and scheduling systems won’t create reliable velocity in a custom job shop. It will create faster oscillations.

Goldratt said technology is necessary but not sufficient; improvement needs new rules. In job shops those rules are about velocity over local efficiency, the right release order, and the right measures.

 

What works instead: velocity-first control

 

Results change when the rules change.

That’s exactly what we operationalize in Velocity Scheduling System and is embedded in the VSS module (iVSS), the first module of our full ERP replacement.  Simple put, ERP is for reporting; iVSS is for results.

Run a custom shop? Want more velocity, shorter lead times and fewer expedites? Want more on-time shipments?

Don’t bolt AI onto 1980s rules. Change the rules first. Then add intelligence.