ERP and Autonomous Workflows: What's Real

Every ERP conversation in 2025 includes this phrase: "autonomous workflows."

The vision is seductive. A purchase requisition comes in, the system automatically finds the best vendor, creates the PO, schedules delivery, flags it in your calendar. No human needed. Just rules and AI.

We can't do that yet. And I don't think we can do most of it the way people imagine.

Here's what autonomous workflows actually look like in production right now, and where the limits are.

What Counts as Autonomous

First, let me define the term. An autonomous workflow is one where the ERP system takes an action without a human explicitly approving it first. Not assisted. Not suggested. Actually executed.

Most "autonomous" workflows in 2025 are: fully autonomous on happy path, but with extensive fallback conditions that kick a human in if anything looks wrong.

Here's an example from a manufacturing client. A reorder point is hit for a high-volume SKU. The system automatically creates a PO with a specific vendor. That vendor has been reliable. The pricing is locked. The item is standard. The quantity is predictable. Autonomous? Yes. It works about 85% of the time.

But 15% of the time, something is different. The vendor is out of stock (the data wasn't updated). Lead time changed. New compliance requirement. Shipping cost skyrocketed. And for every one of those cases, a human has to step in, override the autonomous decision, and make a new one.

That's not fully autonomous. That's "automated until it's not, then escalate."

Is it valuable? Absolutely. Humans are freed up from routine decisions. But it's not the sci-fi version of "ERP that just works without people."

The Constraints That Matter

Why can't we do full autonomy? Because real workflows have constraints that are impossible to model perfectly.

Business constraints: "We can spend this much this quarter." "We need to keep this supplier happy." "We can't have more than 30 days inventory of this item." Most of these constraints are in someone's head or in an email thread, not in your ERP system. An autonomous workflow either ignores them or halts when it encounters them.

Data constraints: Your ERP data is unreliable. Vendor master data is stale. Lead times are wrong. Prices haven't been updated. Inventory counts don't match reality. You're asking an AI system to make autonomous decisions on data that's not trustworthy. Most intelligent systems will refuse to do it. Some will do it and fail gracefully.

Regulatory constraints: Industries differ wildly. Pharma has different rules than retail. A purchase needs approval if the vendor is in a sanctioned country. If the item is regulated. If the amount exceeds a threshold. If it's the first order from that vendor. Encoding all of that logic correctly is the difference between a working system and a lawsuit.

Process constraints: Workflows in real companies have exceptions. When the exception happens, you need human judgment. "Usually buy from this vendor. Unless their lead time is more than 10 days, in which case use this backup vendor. Unless it's end-of-quarter, in which case use the internal stock if available." That's only three rules. Most real processes have dozens.

Most teams try to hard-code all these rules into their ERP system. That's where everything breaks.

The Pattern That Works

The winning approach in 2025 is: autonomous for the high-confidence cases, escalation for everything else.

You define:

• Which decisions the system can make autonomously (narrow scope)
• What data validation must pass (data quality gates)
• What exceptions trigger human review (escalation rules)
• How escalation happens (do you need one approval, two, an SLA, etc.)

Then you run it and measure. How much actually automated? What percentage escalated? Of the escalations, which ones did the human approve as-is vs. override? This tells you whether your autonomy rules are too loose, too tight, or just right.

The best implementations I've seen operate at 60-75% autonomous. The rest escalate. But those escalations are fast. They hit someone's queue, get approved or modified in minutes, and move forward.

Compare that to "everything manual" which is 0% autonomous but everyone agrees it's approved. The 60-75% autonomous system is faster overall because you've saved the human from deciding on the routine 60-75% while still protecting them on the complex cases.

Where Autonomous Workflows Actually Ship

Reordering high-volume SKUs with stable suppliers? Yes, this works.

Automatically matching invoices to POs when they match exactly? Yes.

Aging receivables aging and flagging overdue accounts? Yes.

Automatically closing fiscal periods and posting journal entries? Yes, but with more careful validation.

Automatically making customer credit decisions? Not really. Too much liability. Most systems require a human sign-off on new credit or credit increases.

Automatically adjusting prices in response to market changes? Not autonomously. Suggested, yes. Autonomous, no. The liability is too high.

The pattern: the workflows that are actually autonomous are the ones with binary or near-binary decisions where wrong doesn't mean expensive. The ones where complex judgment is required stay escalated.

The Future State

I think autonomous workflows will expand in 2025 and 2026, but not in the way people imagine.

What will change is the quality of escalation. Instead of humans making decisions from scratch, they'll be confirming or modifying AI-assisted recommendations. "The system recommends vendor B. Here's why. Do you approve?" That's still not autonomous, but it's faster than "figure it out from nothing."

And the models will get better at understanding constraints. Not through hard-coding rules, but through learning from historical decisions. "Here's what humans decided in similar situations. Here's the pattern." That informs the system without requiring explicit coding.

But the fundamental truth won't change: ERP workflows have to be correct. A manufacturing company can't have a system that's autonomous 99% of the time and wrong 1% of the time because that 1% might be a $100,000 mistake or a compliance violation.

So the autonomy you're going to see is: very confident in narrow cases, explicit escalation otherwise.

That's not as exciting as the pitch. But it's what actually works, and that's where we're headed.