Alison Crawford
Last week we laid out the field guide: agentic AI is not RPA, and it is not a chatbot. The difference is that an agentic workflow detects an issue, reasons about it against your business context, and executes the fix inside your systems of record. It acts.
Which raises the question every serious buyer asks next, usually about ten minutes into a demo. If it acts on its own, what happens when it is wrong? Who is watching? And the honest version of that question, the one people are too polite to say out loud: I am not handing my allocation decisions to a black box.
Good. You shouldn't. And the vendors telling you that "fully autonomous" is the goal, that the mark of a mature system is one where the humans have been removed, are selling you a story that does not survive contact with a real supply chain. Autonomy is not a switch you flip to "on." It is a spectrum. Where a given workflow sits on that spectrum is a decision you make, workflow by workflow, and the ability to place it precisely is the product.
What the spectrum actually looks like
Autonomy is not binary, and treating it as binary is how you end up with a system nobody trusts enough to turn on. A useful way to think about it is four settings, and the same workflow can move between them over time.
- Recommend. The workflow detects the issue and proposes the action. A person reviews and executes. Nothing happens without a human hand on it.
- Approve. The workflow prepares the action and holds. A person approves, and then it executes. The difference from recommend is that the work is already done, the human is a gate, not a laborer.
- Act with notification. The workflow executes inside a defined boundary and tells a person what it did. The human can reverse it. This is the setting most mature workflows live in.
- Act autonomously. Inside tight, pre-agreed limits, the workflow executes and logs. A person reviews in aggregate, not case by case. Reserved for high-volume, low-variance decisions where the cost of waiting exceeds the cost of a rare miss.
No workflow should start at the bottom of that list. The point is not to reach "act autonomously" as fast as possible. The point is to place each workflow where the math actually favors it, and to be able to move it as trust is earned.
Why "fully autonomous" is the wrong goal
The pitch for removing the human sounds like efficiency. In practice it is a liability, for three concrete reasons.
First, supply chain execution runs on exceptions. The routine 80% of a workflow is easy to automate and was probably automated years ago. The value is in the 20% that is unusual: the short-ship that might be a data error or might be a real shortage, the chargeback that is contestable, the allocation call between two accounts that both matter. Those are precisely the decisions where judgment, and accountability, belong with a person who can be asked why.
Second, trust is earned per workflow, not granted to a vendor. A COO does not decide once that "the AI is trustworthy." They watch one workflow handle a hundred exceptions correctly, and then they widen its boundary. A system that cannot show its work, or cannot be dialed back one notch when it gets something wrong, never earns the second workflow. Removing the human removes the mechanism by which trust is built.
Third, someone is accountable to a customer, an auditor, or a board, and that someone is not the software. When a Fortune 500 retailer asks why an order was expedited or an allocation was changed, "the model decided" is not an answer anyone can stand behind. The human in the loop is not friction. They are the accountable party, and the system exists to make their judgment faster and better informed, not to replace it.
Human-in-the-loop is the feature, not the training wheels
Here is the reframe that matters. Keeping a person in the loop is often described as a phase, the cautious early setting you graduate out of once the system proves itself. That framing is backwards for most of what a supply chain does.
For the high-stakes, high-variance decisions that make up the hard part of execution, human-in-the-loop is not a limitation you tolerate on the way to something better. It is the design. The system's job is to compress the time around that decision: surface the exception the moment it appears, pull the unstructured context your system of record cannot read, lay out the options with the tradeoffs already calculated, and stage the action so approval is one step instead of an afternoon. The person still decides. They just decide in minutes, with better information, instead of finding out three days later that a decision needed making at all.
That is what a system of action does. It does not take the human out of execution. It takes the delay, the blind spots, and the manual assembly out of the human's way, so the decisions that need a person get one, faster, and the routine work that doesn't runs on its own.
The right question is never "how much can we automate." It is "where does this specific workflow belong on the spectrum today, and what would have to be true to move it." A vendor who answers that honestly, workflow by workflow, is showing you how execution actually works. A vendor who answers "all of it, autonomously" is showing you they have not run one.
Frequently asked questions
Does human-in-the-loop mean the system is slower?No. The point of keeping a person on the high-stakes decisions is to make those decisions faster, not to add a review step. The workflow does the detection, context-gathering, and option-framing in the background, so the human's part shrinks from hours of assembly to a single approval.
Can a workflow's autonomy level change over time?Yes, and it should. A workflow typically starts at recommend or approve, and its boundary widens as it demonstrably handles exceptions correctly. The autonomy setting is a dial you turn as trust is earned, not a fixed property of the system.
Isn't "fully autonomous" the more advanced product?Only for a narrow class of high-volume, low-variance decisions. For the exception-heavy work that makes up most of supply chain execution, a system that can place each workflow precisely on the spectrum, and keep a human accountable where it matters, is the more advanced product, not the less.
Who is accountable when an agentic workflow acts?A person, always. The workflow executes inside boundaries a human sets, logs what it does, and can be reviewed or reversed. Accountability stays with the operator, which is exactly why the human-in-the-loop design matters for enterprises answerable to customers, auditors, and boards.
Alison Crawford is Director of Product Marketing at Emerix, the system of action for supply chain execution. She has spent her career in supply chain software, with prior roles at RELEX Solutions, Aera Technology, o9 Solutions, and Kinaxis, and was recognized in the 2023 Women in Supply Chain Awards. She writes about the distance between what supply chain software promises and what it actually executes.
Connect with Alison on LinkedIn.