From decision support to autonomous execution
Autonomy in the supply chain is granted one decision type at a time, against criteria agreed in advance, and it can be taken back.
"Should we let the system decide?" has no useful answer for a whole supply chain. It has a good answer for a single decision type: releasing a replenishment order for a low-value packaging item, approving a freight bill that matches the contracted rate, re-sequencing a shift after a short stoppage. Each of these has its own volume, its own cost of a mistake and its own track record. Autonomy should be granted the same way, one decision type at a time, against criteria agreed before the first recommendation goes out.
This article lays out a practical ladder, the evidence needed to move up it, the triggers that move a decision back down, and the decisions that should stay with a named person.
The autonomy ladder
Each rung changes who acts and who checks. The rungs apply to a decision type, never to a whole system. The same organization can have freight-bill matching on rung four and supplier awards on rung two.
| Rung | What the system does | What people do | Evidence needed to reach it |
|---|---|---|---|
| 1. Observe | Reads data and works out what it would have recommended. Users see nothing yet. | Decide as usual. | Read-only connection in place. The decision type and its owner are defined. |
| 2. Recommend | Proposes an action with options, evidence and a deadline. | Accept, edit or reject, giving a reason for every edit or rejection. | Shadow results reviewed. Recommendations match good decisions often enough to be worth reading. |
| 3. Approve to execute | Prepares the transaction and, once a named person approves, writes it through the system of record. | Approve or reject each action. | Stable acceptance rate. Rejections understood. Write path and reversal tested. |
| 4. Execute within limits | Acts alone inside agreed limits on value, quantity, counterparties and timing. Anything outside goes to a person. | Review exceptions and a sample of automatic actions. Own the limits. | Sustained record at rung 3. Outcomes at or better than baseline. Limits signed by the decision owner. |
| 5. Self-directed with audit | Chooses among options within a wider policy and reports after the fact. Every action leaves a full record. | Audit regularly, set policy, handle escalations. | Long record at rung 4 without serious errors. Audit trail proven. Executive sign-off. |
Many decision types will never need rung five, and many will rightly stay at rung two or three. The ladder's job is to make the current rung visible and deliberate.
Granting autonomy per decision type also protects the whole effort. When autonomy is granted to a system as a whole, the first serious mistake tends to end it everywhere. When it is granted per decision type, a mistake in one area demotes that area and leaves the rest running. It keeps the conversation concrete too. The buyer who owns low-value replenishment can judge the evidence for low-value replenishment. Nobody can judge the evidence for "the AI" in general.
Promotion criteria, written in advance
Criteria agreed after the results come in tend to bend toward whatever happened. Write them down per decision type before the system makes its first recommendation. They should cover:
- Volume: the minimum number of decisions observed at the current rung.
- Agreement: the share of recommendations approved without change.
- Outcome: results against the agreed baseline, such as stock-outs, days of cover or cost.
- Errors: no serious error in the review period, with "serious" defined up front.
- Limits: the exact boundaries that will apply at the next rung.
- Sign-off: the decision owner, plus finance for any decision that commits money.
Write limits in business terms the owner can check: value per order, the approved supplier list, maximum quantity relative to normal demand, permitted days and times. A limit expressed only as a model confidence threshold is hard for anyone outside the data team to audit.
Illustrative example. Replenishment orders for C-class packaging items move from approve-to-execute to execute-within-limits after 12 weeks in which at least 95% of recommendations were approved unchanged, stock-outs stayed at or below baseline and no order exceeded $5,000. At the new rung, the limits are orders up to $5,000, approved suppliers only and no change to the delivery address. Everything else still goes to the buyer.
Demotion criteria, just as explicit
Autonomy that cannot be withdrawn is not governed. Agree in advance what moves a decision type back down a rung, and make the move automatic:
- A serious error, such as the wrong supplier, the wrong ship-to address or a quantity off by a factor of ten.
- Outcomes drifting below baseline for an agreed period.
- A change in the data underneath: an ERP upgrade, a master-data migration, a new supplier or a feed that has gone stale.
- Conditions the system has not seen before, such as a product launch, a sudden price shock or a new regulation.
- A jump in the rate at which people reject or edit recommendations.
Re-promotion follows the same criteria as the first promotion. Separately, someone with authority should be able to stop automatic execution for a decision type at once, and that stop should be logged with a reason. Treat demotion as routine. It is the mechanism that makes promotion safe to grant in the first place.
What should stay with a named person
Some decisions should not be automated at any rung, whatever the track record:
- New commitments to outside parties beyond agreed terms: onboarding a supplier, changing contract terms, changing customer prices.
- Anything touching safety, quality release or regulatory compliance: releasing a batch on quality hold, overriding a recall block, shipping restricted goods outside the rules.
- Decisions about people, such as staffing and shift assignments that affect pay.
- Large, hard-to-reverse commitments above a value ceiling set by finance.
- Financial write-offs and postings outside policy.
- Changes to the system's own limits, promotion criteria or permissions.
The system still has a role in these decisions. It can gather the evidence, lay out the options and draft the documents, so the person's time goes into the judgment itself.
What every rung needs
Three things make the ladder work at every level.
A record for every decision. The signal, the options weighed, the choice, the evidence, who approved it or which policy allowed it, what was written where, and the outcome measured later. Without that record you cannot promote, demote or explain anything.
A single named owner for each decision type. The owner sets the limits, signs promotions and answers for results. A committee cannot do that job.
A tested reverse path for each action. How to cancel the purchase order, reverse the transfer or restore the previous sequence. Test it before the decision type reaches rung three.
Finding your first candidate on Monday
List ten decisions your team makes every week. For each one, note four things:
- How many times a week it is made.
- The typical value of one decision.
- How easily it can be reversed.
- How often people change what the current system or spreadsheet suggests.
The best first candidate is frequent, low in value, easy to reverse and rarely changed by people today. Write its promotion and demotion criteria, name its owner, and start it at rung one.
FLX4 is engineered for autonomy earned this way. OpsNest routes actions through an approval inbox before anything executes and includes an autonomy setting with a kill switch. Its agents start at recommend. In TransNest, every write an agent proposes waits for a person's approval before it runs.