Your copilot suggests reorders on its own: when to trust it (and when not to)

What's coming: autonomy grows, wave after wave

D365 Supply Chain Management's 2026 release wave 1 pushes toward agents that act: AI-assisted picking, automatic inventory rebalancing across warehouses, demand forecasts linked to price changes. On paper, less manual work and faster decisions. In practice, it means the question "how much do I let the system do?" stops being theoretical: every company on D365 will face it within the next twelve months.

The two reactions I see inside companies

On projects I meet two extremes. On one side, teams that accept the agent's suggestions wholesale, because the time saved is worth the risk: approval becomes a ritual click, and real control no longer exists. On the other, teams that want to see every single reason behind every reorder, even at the cost of losing nearly all the speed advantage: the agent becomes a proposal generator nobody uses.

Neither position is wrong in itself. What's wrong is choosing it by company culture ("we're cautious" / "we're innovators") instead of analysis: trust in an agent isn't a philosophy, it's a per-process decision.

The framework: three questions for every delegable decision

Before delegating a class of decisions to an agent, three questions with measurable answers:

• How much does a mistake cost? A wrong reorder of €200 worth of fasteners gets absorbed; a wrong reorder of €80k of raw material with a 4-month lead time doesn't. The cost threshold below which the agent decides alone is the first thing to write down — in euros, not principles.
• Is the logic explainable? If the agent can say "reordering because the forecast rose 18% after the supplier's price change", human approval takes ten seconds and means something. If the proposal is a black box, every approval is an act of faith — and acts of faith in supply chain end badly.
• Is the action reversible? A transfer between internal warehouses can be undone; a confirmed order to a non-EU supplier can't. Reversible actions tolerate autonomy; irreversible ones require the checkpoint even when cost and explainability check out.

Practical guardrails in D365

Translated into configuration, the framework becomes four concrete guardrails:

• Value-based autonomy thresholds: below €X the agent executes, above it proposes and waits. The CFO sets X, not IT — and it differs by product category.
• A complete audit trail: every agent action logged with inputs, logic, and who/what approved it. When the wrong reorder lands (it will), "why did it do that?" must have an answer in minutes, not weeks.
• Visible reasoning on every proposal: the planner must see the why next to the what. It's the single factor that moves adoption — explained suggestions get used, bare ones get ignored.
• Periodic boundary review: every quarter, look at the override rate. If planners correct the agent less than 5% of the time on a decision class, the autonomy threshold can rise; if they correct it a third of the time, the problem isn't trust — it's the model.

Where to start

The practical sequence: inventory the decisions your copilot can already propose in D365; classify them with the three questions (cost of error, explainability, reversibility); delegate low-cost reversible decisions first; measure the override rate for a quarter and move the boundary based on data, not feelings.

The end state isn't "trusting AI" or "not trusting it": it's having an explicit, written, periodically reviewed boundary between what the agent decides and what it proposes. Companies that have one turn autonomy into an advantage; the others oscillate between the ritual click and paralysis.