The model isn't slow. The sign-off queue is. The next gains in supply chain AI will come from redesigning who is allowed to decide what — not from a more accurate forecast.

Approval latency is the delay between an AI system recommending a supply chain action and a human authorizing it to execute — the time a decision spends sitting in a sign-off queue. For two years the industry has framed the bottleneck in supply chain AI as a technology problem: not enough data, not enough accuracy, not a mature enough model. That framing is now mostly wrong.

The constraint has moved. Forecasting models are already good enough to drive real decisions; what's slow is the human approval step wrapped around them. A forecast that is 90% accurate but takes 72 hours to clear procurement, logistics, and production has already lost most of its value in a promotional or seasonal window. The next step-change in supply chain AI will not come from a better algorithm. It comes from shortening — and in defined cases removing — the approval queue.

The accuracy story isn't wrong. It's just finished.

The conventional wisdom earned its place. Better models genuinely moved the needle, and for a while accuracy was the binding constraint.

McKinsey's research on autonomous supply-chain planning found that consumer-goods companies adopting real-time, integrated planning saw revenue rise up to 4%, inventory fall up to 20%, and supply-chain costs drop up to 10% (McKinsey). Adoption is following the capability: Gartner expects 50% of cross-functional SCM solutions to use intelligent agents to autonomously execute decisions by 2030, up from roughly 5% of enterprises using such features in 2025 (Gartner).

So the prediction problem is largely solved for the decisions that matter most. The question stopped being "is the forecast good enough to act on?" and became "how fast can we act on it?"

The bottleneck moved from prediction to permission.

The thing slowing supply chain AI down today is not the model — it's the wait for a human to say yes.

McKinsey describes one branded food-and-beverage company that needed more than five days to produce a demand plan and more than two days to produce a dispatch plan (McKinsey). The model can generate the recommendation in seconds; the latency lives in the hand-offs and approvals between the recommendation and the action. It is also why Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 — not because the models fail, but because of escalating costs, unclear business value, and inadequate risk controls (Gartner). Those are governance problems, not accuracy problems.

In our work with enterprise CPG operators, the same pattern repeats: a recommendation is technically ready in minutes and operationally stuck for days, because no one has defined what the system is allowed to do on its own.

"The model was never the bottleneck. The sign-off queue was."

When the approval step is the constraint, tuning the model for another two points of accuracy changes nothing. You are optimizing the fast part of the system.

Stop tuning the model. Redesign the decision rights.

The fix is organizational, not algorithmic: decide in advance which decisions an agent can execute autonomously, which need a human to confirm, and which stay fully human — then compress the approval path for the first two.

This is where the industry is actually heading. Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from 0% in 2024, and predicts 60% of supply chain disruptions will be resolved without human intervention by 2031 (Gartner). BCG estimates agentic systems already accounted for 17% of total AI value in 2025, rising to 29% by 2028. None of that materializes if every agent action still queues behind a manual sign-off.

The practical move is to treat this as a sequencing problem before a software problem. The first question is not what tool — it is what decision needs to happen, in what window, with what authority. Low-risk, high-frequency, easily reversible decisions — a routine reorder within a set band, a reroute under a cost ceiling — are where autonomous execution pays off first. High-stakes, low-frequency, hard-to-reverse calls stay human. The point of human-in-the-loop is to put judgment where it changes the outcome, not to make a planner rubber-stamp a queue.

Heizen is an AI-native software delivery company that builds supply chain systems for enterprise CPG and manufacturing companies, and the recurring lesson is that authority design — not model quality — decides whether an agent creates value.

The takeaway

If your supply chain AI is not delivering, the instinct is to question the model. Check the queue first. The decision was probably ready hours ago; it was just waiting for permission. Before you let an agent act on its own, though, the harder question is what it needs to earn that authority.