Supply chain execution runs on hundreds of rules: promising rules, allocation fences, carrier choices, validation checks... The network, customers, and carriers change faster than any manual review can keep up with, so some rules fall through the cracks. KPIs show that something is drifting, but not where the cause is. And fixing each individual occurrence of a recurring problem guarantees it returns, because the cause is not fixed.
The setup that best prevents these problems from coming back combines AI agents and teams. Agents analyze the full context of every order and trace each pattern to its cause. Teams investigate with them and decide based on the agents' recommendations, which stay within the strategy and arbitration rules. Once the cause is addressed, the problem no longer occurs.
Teams keep fixing the same orders by hand because their first job is to answer the immediate need, so the customer promise is kept. They don’t have time to investigate, and an investigation would cost more than the single case it explains. The order gets fixed, and the cause of the problem stays in place.
For example, a manufacturer or a B2B distributor might see recurring problems such as:
Each one takes minutes to correct, so it gets corrected, every time it comes back.
Problems like these fall between the two ways a supply chain normally handles trouble. Routine, such as an in-stock order promised, allocated, and shipped as planned, is handled by the rules. True disruptions, such as a port closure or a supplier failure, get crisis teams. Kbrw calls what falls between them the missing middle: problems too frequent to be exceptions and too small to justify an investigation.
Some of these problems are non-exceptions: recurring situations handled one case at a time, by hand or as separate exceptions, because nobody sees the link between them. Nobody notices that one common cause would remove many occurrences at once.
Because these cases are small and rarely linked to one another, the usual ways of finding causes miss them. KPIs show the drift but not the cause. Exception management can fix a cause when the cause is obvious, but it treats each case on its own. When the cause isn’t obvious, teams can investigate and fix it, but it takes hours of expert work, so it is rare. Part of that work is following the problem across the order journey, because the cause is often in a different step from the symptom. A promise problem can come from allocation, and a shipping problem from a sourcing rule.
Left in place, these causes erode the perfect order rate. At the median, only 88% of orders are perfect, according to APQC’s Open Standards Benchmarking data from 13,590 organizations. APQC counts an order as perfect when it is delivered on time, complete, damage-free, and with accurate documentation. About one order in eight falls short somewhere.
“A recurring problem never looks expensive, because each occurrence takes ten minutes. Multiply those ten minutes by every order, every week, and you find a full-time job that nobody was ever hired to do.” - Philippe Romano, CEO & Co-Founder, Kbrw
AI agents should take on the investigation that teams can’t afford, because they are fast and exhaustive. They work best with your teams, because part of what explains a recurring problem is recorded in no system. Agents find the pattern and trace it to a cause; teams bring the missing context and decide what changes.
Agents can take this on because, for them, being thorough costs little. An agent is not smarter than your best teams. It can simply afford to be thorough everywhere, all the time. For a person, crossing the ERP, the order management system (OMS), stock and capacity positions, carrier data, and customer agreements takes hours for a single case. An agent can do it for every case, every day.
Agents analyze the full context of every order: rejected, late, corrected by hand, promise missed, demand lost, etc. They surface issues nobody saw and link anomalies nobody had connected, which is how non-exceptions come to light.
Agents also hand teams a short list of qualified signals, not a flood of alerts. Each signal describes a recurring situation, how often it occurs, and the pattern the agents recognized, which gives a first lead toward the cause.
Teams then investigate with the agents, starting from that lead. They challenge the agents' hypotheses and redirect the analysis. Part of what happens in a supply chain is recorded nowhere, so every piece of feedback helps the agents understand the wider context. When a workaround or a local event is behind the problem, teams can bring the agents a cause they could never have guessed. Because they worked on the investigation, they understand the change the agents propose and can judge it with full knowledge.
This collaboration matches where analysts expect automation to stop for now. Gartner predicts that by 2031, 60% of supply chain disruptions will be resolved without human intervention. In the same March 2026 release, Gartner advises restricting full automation to low-risk decisions for now. For higher-stakes decisions, it says AI is "better used to augment human judgment."
Kbrw’s agentic OMS runs this investigation continuously with your teams, starting from what actually happens to orders and customers. It is part of Kbrw’s Agentic Execution Engine.
A recurring problem stops when its cause is addressed, not when the order is fixed. That takes a loop where people and agents collaborate, run continuously. Agents recommend changes within the strategy and arbitration rules, teams decide, and the situation no longer occurs.
Because it runs continuously, the loop extends Lean and Six Sigma rather than replacing them. Those methods work project by project on a few big problems, while the missing middle needs continuous and exhaustive investigation.
Kbrw’s loop has four steps, each with a role for agents and a role for teams:
The four steps of the continuous improvement loop in execution
| Step | What agents do | What teams do | Output |
|---|---|---|---|
| Discover | Surface recurring frictions, non-exceptions included, and run their own analyses to surface issues nobody saw | Start from a KPI or launch their own investigations | A short list of qualified signals |
| Pattern | Correlate occurrences, analyze the context, and put forward hypotheses | Bring what no system records, challenge, and redirect | A documented problem with a shared cause |
| Decide | Recommend solutions within the strategy and arbitration rules | Review, adjust, and decide based on the agents' recommendations, for affected orders and a lasting rule change | An approved change, consistent with the strategy |
| Deploy | Change the rule, remember why, and keep it under watch | Handle the actions that need a person | The situation no longer occurs |
In Deploy, every decision is kept in the agent’s memory with its context: what was seen, what was decided, and why. When a rule later causes a problem, teams know why it exists, so they can change it with confidence instead of adding another rule on top. Today every change is approved by a person. Autonomy extends scope by scope, following business impact and criticality.
Kbrw tracks whether the loop works with time to definitive resolution. Time to definitive resolution measures how long it takes until the cause of a recurring problem is settled and the problem stops coming back, not how long it takes to handle the affected order.
For example, one recurring problem might follow two very different timelines.
The usual cycle:
Starting from what actually happens:
Supply chain execution problems keep coming back while each occurrence is fixed where it shows. Teams absorb them because the customer promise comes first, and investigating one case costs more than the case itself. AI agents change that cost. They analyze the full context of every order, surface what nobody flagged, and trace each pattern to its cause. Teams investigate with them and decide based on their recommendations. The problem stops when the cause is found and addressed.
To see how the loop works across the order journey, with a worked example, download the whitepaper Start from reality. Stay on course.