Walk into almost any mid-market manufacturing or distribution company today and you will find at least one AI pilot running somewhere. A demand forecasting model. A conversational interface for ERP queries. An anomaly detection alert. They have the tools. They have the data. The pilots look promising.
Then ask how many of those pilots are running in production, shaping how decisions actually get made. The answer is almost always: not many. Sometimes none.
Yet cost-efficiency remains the top priority for the third consecutive year — the AI is not translating into savings
The disconnect is not a technology problem. The technology works. The forecasting models are reasonably accurate. The conversational interfaces do answer questions from live data. The anomaly detection does flag the right signals.
The problem is that the technology lands in an organisation that has no formal model for what to do with its output. No defined decision owners. No structured handoffs between the AI recommendation and the human who acts on it. No audit trail for what was decided, by whom, and why. The AI becomes one more input into the weekly coordination call — and the weekly coordination call stays.
What stops AI from scaling
After working with mid-market manufacturers and distributors across multiple AI deployments, we have identified three structural gaps that determine whether a supply chain AI investment stays in the pilot or reaches production. They have nothing to do with model accuracy or data quality. They are all organisational design problems.
No decision ownership
The AI surfaces a signal — a stockout forming, a supplier overdue, a production order at risk. The alert goes to a shared channel. Everyone sees it. Nobody is assigned to act on it within a defined timeframe. Two days later, the signal is still there. The stockout happened.
Decision ownership is the single most important structural element of a scalable AI deployment. Every AI output needs a named owner: the person responsible for evaluating it, acting on it, or escalating it within a specified window. Without this, the AI generates insights that disappear into the noise of operational life.
Assign decision ownership by category before any AI goes live. Stockout risk: owned by VP Operations, response SLA 4 hours. Budget breach: owned by CFO, response SLA same day. The AI is not the decision-maker — it is the trigger. A named person owns the decision.
No approval governance
AI can draft a purchase order, a reschedule, a transfer order. But who approves it? At what value threshold? Through what channel? With what documentation for the audit trail?
Most pilot deployments leave this undefined. The AI recommends. A person approves somewhere, somehow — usually via email, or a verbal yes in a meeting, with no structured record. That approach collapses at scale and creates compliance exposure in regulated industries.
Define approval tiers before deployment: values under £10K approved by Ops Manager via Teams card, £10K–£50K requires VP Operations, above £50K requires CFO. Every approval is logged with timestamp, context, and approver identity. The governance model is the infrastructure — the AI operates within it.
No cross-functional coordination layer
Most operational decisions require input from more than one function. A stockout resolution involves purchasing, finance, and operations. A production at-risk scenario involves planning, procurement, and the plant floor. AI can surface the signal, but if there is no structured workflow for cross-functional coordination, the decision still takes two days — now with an AI notification attached to the email chain.
Map the cross-functional decision workflows before deployment. For each signal type: who needs to be notified, in what sequence, with what information, and who makes the final call. The AI drives the signal into the workflow, not into an inbox.
The execution layer that makes AI scale
These three gaps share a root cause: treating AI deployment as a technology project rather than an operating model redesign. The forecasting model is the easy part. The hard part is changing how your organisation makes decisions — and that requires the same discipline you would apply to any operational transformation.
The organisations that get supply chain AI right do not just deploy a model. They define the decisions it needs to support. They assign ownership for those decisions. They build the governance infrastructure — approval tiers, audit trails, cross-functional coordination flows — before the AI goes live. The technology is the last step, not the first.
Hackett Group, 2026 Supply Chain Key Issues Study
76% report 25%+ operational improvement when the operating model is right. That is a significant outcome — and it is available to any mid-market manufacturer that treats AI deployment as an operating model project rather than a software installation.
What the operating model looks like in practice
An organisation that has built this operating model handles a stockout signal differently from one that hasn't. The difference is not the AI — it is the governance layer around it.
Without a decision operating model — a stockout alert fires at 2 PM Tuesday. It routes to a shared ops channel. The buyer responsible for that category is in meetings. The alert sits. It comes up in the Friday S&OP call, gets debated, and a decision is made Monday. The PO is placed Tuesday. Five business days after the AI saw the risk, the response begins. By that point, emergency freight is already on the table.
With a decision operating model — the same alert fires at 2 PM. It routes to the named owner (VP Operations) with a 4-hour SLA. The owner reviews the costed recommendation in Microsoft Teams — exposure value, draft PO, supplier lead time — and approves with a single click at 4:30 PM. The approval is logged, the PO posts to Business Central, the supplier is notified before 5 PM. Total time: 2.5 hours. The same decision. A fundamentally different operational outcome.
What to measure
The operating model is measurable. These are the four KPIs that tell you whether your decision governance is working:
Decision cycle time — from signal detected to approved action. Target: under 4 hours. Most mid-market manufacturers running without a governance model average 2–3 days.
OTIF rate — On-Time In-Full delivery percentage. A 1% improvement on a $50M revenue base is $500K in penalty avoidance and customer retention. Decision latency is one of the top three controllable drivers.
Emergency freight spend as % of COGS — the most visible financial signal of slow decisions. Each emergency freight event is a decision that arrived too late. Target: below 1.5%.
Decision ownership coverage — what percentage of your AI signal types have a named owner with a defined response SLA. If this is below 80%, your governance model has gaps that your AI cannot compensate for.
OpsGrid — in live beta — is IntelliconnectQ's decision infrastructure for Dynamics 365 Business Central, built around these measures. Every signal has an assigned owner. Every approval routes through defined tiers. Every decision is logged with timestamp, context, and approver. Nothing writes to Business Central without explicit human approval.
For the underlying framework — the distinction between decision intelligence (surfacing signals) and decision infrastructure (governing what happens next) — see Decision Infrastructure vs. Decision Intelligence. The operating model described in this article is the output layer that makes intelligence operational.
If you recognise the gap described here — AI pilots that produce insights without producing decisions — the question is not what AI to buy next. It is: do you have the governance infrastructure in place to act on what it tells you?