Insight Is Cheap Now. Action Is Scarce.

In conversations with supply chain leaders this year, we keep hearing the same three admissions, usually a few minutes in, once the polished version of the story gives way to the real one.

The first is that the systems aren’t going to converge. Plants arrive through acquisition with their own ERPs attached. Regions standardize on different platforms for good local reasons. Customizations outlive the people who wrote them. A clean, single-vendor stack is not a destination anyone is actually reaching.

The second is that having the data is not the same as having the understanding. The purchase orders, supplier history, and schedule changes are all recorded faithfully. What’s missing is what happens next: turning a signal into a decision, and the decision into action back inside the systems where the work gets done.

The third is that nobody wants to wait for a data consolidation program before AI delivers anything. Those programs take 18 to 36 months, and by the time value shows up, the cost has long since hit the P&L.

Most of the industry now agrees on those three points. So, the question is no longer whether intelligence should sit across the systems a company already runs. It’s how that layer is built, who owns it, and what it does when a disruption crosses more than one function. That’s where the choices start to matter.

Where the data layer lives, and whether you can leave

Any AI worth deploying needs data that has been connected, reconciled, and given business context across ERP, planning, warehouse, and supplier sources. The question is where that work lands and whose it is.

A data foundation built for AI should read from source systems where they already sit, with nothing migrated. It should store data in open table formats any engine can read. And it should deliver tailored data to each consumer, because a forecasting model, an order-promising agent, and a dashboard all need different shapes of data, and forcing them onto one flattened table serves none of them well. Lineage, permissions, and audit should be built into the foundation rather than added per application. And if you ever remove it, your infrastructure should be exactly where it was before. If a data layer can’t pass that test, it’s another migration with better marketing.

Functions don’t lose money. Handoffs do.

It’s tempting to deploy AI the way the org chart is drawn: one agent for procurement, one for production planning, one for customer service. It’s a tidy way to package things. It’s also exactly where supply chain value leaks.

Take a supplier whose deliveries are starting to slip. Catching that three weeks early is useful. But the decision that matters isn’t only whether to expedite or re-source. It’s what the shortage does to the master schedule, which customer orders are now at risk, whether safety stock at another site covers the gap, and which promise dates need to move before a customer calls. Those are four teams’ decisions. In most companies, they get reconciled in a meeting, a spreadsheet, and a few phone calls.

Agents organized by function inherit those same seams. Agents coordinated around decisions, reconciling across forecasting, planning, inventory, and on-time-in-full performance the way a well-run planning team would, can re-plan demand, production, inventory, and supply together from a single signal. The value comes from the coordinated system, not any one agent.

Governance tells the AI what it’s allowed to do. Judgment tells it what you would do.

When the conversation turns to AI that takes action, every CIO asks the same things: What can it touch? What can it change without approval? Can I trace every action back to who authorized it and why? Those are the right questions, and any system that can’t answer them isn’t ready for production.

But permissions and audit only answer “is this action allowed?” They don’t answer “is this what we would do?” Every supply chain runs on rules no system ever captured: which customers get protected when capacity is short, which suppliers get a second chance, when a planner overrides the forecast and why, when to escalate and to whom. That logic lives in people. When they retire, it walks out the door.

What’s missing in most AI architectures is a layer that captures that judgment and keeps learning it. Every approval, rejection, override, and escalation becomes a signal. Routine cases build a reliable baseline quickly. The edge cases, the exceptions and conflicts that need a planner’s call, are where institutional knowledge actually lives, and where the value is. Governance keeps the system inside your limits. Captured judgment makes its recommendations reflect how your senior planners actually decide. And that judgment should belong to the company, never shared across customers.

There’s a quieter benefit, too. Much of what companies call ERP customization debt is really business judgment frozen in code. When that judgment lives in a layer above the ERP, the next modernization doesn’t require rediscovering it.

The loop doesn’t close at action. It closes at outcome.

Signal to action is the right frame, but it needs a step on each end.

Before the action, the planner should see the full picture: the recommended action, the data and constraint that triggered it, the alternatives considered, and the expected impact. Approval with the why attached is what lets people decide faster instead of second-guessing.

After the action, the approved decision should write back into the ERP, warehouse, or planning system, and the outcome should be measured against what was projected. That result feeds back into the system’s judgment, which is what makes it sharper next cycle. Without that step, you have a faster workflow. With it, you have one that compounds.

It’s also how teams move decisions to automatic responsibly. Not by granting broad authority up front, but by letting projected outcomes prove out cycle after cycle, until the team chooses which calls can run on their own. Trust gets earned one decision type at a time, and the team holds the dial.

Questions worth asking

If you’re evaluating anything in this space, including what your current vendors are adding, these are the questions I’d bring:

Is the data foundation portable, built on open formats, and removable without leaving a mess?
Are the agents coordinated across decisions, or packaged by function?
Where does your company’s decision logic live, who owns it, and does it improve with every approval and override?
Does the loop end at the write-back, or at a measured outcome?
Who decides what runs without approval: vendor defaults, or your team based on results?

And start with the decision that hurts most. One industrial manufacturer took master production scheduling from 24 spreadsheets and 18 hours to 15 minutes, not by replacing a system, but by coordinating the ones already in place. They started with that one decision and expanded from there.

Insight is cheap now. Action is scarce. The companies that close that gap will be the ones that coordinate across functions, keep their own judgment, and learn from every outcome, on top of whatever systems they already run.

A2go is a supply chain decision intelligence company that builds coordinated AI agents above the systems companies already run. To map your highest-pain decision and a path to first impact, visit A2go.ai