Create a free Supply & Demand Chain Executive account to continue reading

Why Data Cleanup Projects Fail Manufacturing Operations

If deliveries keep arriving late, increasing the lead time setting without investigating the cause can normalize the delay instead of addressing it.

Lean Dna Richard Lebovitz Headshot
Digital chain link with binary code and glowing blue and orange lights representing blockchain or cybersecurity technology
Chatchanan AdobeStock_923084100

Manufacturing companies often fix data problems without investigating root causes, which hides real operational issues instead of solving them. AI agents work best when they investigate underlying assumptions, gather evidence, and recommend changes that validated analytics can verify—turning data maintenance into measurable business improvements.

  • Data cleanup alone is insufficient; operating assumptions like lead times, safety stock, and replenishment policies require ongoing maintenance, not one-time fixes.
  • AI agents should investigate before correcting—pulling records, testing competing explanations, and preparing evidence for human review rather than simply updating numbers.
  • Different problems need different solutions: a mis-recorded supplier commitment is a data error, while a safety stock target is a policy choice requiring distinct evidence and analysis.
  • Stretching lead times to cover inspection backlogs normalizes delays and masks root causes, leading to hidden costs in expedites, excess safety stock, and late orders.
  • Pair agentic investigation with validated analytics, set clear boundaries on agent access and actions, start with small pilots, and measure business impact on inventory, shortages, and delivery performance.

Manufacturing executives want AI to reduce excess inventory, improve customer delivery, and use their teams better. But AI can only be as good as the assumptions it works from.

Operational data can be complete and consistent and still be wrong for how the business runs today, such as a three-week lead time for material that now takes six. An outdated assumption shapes how much cash the business commits, whether production has the material it needs, and how much of a buyer’s day goes to chasing problems instead of preventing them.

But simply updating the data to match reality can hide the real problem. If deliveries keep arriving late, increasing the lead time setting without investigating the cause can normalize the delay instead of addressing it.

Data quality is the foundation of advanced analytics. The next step is to expect AI to help maintain that foundation. Manufacturers can use AI agents (software that investigates within limits people set) to examine unreliable assumptions and recommend defensible changes, rather than carrying the entire data maintenance burden themselves.

Data cleanup is not enough

Most data cleanup is a project: a one-time effort to correct item master, policy, supplier, and bill of material data. That matters, but projects end, and the operating assumptions that drive daily decisions keep changing: demand forecasts, supplier lead times, safety stock, and replenishment policies. Those assumptions need ongoing maintenance, and they are not all the same kind of problem. A mis-recorded supplier commitment is a data error. A safety stock target is a policy choice that should reflect demand variability, replenishment performance, and the service objective. Both need review, but they require different evidence.

Investigate before you correct

An alert tells someone to look. An agent should do the looking: pull the relevant records, test competing explanations, and prepare a proposed fix before asking the responsible person to review it. Judge an agent by the investigation it completes, not by how fluently it answers questions.

Buyers carry a supplier’s real lead time in their heads: six weeks while the system says three, compensating with calls and expedites.

Every delivery leaves a trail of dates: order release, supplier acknowledged ship date, actual ship date, dock receipt, and inspection release. An agent can compare those dates across comparable orders to identify where delays accumulate. Three places to investigate are supplier performance, transportation, and receiving inspection. More than one may contribute, and each has a different owner.

If the supplier ships after the agreed date, purchasing investigates capacity and delivery performance. If shipments leave on time but arrive late, logistics examines transit assumptions. If material arrives on time but waits for inspection, receiving and quality own the problem.

Stretching the supplier’s lead time to cover an inspection backlog makes the plan look right while the backlog keeps growing, and the cost shows up later as expedites, extra safety stock, and late orders. The factory may need a temporary planning allowance while it reduces the backlog. Record that allowance separately, with an owner and review date, so today’s constraint does not become tomorrow’s permanent assumption.

Pair agentic investigation with validated analytics

The same approach extends to safety stock and order policies, and it works because each part has a distinct job. The agent investigates: it gathers evidence, identifies related issues, and coordinates review. Validated analytics (tested against past results and approved by your analyst) calculate the optimized target inventory and test the implications of proposed changes. A language model’s explanation cannot substitute for that calculation. A person decides.

Suppose a part repeatedly runs short despite a high safety stock setting. The cause might be an understated lead time, demand variability, or an order interval that limits replenishment frequency. Increasing safety stock without examining those relationships can commit more cash while leaving the cause unresolved. For example, adding 500 units of safety stock to a part that costs $40 ties up $20,000 in cash. Analytics can then analyze the last twelve months of demand and receipts to test whether the higher target would have prevented the last shortage.

Demand needs similar care. When shipments fall below forecast, examine customer orders, backlog, and cancellations. Customers may be ordering less, or shortages may be preventing on-time delivery. Treating constrained shipments as evidence of weaker demand can deepen the next shortage.

Trust requires evidence and accountable decisions

An executive should ask how the system handles an incorrect diagnosis. The answer should be a change that can be traced and reversed. Every proposed change should come with its sources, assumptions, alternatives, and expected effect on inventory and service levels. Reviewers need enough evidence to challenge the recommendation, including what remains uncertain.

Set the boundaries first: which data the agent can access and which actions it can take. Keep the validation rules that already work. Low-risk corrections (a supplier name entered two different ways, say) can run automatically under approved rules. Changes to lead times, safety stock, or replenishment policies need approval based on their potential effect on inventory and service.

Verification has two parts: confirm that the authorized update reached the system, then assess whether operating results support the diagnosis. A successful ERP data update proves only the first.

Start small and measure business impact

Start with one recurring problem at one factory, chosen for its effect on production, excess inventory, or expedite workload. For example, a pilot might cover the ten parts that caused the most expedites at one plant last quarter. Require a clear owner and enough evidence to investigate it. Establish the current effort and recurrence rate before introducing an agent. Agree on the inventory and service objectives together before evaluating a proposed reduction.

Measure investigation time, verified resolution, and recurrence. Track inventory, shortages, expedites, and delivery performance separately, accounting for other business changes.

A decision nobody is making

Every lead time, safety stock setting, and order interval that nobody revisits is a decision the business keeps making without anyone deciding it.

AI can increasingly take on the investigation that keeps those decisions current. The goal is less excess inventory without compromising delivery, and teams that spend their time improving the supply chain instead of reconstructing the same problems.

A data fix is complete when the record changes. An operating improvement is demonstrated when the recurring problem declines and the business can measure the result.

Page 1 of 322
Next Page