Trust before automation: how to introduce AI in planning, step by step

At Nitator Automotive, a supplier of sheet metal components to heavy vehicle manufacturers, around 50 delivery schedules arrive every day. Together they hold around 52,000 order lines. No planner can analyze that by hand, and today’s analysis is outdated by tomorrow. AI can help. But in an industry where a wrong call is expensive, the real question isn’t what the technology can do. It’s when you can trust it.

Forecasts you can’t quite rely on

Delivery schedules from OEMs are the starting point for almost every planning decision a supplier makes. Production, material purchasing, packaging, staffing. But anyone who works with them knows the plans rarely hold all the way to the actual order.

Some deviations are random, but most are not. In some supply chain lines, customer forecasts run systematically and periodically high, sometimes by as much as 20 percent. Ahead of the summer holidays, many customers add an extra margin out of fear that deliveries won’t keep up in August. Twelve-month forecasts are routinely adjusted downward as the delivery date gets closer.

The consequences are familiar: safety stock that is higher than it needs to be, capital tied up in inventory, and late changes that cause scrambling on the shop floor. And of course, a negative impact on the final financial result.

The patterns that cause the deviations are there in the data, but with tens of thousands of order lines a day, nobody has the time to find them by hand.

Why you can’t just switch AI on (and why it’s worth the wait)

It would be easy to call this a technology problem. It isn’t. The tools already exist. What’s missing in the world of supply chain management is the confidence to let refined tools affect real decisions.

That caution is rational. Automotive is, as Charles Gilbert, Logistics Manager at Nitator Automotive, describes it, a low risk, low reward industry. Customers order components that look more or less the same for decades, and the payoff from a change can take a long time to show. When feedback is slow and mistakes are costly, you hold back until the value is proven.

On SCE Talks, Charles described where he wants to get to:

“My hope is that we’ll trust the forecasts from Forecast AI enough to feed them into our ERP system as the truth, so our ERP starts working from those parameters.”

There is a human side to it too. We tend to forgive a planner’s mistake more readily than a system’s, even when the system is statistically better. So, the answer is not to push AI harder into the process. It is to let it earn its place.

Build a better forecast before it reaches your ERP

Forecast AI is built around the idea that ownership, control and incremental testing is the key to make AI enhance your forecasts. It works in the flow where PipeChain already receives your customers’ delivery schedules and converts them for your ERP system. Before a schedule is imported, the AI model takes a look at it.

For each item, the model goes through every delivery schedule received and compares what was forecast with what was actually ordered. Over time, it learns the ordering patterns and builds a forecast of its own. That makes it possible to compare the OEM’s forecast with the model’s, and see which one has historically been more accurate.

Forecast AI then highlights the items where the model has the better track record. Here’s the important part: you decide what happens next: which period of call-offs to adjust, and by how much, based on your own lead times. Only call-offs within that window are adjusted before the schedule is read into your ERP.

The model is trained, but you set the parameters. Ownership, and the level of tolerance, stays with you.

Three steps from decision support to automation

Trust isn’t built for the whole operation at once. It is built per item and per time horizon. That is why we recommend introducing AI in planning in three steps.

Step 1: Start far out, with decision support 4–12 months ahead

Forecasts four to twelve months out shape capacity and inventory planning, but not next week’s shipments, so the AI can inform decisions without driving them.

This is where the learning starts. Compare the model’s forecast with actual outcomes month by month to see where it’s right, where it’s wrong, and why.

Step 2: Move closer, with recommendations 1–4 months ahead

Once the model has proven itself on certain items, it can move closer to delivery, where recommendations directly affect purchasing, packaging and production planning.

The key is to be selective: if a customer’s bracket forecast has run 15 percent high every June for three years and the model has learned to correct for it, use the model. For a sporadic order with no clear pattern, keep the OEM’s forecast. The closer you get to delivery, the more the evidence has to support it.

Step 3: Automate gradually, where trust is proven

The final step is to let adjusted forecasts flow into the ERP system without manual handling, for the items and periods where trust has been earned. Not everything, and not all at once. Automation grows as the track record does, and so does the trust within the organization.

Control stays in place throughout. Settings are made per item and can be tightened or loosened as conditions change. AI doesn’t take over the decision. It takes over the routine, so planners can spend their time where judgment really matters.

Where we are today

Forecast AI is available in beta and is being developed together with suppliers in the automotive industry, Nitator Automotive among them. What we have learned so far confirms the approach: value comes from starting small, measuring honestly and letting the results lead the way.

It is also worth being clear about what Forecast AI is, and isn’t. It doesn’t start out as the truth about future demand. It earns that role, item by item. It is a way to reduce risk, by giving you a broader and more consistent basis for the decisions you already make.

Start where the risk is lowest

AI in planning doesn’t have to begin with a big decision. It can begin with a simple question: which of your forecasts do you trust the least, and how far ahead do they reach?

That is a good place to start. Far from the shop floor, with time to learn, and with full control over what happens next.

We are looking for more suppliers who want to test Forecast AI in beta. If you’d like to talk about what your forecasts look like today, and where AI could make a difference, let’s start with a short conversation.

Book a meeting: 30 minutes about your forecasts