A demand forecasting pilot usually starts with a dataset that looks clean: three years of orders, receipts and delivery dates pulled from the ERP. Much of that history was never generated by a system. Someone typed it in from a purchase order PDF, a carrier invoice or a scanned delivery note, often at the end of a long shift. The model can't tell which fields were keyed by hand, so it learns from the typos as faithfully as from the real demand signal.

That's one reason supply chain AI struggles to get past the pilot stage. Gartner director Snigdha Dewal said in a 2026 webinar that only 17% of AI pilots launched by supply chain organisations scaled successfully, as reported by CCJ. The order of work matters more than most roadmaps admit. Start where data enters the business. Automating document intake pays for itself in saved hours and avoided errors, and it gives every later model cleaner history to learn from.

Forecasting models inherit every typing error upstream

A forecast is only as accurate as the history it learns from, and that history has its own supply chain. The delivery date on a purchase order becomes a lead-time record. A quantity on a goods receipt becomes a stock movement, while a charge on a carrier invoice becomes a cost per lane. When any of those values is entered manually, the error propagates into every report and model that reads it.

Some mistakes are obvious, like a pallet count with an extra zero, and somebody catches them. The ones that damage forecasting are dull. Think of cases recorded as units for one supplier, a promised date entered instead of the actual arrival, or a fuel surcharge booked against base freight. None of them breaks the ERP, yet each one shifts the pattern a model sees.

Research from GEP and the University of Virginia's Darden School of Business points the same way. Its 2026 survey of 180 senior supply chain executives found that fewer than 1 in 10 organisations had scaled AI pilots into enterprise-wide operations. The organisations that did scale were several times more likely to invest in automated data cleaning and digital audit trails. Operational discipline separated the two groups more clearly than the choice of technology did.

What document errors already cost before any AI arrives

Document errors cost money long before anyone builds a model on top of them. Deep Current surveyed 162 freight forwarders, 3PLs and carriers, and trans.info reported the results in November 2025. Of the logistics executives asked, 57% said document errors had directly caused shipment delays in the past year. Another 36% reported compliance penalties linked to paperwork. The causes were mundane: missing HS codes, mistyped consignee addresses and incomplete fields.

Only 24% of the same respondents had digitised internal document handling end to end, and 61% still relied on email and spreadsheets to work with overseas partners. So most of the cost lands in operations budgets as overtime, rework and credit notes. It rarely shows up in the IT line where AI projects get approved. Worth saying: that split makes document automation easy to underrate in a business case, because the savings belong to a different manager.

Document intake is where automation pays for itself

Automation pays off where work is manual, repetitive and expensive to get wrong, and document intake meets all of those conditions. Purchase orders, advance shipping notices, bills of lading and carrier invoices arrive in hundreds of layouts. The fields a business needs from them, though, barely change from month to month. A team can measure the baseline before writing any code: minutes per document, the share of documents corrected later and invoices disputed each month. That baseline turns the business case into arithmetic, not a promise.

The payback is also quick to read. Extraction accuracy becomes visible within weeks, because every field a person corrects is a labelled example the system can learn from. A forecasting model is different, since its value only shows after a full season of decisions. And the planners benefit anyway, as cleaner receipts and lead times flow into the history their models will train on.

A growing share of the requests we receive at Cortance are for process automation and AI agent development. I take that as a sign that buyers have stopped asking what AI could do in general. They now ask which piece of manual work it should remove first. For most supply chain teams, document intake is a sound answer, provided they're honest about the volume they actually handle.

When forecasting should still come first

Forecasting deserves the first slot when the data problem is already solved. Some companies receive most orders, ship notices and invoices through EDI or supplier portals, so very little is keyed by hand. Others have finished an ERP programme that moved suppliers onto structured feeds. In those cases, a document project would automate a small remainder, and the larger gain sits in planning. A quick sign is the share of inbound documents that still arrive as PDFs, scans or email attachments. When that share is small, document work won't move the numbers much.

Volatility is the other exception. Picture a seasonal retailer, or a spare-parts distributor with lumpy orders. When the main pain is demand swings nobody can explain, poor forecasts may cost more than bad paperwork does. Even then, a short audit of where the forecast's inputs come from costs little. Still, if most of those fields arrive structured, start with the model and don't look back.

The part pilots forget: who keeps extraction accurate after go-live

Extraction accuracy decays after go-live unless someone owns it. Carriers redesign invoice templates and suppliers introduce new surcharge codes, usually without telling anyone. No alarm goes off when that happens. The system keeps producing confident output while accuracy slips, and the first sign is often a reconciliation that no longer balances.

Most of the fix is routine work with a clear owner. Low-confidence fields should go to a named reviewer who corrects them, so people spend their time on judgement instead of retyping whole documents. A monthly sample checked against the source documents turns drift into a number somebody watches. New layouts need an owner and a deadline too, because a template change left in a general support queue can sit there for weeks.

Staffing decides the outcome here more than the model does. Gartner published a survey of 140 senior supply chain leaders in April 2026. In it, 56% of chief supply chain officers called integrating AI with legacy systems and processes a major challenge. Half said they had limited internal expertise to implement and manage AI. A document pipeline sits on exactly that seam, between new models and old ERP fields. Teams without those skills in-house can bring inengineers who build document pipelines with human review for the first build. Later they can add engineering capacity one stage at a time as the next document type comes into scope. The process owner, however, should stay on the payroll.

A one-week test before you pick the first use case

Before approving a forecasting pilot, pick one input the model will rely on, such as supplier lead time. Trace a sample of those records back to the documents they came from. Count how many were keyed by hand and how many needed a correction later. If that share is large, the forecast is mostly learning your data-entry habits, and document intake is the better first project. Where the inputs arrive clean and structured, you've earned the right to start with the model. Either way, the choice rests on evidence from your own records instead of a vendor's roadmap, and it's the cheapest piece of supply chain AI research you'll run this year.