Automating inventory: the supply chain data nobody trusts

Inventory is the part of the supply chain where the system and the warehouse disagree, and every downstream calculation inherits that disagreement. Automating the reporting on top of an inaccurate count produces faster wrong answers, which is the outcome to avoid.

Below: why the figures diverge, which tasks automate cleanly, what has to be true before forecasting, the supplier side, and how to measure whether it is working.

Ask two people in the same company how much of an item is in stock and you can get two answers without either being careless. One reads the system; the other walked the aisle.

That gap is the practical starting point for any supply chain automation, because it determines whether anything built on top of the data can be trusted.

This is worth resolving before choosing any tool, because almost every supply chain capability on offer — replenishment, forecasting, allocation — assumes the count is approximately right and degrades quietly rather than failing when it is not.

Why the figures diverge

Timing is the largest cause and the least dramatic. Goods received on Friday and recorded on Monday are physically present and systemically absent for three days, and any calculation run in that window is wrong through no error.

Then there is unrecorded movement — samples, breakages, internal transfers, returns processed at a different location — each individually small and collectively enough to make a count unreliable.

And unit-of-measure confusion, which is the one that produces the spectacular errors: a case counted as a unit, a kilogram recorded as a gram. These are rare and expensive, and they survive precisely because they are rare enough that nobody builds a check for them.

Which tasks automate cleanly

The ones with a definite rule and a verifiable output. Consolidating stock positions across locations into one view. Reconciling receipts against purchase orders. Generating replenishment proposals from defined thresholds. Chasing suppliers on overdue confirmations.

What these share is that a person can state the rule in a sentence and check the result. Each of them also consumes a surprising amount of qualified time today, usually spread thinly enough that nobody has measured it.

The tasks to be careful with are the ones where the rule exists but the data does not support it — automatic reordering against a stock figure that is systematically wrong will order confidently and wrongly, at volume, without anyone reviewing it.

What has to be true before forecasting

Demand forecasting is the capability everyone wants and the one most often built on unsuitable foundations. It needs three things: history that reflects demand rather than sales, a stock figure that is approximately right, and lead times that are measured rather than assumed.

The first is the subtle one. Sales history records what was sold, which excludes what customers wanted and could not get. Forecasting on sales alone teaches the model to reproduce the stockouts, and the effect compounds each cycle.

The third is the one that quietly invalidates the output in this region. A lead time recorded as thirty days when imports actually take forty-five in a bad month makes every safety stock calculation optimistic in exactly the periods when it matters.

For importers in Colombia and Mexico this is rarely a small correction. Port congestion, customs inspection and inland transport each add variability that the planning parameter usually treats as a constant, and the difference between the assumed lead time and the observed distribution is the single input most worth measuring before any forecasting tool is evaluated.

The supplier side, which is mostly documents

A large share of supply chain administrative work is document handling: order confirmations, shipping notices, customs paperwork, invoices arriving in a dozen formats from a dozen suppliers.

This automates well because the documents are structured enough and the volume is real. It is also where the largest measurable time saving usually sits, ahead of anything analytical, and it requires no forecasting capability at all.

The realistic scope is per-supplier rather than universal. The five suppliers that account for most of the volume are worth automating properly; the long tail is not, and trying to cover it is how these projects lose their schedule.

Where the automation touches customs and imports

Any inventory that crosses a border acquires a second set of documents and a second set of timings, and both are usually held outside the systems the operations team can see. The customs broker knows where the shipment is; the ERP knows what was ordered.

Automating the bridge between those two — pulling status, matching it to the purchase order, and updating the expected arrival — removes a daily chase that in most importing companies is somebody's morning. It also produces the first honest lead-time history the organisation has ever had, as a by-product.

The constraint is that the broker's data arrives in whatever form the broker sends it: a spreadsheet, a portal, an email. That is a document automation problem rather than an inventory one, which is a useful thing to recognise early because it changes who should build it.

Cycle counting is the unglamorous fix

Where the count is unreliable, the remedy is not a better system. It is counting more often, in smaller slices, prioritised by value and movement rather than counting everything once a year.

This is not automation and it is the prerequisite for it. An organisation that adopts cycle counting typically finds its accuracy improves enough within two quarters to make the automatable calculations worth automating.

It also produces something more useful than a number: a record of where the discrepancies concentrate. Those locations and item classes are where the process is actually broken, and that is a shorter list than "inventory accuracy" implies.

The counting itself is worth automating only at the edges — generating the count lists, routing them, capturing the results and flagging the variances above a threshold. The count is done by a person, and trying to remove that is how organisations end up with an accurate system and an inaccurate warehouse.

How to measure whether it is working

Not by hours saved alone. The measures that matter are stock accuracy by location and class, the share of orders that could be fulfilled from available stock, and the value of stock that has not moved in a defined period.

That third one is where the money usually is. Excess stock is a cost that does not appear as a line item and is funded quietly, and it is one of the few supply chain figures where an improvement can be stated in currency without an argument about assumptions.

Reviewing these monthly, per location, is enough. Reviewing them in aggregate hides exactly the concentration that makes them actionable.

One caution on the first of them: accuracy measured as a percentage of items is flattering, because most items are low value and rarely move. Measured by value, or restricted to the items that turn over, it produces a number that is worse and considerably more useful.

Where to start

With the count, honestly measured on a sample, and with the document volume from the largest suppliers. Those two determine whether the first project is a data problem or an administrative one — and it is usually the second.

A process automation assessment establishes both and calculates the return with local costs, so the automation plan begins where the data already supports it rather than where the presentation was most persuasive.

Frequently asked questions

Why does the system disagree with the warehouse?

Mostly timing — goods received and recorded days apart — plus unrecorded movement such as samples, breakages and transfers, and occasional unit-of-measure errors that are rare and expensive.

Which supply chain tasks automate cleanly?

Consolidating stock across locations, reconciling receipts against purchase orders, generating replenishment proposals from thresholds, and chasing overdue supplier confirmations.

What is needed before demand forecasting?

History that reflects demand rather than sales, a stock figure that is approximately right, and measured lead times. Forecasting on sales alone teaches the model to reproduce past stockouts.

Why do lead times matter so much in the region?

Because a lead time assumed at thirty days when imports actually take forty-five in a bad month makes every safety stock calculation optimistic exactly when it matters most.

What fixes an unreliable stock count?

Cycle counting — counting more often in smaller slices, prioritised by value and movement. It is not automation, and it is the prerequisite for it.

How should the result be measured?

Stock accuracy by location and class, the share of orders fulfillable from available stock, and the value of stock that has not moved. The third is usually where the money is.