What actually breaks when you automate purchasing

Operations · 12 August 2026 · 8 min read

Purchasing looks like the obvious first department to automate. The rules seem clear: stock drops below a level, an order goes out, goods arrive. Most automation projects start here and most of them stall in the same place.

The rule is not the problem

Setting a reorder point is trivial. Any system does it. What breaks is everything the reorder point assumes: that lead times are stable, that the supplier ships what was confirmed, that the price on the order is the price on the invoice, that a partial delivery is rare.

In a real operation none of those hold. The purchaser you are replacing was not calculating reorder points in their head. They were absorbing variance.

You are not automating a calculation. You are automating a judgement about who can be trusted this week.

What to automate first

Start with the parts where variance is low and the decision is genuinely mechanical:

• Price checks against agreed terms — flagging, not deciding
• Order confirmation matching, which is pure comparison
• Chasing overdue confirmations, which is a timer and a message
• Building the order draft, leaving the send to a person

That last one matters more than it sounds. A system that prepares and a human who approves gets you most of the time saving with none of the trust problem. Move the approval later, once the flagging has been right for a quarter.

Where AI earns its place

Not in the reorder calculation. In the reading: supplier emails, PDF confirmations, invoices that never match the format they matched last month. That is where a person was spending their afternoon, and that is where a model is genuinely better than a rule.

The measure

If your purchasing automation has not reduced the number of surprises, it has not automated anything. It has just made the same surprises arrive faster.