Why we didn't use an AI agent for order entry

Everyone is talking about AI agents at the moment: AI that decides for itself which tools to call, in what order, to get a job done. It's an obvious question for order entry. Why not give an agent access to Kinetic and let it work out how to enter the order?

We tried it. For order entry, the agent was slower and cost more than the fixed pipeline we already had. So lowki.ai's order entry doesn't use one today.

What a fixed pipeline means

lowki.ai's order entry works the same way every time:

1. An order arrives as an email, PDF or spreadsheet.
2. The AI reads it and turns it into a fixed data structure: customer, ship-to, lines, part numbers, quantities, prices, dates.
3. The customer and parts are matched against your Kinetic data, and the order is created through Kinetic's own business objects.
4. Anything it's unsure of is flagged for a person to check. We recommend a hold BPM so every order is reviewed before it's processed.

In order entry today, the AI's job is understanding the order. It doesn't choose what to do next, it has no tools, and it can't call anything in Kinetic. The process around it is fixed and the same every time.

Why the agent lost

Order entry is a fixed-shape job. Every order needs the same steps in the same order. An agent has to rediscover those steps each time: look up the customer, look up the parts, work out which calls to make, check its own work. Every one of those decisions is another round trip to the model, which takes time and costs money. You're paying the AI to work out something you already know.

It also makes the results harder to trust. A fixed pipeline does exactly the same thing with every order, so when something goes wrong you can see where. If the AI's output doesn't fit the structure, the code rejects it and a person looks at it. An agent can take a different route on Tuesday than it did on Monday, and explaining why is a lot harder.

And for an ERP write, predictable beats clever. Your auditors, your finance team and your IT manager all want to know what the system will do before it does it.

Where agents do make sense

This isn't an anti-agent post. Agents are good at jobs that don't have a fixed shape: investigating an exception, working out why an invoice won't match, triaging a document nobody has seen before. Those are worth exploring, and we are.

The rule we work to is simple. If you can write the steps down, write them as code. Use AI for the part you can't write down, which for order entry is reading the document.

A question worth asking any vendor

If you're looking at AI order entry, ask what the AI is actually allowed to do. Does it only read documents, or can it decide what to call in your ERP? Neither answer is automatically wrong, but you should know which one you're buying, and what it costs per order.

Related: Sales order entry · Built on Kinetic, not beside it · AI Order Processing Is Not OCR