Rules first, models second: how Alfred zooms in on the right value

When we started designing Alfred, the obvious idea was on the table: send every invoice to an artificial intelligence model, ask it for the data and be done.

In a demo, it works beautifully. In a company’s day to day, three problems show up.

The cost grows with every invoice. And invoices never stop coming.

The same question doesn’t always get the same answer. If an invoice produces one result today and a different one tomorrow, nobody can explain why an amount ended up in the accounting system.

Mistakes look convincing. An invented value in the right format is the worst possible error: nobody suspects it.

So we turned the question around. Instead of “what can AI do?”, we asked: “how would a very careful person get to the right value?”.

Like a photographer zooming in

Think about how you look at an invoice you’ve never seen. First, from a distance: you recognize the shape, where the total usually sits, where the supplier’s logo is. Often that’s enough.

When it isn’t, you get closer. You read more carefully, compare, look for clues. And only for what still isn’t clear do you stop and really think.

Alfred works the same way. It looks at each invoice in finer and finer stages, and each stage focuses only on what the previous one couldn’t solve. The simple things get solved quickly and predictably. Artificial intelligence comes in at the end, with less work and more context, exactly where it adds the most.

It’s a zoom: from the full picture to the exact detail.

Every value has to make sense

Finding a value isn’t enough. It has to make sense for your business.

If your company’s purchase order numbers always follow a certain shape, a value that doesn’t is discarded, even if it looks reasonable. If a total doesn’t match its taxes, something is wrong. Alfred checks every value against those rules before accepting it.

It’s the difference between “I found something that looks like a total” and “this is the total”.

What you get from this approach

• Consistent results. The same invoice produces the same result, today and a year from now.
• Predictable costs. The most expensive part of the process is only used when it’s needed.
• Data you can explain. Every value knows where it came from.
• Real autonomy. The system doesn’t need someone watching over it to work well.

The biggest lesson

Artificial intelligence works better as a piece of a good system than as the whole system. Surrounded by stages that prepare it and rules that validate it, it’s extraordinary. On its own, it’s a source of surprises.

That’s the judgement we apply in all our applied AI projects. And if you want to see it working with your own invoices, meet Alfred.