What AI First Should Mean in Procurement Software

Every procurement tool I see this year calls itself AI first.

Some of them mean it. Some of them added a chat box to a dashboard. From the outside, the two look the same.

So here is what I think AI first should mean in procurement software. Not as a slogan. As a set of choices you can test.

Start with what AI is actually good at

Language models are very good at a few things that procurement has always done by hand.

Reading. An invoice line that says "SVC CHG MAINT Q3 BLDG 4" means something. A model can work out what.

Understanding messy structure. Every supplier price list, contract tracker and plant spreadsheet has its own layout. A model can read the sheet, work out which column is the unit price, and ask about the one it cannot tell.

Explaining. A result that takes an analyst ten minutes to explain can be put into three plain sentences.

Doing the tedious work nobody wants. Grouping five spellings of the same supplier. Spotting the line that was classified in the wrong place. Proposing where a new category belongs.

That is a lot of value. It is also a specific kind of value. It is about language and judgment on messy inputs.

Then be honest about what it should never do

A language model predicts the next word. It does not calculate.

Ask a model for your total spend with a supplier and it will give you a number. It will sound confident. Ask again and it may give you a different one.

In procurement, numbers go in front of finance. A savings figure has to survive a CFO asking where it came from. A model's arithmetic will not survive that question.

So the first rule of AI first, in my view, is this. Deterministic code produces every number. AI never does.

In SpendCraft, the AI reads, explains and suggests. A query against your data produces the figure. Same question, same data, same answer. Every time. When data is missing, it shows as missing. It is never estimated.

AI first means the data work is done by AI too

Most spend tools put their AI on top. A chat box over a dashboard over data that is still a mess.

That is backwards. The hardest part of procurement analytics has always been the data. Supplier names that do not match. Categories that never fit. Taxonomies nobody owns.

So that is where AI should work first.

In SpendCraft, Caddies do this work alongside your team. A Caddie is an assistant built for one hard problem. Some of them are built for the data itself:

• Building and shaping your taxonomy. A Caddie proposes the top levels from how your organization is set up and who owns which categories, then the lower levels from your own spend. You confirm what is new. You keep refining as you go.
• Correcting classification. When a line is in the wrong place, a Caddie helps you fix it, and the fix becomes a rule, so the same line does not land in the wrong place next month.
• Normalizing vendors. A Caddie proposes which supplier names belong together, and you confirm. Ten names become one supplier, and new aliases join it as they arrive.

In every one of these, the AI proposes. A person confirms. That is not a limitation. It is the point.

AI first means the work does not end at the chart

A dashboard shows what happened. Procurement needs help with what to do next.

A renewal is due. A price increase letter arrives. An invoice may have been paid twice. Each of those is a decision, and each one used to mean exporting data to a spreadsheet and starting from scratch.

This is the other place AI belongs. Not answering a question in a chat window, but working a problem with you.

A Caddie brings the evidence for one specific situation: the price history, the alternatives you already buy from, the invoices behind a number. It reads the files you bring in whatever shape they are in. It may suggest a decision and shows why. Any change to a number waits for you to accept it. Only your confirmed choice counts.

AI first means the product sets itself up

The old pattern for enterprise software was months of configuration before anyone saw anything useful.

AI first should flip that.

When your spend data lands in SpendCraft, Spaces are generated for you. Where money is moving, what decisions are coming up, what may be owed back, how healthy your data is. They work on day one from spend data alone, before anyone uploads anything else.

They are ordinary Spaces. Keep them, narrow them to one category, or remove them. And when you work something out in a Caddie, it goes into a Space where it is kept and tracked over time, with Scouts watching the data underneath and telling you when something changes.

Nobody has to build the first view. The product does that. People decide what matters from there.

AI first means the decision stays with people

This is the part I care about most.

The pressure in AI right now is toward autonomy. Let the agent decide. Let it act.

In procurement, I think that is the wrong goal. Supplier decisions carry context no dataset holds. A relationship that matters for reasons outside the spend. A risk a category manager heard about on a call. A site that needs a local supplier.

A good caddie on a golf course reads the course, recommends a club, and hands it over. The player swings.

That is how I think AI should work in procurement. Propose with evidence. Leave the call with the person who owns it.

Five questions to test any "AI first" tool

If you are evaluating tools, these are the questions I would ask.

1. Ask the same question twice. Do you get the same number both times?
2. Ask where a figure came from. Can the vendor trace it to transactions, or does it come from the model?
3. Ask what happens when data is missing. Does the tool say so, or fill the gap?
4. Ask who does the data work. Is AI cleaning and classifying the data, or only answering questions on top of it?
5. Ask who makes the decision. Does the tool propose and wait, or act on its own?

The answers tell you more than any demo.

AI where it helps. Math where it counts.

AI first should not mean AI everywhere. It should mean AI first in the places it is genuinely better than people: reading, structuring, explaining and doing the tedious work. Math where the numbers live. People where the decisions are made.

That is the line we drew when we built SpendCraft. I wrote more about the architecture in How SpendCraft Thinks About AI.

The Savings OS.