You Probably Don't Need the Flagship AI Model
Same output, $800 a month or $40. Why most procurement work doesn't need a flagship AI model, and what to build instead: your logic, library and RAG.
THOUGHT LEADERSHIP
Bogdan Asproiu
11/10/20264 min read


A procurement tech podcast did the maths this week on something most teams haven't looked at yet. It priced 5 everyday procurement tasks on a flagship AI model and on a right-sized one. Same output, roughly $800 a month against $40. Depending on the model you pick, the same task can cost up to 125x more.
I agree with all of it. Tokens are turning into a spend category and somebody in the team should own it. But I think the conversation stops one step short. It treats the bill as a pricing problem: cheaper model for the simple work, premium model for the hard work. I'd look at it first as a specification problem, and then as make vs buy.
We would challenge this spec in any other category
If a stakeholder asked for the top-spec laptop for every employee, including the ones who only open email, no category manager would sign it off. We'd ask what the work needs and buy to that. With AI most teams skip the question. The flagship is the default because it's the one in the demo, and nobody gets blamed for buying the best. Similarly for E1 vs E5 O365 licenses - some need Word and Outlook, others Teams and Power BI.
Most procurement work doesn't need it. Classifying spend lines, pulling renewal dates and notice periods out of a contract, comparing 3 quotes against a rate card, drafting the first version of an RFP section. This is structured, repeatable work with a right answer. A mid-sized model handles it fine if you give it the right material.
What the big model is being paid for
That last condition is where the money goes. A general model knows nothing about your suppliers, your contracts, your policy or how your business buys. So you make up for it in the prompt. You paste the whole contract, the policy and the background, and you pay for them again on every follow-up, because each follow-up re-bills the context you already sent (the episode makes this point well).
So a good part of the flagship bill has little to do with intelligence. You're paying a very large model to read the same 80 pages again, and to guess at context that already sits somewhere in your shared drive.
Procurement has a name for the fix and it's demand management. If the system finds the 3 clauses that matter and sends only those, the token count drops before the choice of model even comes up.
Build the part that is yours
This is where I go further than "pick a cheaper model". There are 4 things I'd build and own:
1 - Your logic. How you assess a supplier, how you build a negotiation position, where your approval thresholds sit. Write it down as steps the system follows, so you're not hoping a bigger model works it out from scratch each time.
2 - Your library. Contracts, templates, rate cards, past RFPs and lessons learnt, cleaned and tagged. There is no AI first without data first. A flagship on top of a messy shared drive gives you expensive partial answers.
3 - Your RAG. The retrieval piece that finds the right pages in the library for the question asked, so the model reads 2 pages and not 200.
4 - Your context. Who is asking, which category, which market, what was agreed last time. The system should keep this so nobody has to retype it.
With those in place the model at the end has an easier job. It reads a small amount of the right material and writes a clear answer. Mid-sized open-weight models do that well, and you can run them on your own hardware, where there is no per-token bill and the contracts stay in the building.
That's how we run Procurement Union. The companion sits on an open-weight model on our own servers, with the library, the retrieval and the reasoning method built around procurement. I wrote about why the archive matters more than the model in [LINK: The Archive Beats the Ensemble]. When a better open model comes out we swap it in, and the library keeps its value.
It isn't free. The hardware is capex and cleaning the library is real work (most of the work, from my experience). But the cleaned library stays yours whichever model you use next year, and a token bill leaves you with nothing at the end of the month.
Where the flagship still earns its fee
There are cases, and I'd rather name them than pretend otherwise. A one-off, open-ended question in an area where you have no library yet. Exploratory work on a new category or an unfamiliar market, where you don't know what to retrieve. A pile of long, messy documents nobody has structured.
Pay for it there, per use, and know that you're doing it. I'd treat the flagship like a specialist consultant: you bring one in for a specific problem, and you wouldn't keep one on retainer for data entry.
What I would do this quarter
1 - List the AI tasks your team runs and tag each one as repeatable or open-ended.
2 - Ask your vendors (and IT) which model sits behind each tool and how follow-ups are billed.
3 - Take one repeatable task, clean the documents behind it, and test it on a smaller model with retrieval. Compare the outputs side by side before believing anyone, me included.
Tokens as a spend category is the right conversation to have. I'd run the same checks on it that we run on every other category: is the spec right, is the volume right, and should we be making this in-house. For most procurement work I think the answers point to a smaller model and a library of your own.
