Let AI negotiate the spend you never get to.

wop. field notes

Hey {{first_name|there}},

Your team buys the same things from established suppliers. The quote comes in, someone checks the basics, and the order goes through. There might be a better deal available, but running another negotiation competes with everything else procurement has to get done.

I'm talking here about buying autonomous negotiation software for those routine purchases, before the purchase order is raised. The ambition is to give more of that spend a proper commercial conversation without making a buyer supervise every exchange.

For me, that's a much more useful starting point than deciding we need “an AI agent”. It gives us a job to evaluate and a reason the business should care. Then we can get into whether the software is actually capable of doing it.

Start with the deal you want it to improve.

Choose a group of purchases where you understand the requirement, the suppliers and the current buying process. Work out what would make those deals better. That could include price, payment terms or delivery arrangements, depending on what matters to the business.

Be specific about the trade-offs. Say a supplier offers a lower unit price if you double the order quantity. If you don't need the extra stock, that price reduction may be a fairly expensive achievement. The system needs to understand how the quantity changes the value of the deal.

Work through which terms can move, which limits are fixed and how acceptable combinations should be valued. Agree what the software may conclude autonomously and what should come back for a decision. Those are commercial choices procurement should make with the relevant people in the business.

They give you a useful basis for evaluating the product. Writing them down doesn't, by itself, give a language model the ability to negotiate them well.

Find out what actually makes an offer.

Ask the vendor to follow one supplier conversation through its product. How does it interpret a counteroffer, calculate the value of the terms and decide what to offer next? What checks run before an offer reaches the supplier?

An LLM choosing an offer from written instructions and a negotiation engine calculating offers within programmed rules need different scrutiny. A product can also combine them. “AI-powered” is doing a remarkable amount of work if nobody can explain which part makes the commercial decision.

I work at Pactum, so that relationship is worth being clear about. Its published architecture describes rule-based offer calculations using defined negotiable terms and values, with LLMs supporting other functions. It illustrates why we need to look at the components doing the work, rather than assume every agent is an LLM following a brief.

Software-enforced limits and commercial competence are separate things to establish. A product can stay within every limit and still make a poor deal. Equally, a persuasive explanation from a model doesn't show that a prohibited offer will be blocked.

Make the pilot prove the commercial case.

Give the vendor representative cases from the purchases you've selected, with sensitive data handled through your approved process. Include a straightforward deal, a counteroffer that changes the economics, and a supplier who won't agree. Decide beforehand what an acceptable outcome looks like for each. Walking away can be the right result.

For the extra-stock example, watch whether the product recognises the changed quantity, evaluates the whole commitment and makes an appropriate counteroffer or stops. Follow the agreed terms into the purchasing system too. Establish whether the product transfers them automatically, needs an integration or leaves someone to re-key them. That work belongs in the evaluation.

Then use a limited live pilot to assess repeated performance, once the agreed technical, security and commercial checks are satisfied. Compare the outcome with a credible baseline for equivalent goods, quantities and terms. Include software and operating costs when judging the benefit. A supplier's opening quote on its own can flatter the savings story.

Keep track of the buyer time spent checking, correcting and intervening, as well as the negotiations completed and the commercial results. Ask suppliers what the experience was like. A process that creates new chasing for either side has given you something else to fix before expanding it.

Use the results to decide where it should run, where it needs more work and where it doesn't earn its place. Expand the scope when the evidence supports it, and keep reviewing performance as the suppliers, terms or product change.

This is the standard I'd want from procurement AI: more spend gets commercial attention, acceptable deals reach the business, and the benefit survives the cost of running the system. Procurement sets the commercial direction and judges the results. The software has to earn the work we hand over.

Daniel