I Said No Before We Talked Price. Walmart's AI Wouldn't Have.

I've been spending a lot of time lately reading into how large procurement systems are changing — the ways big buyers manage huge volumes of supplier contracts, and how AI is starting to get used to review contract performance and supplier relationships at a scale no procurement team could manage manually. That's how I landed on Walmart's negotiation system. Not because it's the newest thing out there, but because it's one of the few examples with real numbers attached, and because it sits close to work I've actually done — years negotiating freight and 3PL contracts by hand, building a business from $0 to $10.4M in peak revenue. A lot of that was high-volume, low-value, repetitive supplier deals, which is exactly the category Walmart has now handed to a machine.

Here's what I think Walmart got right. The system runs conversations with roughly 2,000 suppliers at once, and the results are public: a 3% average cost gain, payment terms pushed out by an average of 35 days, and 68% of suppliers it approached closing a deal with a machine. Suppliers rate the experience well too — 83% called it easy to use. That's not a vanity metric. It's Walmart, Maersk, Veritiv, and a growing list of Fortune 500 procurement teams doing this, and it's now a Harvard Business Review case study. For tail spend, this makes total sense to me — Walmart has thousands of small suppliers no procurement team could ever negotiate with one by one, and the logic of putting a bot on that is sound. I'd have used it myself if it existed when I was doing this.

I've had a few conversations recently with people working in procurement inside government departments, and the pattern that keeps coming up is the same one Walmart is solving for — too many contracts, not enough people, and no time to give every supplier relationship proper attention. So it's easy to see why something like this is appealing. But the more I read into how Walmart's system actually works, the more I kept coming back to a negotiation I ran that this system, as built, couldn't have run. Not a bigger deal, not a more sophisticated one — just a different kind, where the thing that won it wasn't inside the parameters the bot is designed to optimise. I want to walk through that moment, and be honest about where it leaves the 3% headline.

What the mechanism actually is

Mechanically, Pactum's agents work through payment terms, discount percentages, and rebate structures across thousands of supplier relationships at once, log every exchange for audit, and run at 2am without getting tired or talking themselves into a worse deal because they liked the rep on the call.

There's research from INSEAD's Negotiation and Conflict Management Collaborative that lines up with this: AI-automated negotiation right now stays confined to small-value, few-issue, repetitive, long-tail deals. Partly to limit the damage if the AI gets something wrong. Partly because the more complicated negotiations lean on things nobody's figured out how to automate yet — trust-building being the big one.

That tells you where the 3% is coming from. High volume, thin margin for error, low relationship stakes per deal. Which is a different animal to the one I want to talk about.

The deal I turned down before we'd negotiated a single dollar

When we were looking for a 3PL partner who'd actually give us control over our own fulfilment, one of the strongest candidates in the market didn't want to quote us a rate at all. They wanted a profit share instead — a cut of what we made, not a fee for what they delivered.

We said no immediately. Not because the number was wrong — there was no number yet — but because a profit-share model meant handing decisions about our own P&L to someone else's incentives. That rejection reset the whole negotiation before price was ever on the table. Everything after had to be reframed as a service fee, not a revenue split.

An AI agent negotiating tail spend doesn't do this. It works inside the deal structure it's handed — payment terms, discount percentage, volume commitment — and optimises within those. It doesn't reject the structure itself. We did, and that's where most of the value in this negotiation came from, months before we argued about an actual figure.

Where the leverage came from

Before we went back to anyone with a serious counter, we did the unglamorous part — active market research, working out who could actually deliver what we needed and on what terms. That got us two things.

A credible walk-away, for one. We genuinely had somewhere else to go, so our willingness to leave the table wasn't a bluff. It didn't need to be — and I think that's the kind of thing a supplier picks up on over months of back-and-forth, in a way that doesn't come through in a chat window. And second, we knew exactly how much volume and warehouse space we could commit if the terms made sense.

That let us change the shape of the negotiation. Their opening rate was priced for a single location, nothing on the table for scale. We didn't just push back on the number — we proposed a different pricing model, a sliding scale, where more volume and space from us earned a better rate across the board. That's a structural counter, not a discount request, and it only worked because we'd turned up with a real commitment behind it.

Three to five months, not three exchanges

This wasn't one session, or three. It went back and forth for three to five months. A lot of that time was spent building enough trust for them to believe our volume commitment was real — as much as it was about the numbers themselves. They needed to believe two things at once: that we knew where the market's floor actually was, and that we'd walk if we didn't get somewhere workable.

That kind of patient, relationship-based pressure over months, backed by a real alternative, isn't something the current generation of negotiation AI is built to do. Not because the maths is hard — because trust isn't a variable you can hand to a system with no ongoing relationship to the person on the other side, and nothing real at stake if it walks away.

Where this actually lands in government procurement

It won't land as autonomous tender negotiation any time soon. Australia's Department of Finance has said as much directly — AI will not automate decision-making in a tender process. That's less a technology limit than an accountability one. A public tender needs a human who can be held to account for the outcome, and no agency is handing that off to a bot mid-adoption-curve.

That doesn't mean AI stays out of it, though. It arrives sideways first. I've seen procurement teams already using AI to rehearse a negotiation before the real one — running a simulated negotiation against an AI playing the supplier, to surface tactics they hadn't thought of. And on the supplier side, there are AI tools built specifically to help small businesses write and pre-check government tender bids, already commercially available here in Australia. That's the asymmetry worth watching: suppliers get negotiation and bid-writing AI before the tender process itself does. If you sell into government, or plan to, that's arriving well before any agency has an autonomous negotiation agent of its own.

What to actually do with this

If you're negotiating against a system like this, or expect to be soon, two things matter more than the technology. Work out whether the deal structure you've been handed is the right one before you argue about the number inside it — a bot optimises within the frame it's given, and if the frame's wrong, winning inside it is still losing. And know what your walk-away is actually worth before you use it as leverage, because the gap between a real alternative and a bluff shows up over a multi-month negotiation whether the other side is a person or a machine trained on a lot of bluffs.

The 3% is real, and for tail spend at Walmart's scale, it's a good number. But the deal that got us a sliding-scale rate and full control of our own P&L didn't come from optimising a number. It came from turning the number down, and spending the next few months proving we meant it.