Walmart's AI Negotiates With 2,000 Suppliers at Once. Here's What Mine Couldn't Do.

Walmart lets an AI negotiate with its suppliers. Not assist — negotiate. The system runs conversations with roughly 2,000 suppliers at once, and the published results are specific: a 3% average cost gain, payment terms extended by an average of 35 days, and 68% of approached suppliers closing a deal with a machine. Suppliers even rate the experience well — 83% called it easy to use. This isn't a startup pitch deck. It's Walmart, Maersk, Veritiv, and a growing list of Fortune 500 procurement teams, and it's now a Harvard Business Review case study.

I spent years negotiating freight and 3PL contracts by hand, building a business from $0 to $10.4M in peak revenue. Most of that work was exactly the category this technology now automates — high-volume, low-value, repetitive supplier deals. So I want to do two things. First, be honest about which parts of that job the bot genuinely does better than I ever could. Second, walk through the exact moment in a real 3PL negotiation where the deal turned on something no autonomous agent, as they're built today, would have done — and why that matters more than the 3% headline number.

What the bot is actually good at

Strip away the marketing language and the mechanism is simple: Pactum's agents work through payment terms, discount percentages, and rebate structures across thousands of supplier relationships simultaneously, log every exchange for audit, and run 24/7 without getting tired, distracted, or emotionally invested in a single deal. For tail spend — the long list of small, low-value, high-volume contracts every large buyer has and no procurement team has headcount to negotiate individually — that's a genuinely good use of the technology. Research from INSEAD's Negotiation and Conflict Management Collaborative describes this precisely: AI-automated negotiation today is confined to small-value, few-issue, repetitive, long-tail deals, partly to contain the damage from the AI getting something wrong, and partly because more complex negotiations depend on things that can't yet be automated — trust-building chief among them.

That's not a criticism of the technology. It's a scope statement. And it matters, because it tells you exactly where the bot's 3% number is coming from: high volume, thin margin for error, low relationship stakes. That's a different category of deal to the one I'm about to describe.

The deal I turned down before we negotiated a single dollar

When we went 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 — 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 single rejection reset the entire negotiation before price was even on the table: everything that followed had to be reframed as a service fee, not a revenue split.

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

Where the leverage actually came from

Before we went back to any provider with a serious counter-position, we did the unglamorous part: active market research, working out who could plausibly deliver what we needed and on what terms. That gave us two things.

The first was a credible walk-away. 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 that distinction is exactly the kind of thing a supplier can sense over months of back-and-forth in a way they can't sense from a chat interface. The second was clarity about our own commitment: we knew precisely how much volume and warehouse space we could put behind the deal if the terms made sense.

That let us change the shape of the negotiation. Their opening rate was priced for a single location, with nothing on the table for scale. We didn't just push back on that number — we proposed a different pricing model entirely: a sliding scale, where a bigger volume and space commitment from us earned a better rate across the board. That's a structural counter-proposal, not a concession request, and it only existed because we'd shown up with a real commitment to negotiate against.

Three to five months, not three exchanges

This wasn't a single session. It went back and forth for three to five months. Over that time, the negotiation was as much about building enough trust for them to believe our volume commitment was real as it was about the numbers themselves. We needed them to understand two things at once: that we knew where the market's floor actually was, and that we'd walk if a workable middle ground didn't turn up.

That combination — patient, relationship-based pressure applied over months, backed by a credible alternative — is precisely what the current generation of negotiation AI is not built to do. Not because the maths is too hard, but because trust isn't a variable you can hand to a system that has no continuity of relationship with the person on the other side, and no real cost to walking 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 been direct about it: AI will not automate decision-making in a tender process. That's not a technology limitation so much as an accountability one — a public tender needs a human who can be held to account for the outcome, and no agency is going to hand that off to a bot mid-adoption-curve.

But that doesn't mean AI stays out of government procurement. It arrives sideways first. Procurement teams are already using AI to rehearse negotiations before the real one — sitting across a simulated negotiation with an AI playing the supplier to surface tactics and positions they hadn't considered. And on the supplier side, AI tools built specifically to help small businesses write and pre-check government tender bids are already commercially available in Australia. The asymmetry is the story worth watching: suppliers get negotiation and bid-writing AI before the tender process itself does. If you sell into government, or plan to, that arrives well before any agency deploys 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 itself. First: work out whether the deal structure you've been handed is actually the right one before you argue about the number inside it — a bot optimises within the frame it's given, and if the frame is wrong, winning inside it is still losing. Second: know what your own walk-away is actually worth before you use it as leverage, because the difference between a real alternative and a bluff is exactly the kind of thing that becomes visible over a multi-month negotiation, whether the other side is a person or a machine trained to spot bluffing patterns in the data it's seen before.

The 3% number is real, and for tail spend at Walmart's scale, it's a legitimately good outcome. 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 rejecting the number entirely, and being willing to spend months proving we meant it.