← All articles
- Automated Negotiation
- Procurement AI
- Supplier Power
- Category Management
Everyone Wants AI Bots Fighting for Them. Nobody Wants AI Bots Fighting Against Them.
Automated negotiation is documented in one setting: high-volume, low-value, many-supplier categories. Where the counterparty holds leverage, software does not change who sets the terms.
· 5 min read
By EXOS Research Team
The appeal of an autonomous negotiation bot is easy to picture: it works nights and weekends, never gets tired of a counter-offer, and runs hundreds of negotiations at once. What's harder to picture is the same bot sitting across the table from someone who has real leverage and no particular reason to play along.
The landlord test
Here's a quick way to check whether a category is a good fit for an automated negotiator: would you send the bot to renegotiate the lease on your office building? Almost nobody would. A landlord with one building and one tenant in a tight market has no reason to engage with a structured, automated counter-offer the way a tail supplier competing for volume does — and a specialised component manufacturer with no qualified alternative, or a cloud provider running a standard enterprise contract, are in the same position. Nothing here is a measured finding about how landlords or cloud providers actually respond to bots — it's a way of testing intuition quickly, not a data point, and it holds up because the underlying reason is simple: a counterparty with leverage sets the terms of the conversation, and software doesn't change who has leverage.
Bots are good at this, and the evidence says exactly where
None of this is an argument that automated negotiation doesn't work. Walmart's rollout of an AI negotiation tool from Pactum reportedly closed deals with 68% of approached suppliers, against a 20% benchmark, for a 1.5% cost saving — a genuinely strong result. The detail worth noticing is the scope: this ran against thousands of small suppliers at once, the tail-spend end of procurement that buyers otherwise don't have the hours to negotiate individually. A December 2024 analysis from INSEAD's business school puts the boundary plainly: automated negotiation, on the evidence available, is currently "limited to small-value, few-issue, repetitive and long-tail negotiations." That's not a knock on the technology — it's a description of where it's actually been shown to work, and it happens to line up with the landlord test rather than contradict it.
What happens when trust actually gets measured
A 2026 study published in Frontiers in Artificial Intelligence surveyed 238 procurement, legal and digital-transformation professionals across the EU and Southeast Asia and found something specific: using more AI in negotiation was linked to higher perceived efficiency, but lower trust in the AI system itself, once the automation started replacing the human touchpoints and explanations that make a recommendation defensible. Trust actually rose with moderate AI use and only fell once automation dominated the interaction — not a blanket rejection of the technology, a curve. In the interviews behind the survey, several participants in cross-border and public-tender negotiations described real discomfort when only one side of the table had access to AI tools at all. That's the closest thing to hard evidence for the landlord-test intuition: not that suppliers refuse bots out of spite, but that trust in AI-mediated negotiation is measurably fragile precisely where power and stakes are highest — which is the same territory a landlord, a sole-source manufacturer or a strategic cloud contract already occupies.
Automating both sides doesn't close the gap
A natural question is whether the problem disappears once suppliers use their own AI tools too — two bots negotiating should at least be a fair fight. Recent large-scale studies of AI-versus-AI negotiation say otherwise: outcomes track the relative strength of the two systems, not a neutral split, and in one benchmark the gap in realised value depending on which side had the stronger model reached nearly 15 percentage points. Putting a bot on both sides of the table doesn't cancel out an existing power imbalance — on this evidence, it can just move the imbalance from the human relationship into the software.
Where the leverage actually comes from
None of this leaves a buyer facing a powerful supplier with nothing to do. It means the lever isn't a better-worded automated message — it's decomposing what the supplier's price is actually built from, challenging a specification that quietly locked out every other bidder, building an alternative that's real enough to be credible, and getting the multi-year pricing terms structured before signature rather than disputed after. That work is what a negotiator does with the time an automated tool would otherwise have spent running the same script on a supplier who was never going to respond to it.
The full evidence, sourcing and the Kraljic-quadrant breakdown of where automated negotiation is and isn't a fit are on the methodology page: Supplier power and automated negotiation: the counterparty decides whether a bot can work. See also the companion piece on process length: The automation boundary.
Sources: Mejri & Skhiri, "Trust by design in AI-augmented procurement systems," Frontiers in Artificial Intelligence (2026); Vaccaro, Aral et al., "Advancing AI Negotiations" (arXiv:2503.06416, 2025); Zhu et al., "The Automated but Risky Game" (arXiv:2506.00073, 2025); Pactum/Walmart figures via Harvard Business School Online and Supply Chain Management Review (2025–2026); INSEAD Knowledge (Dec 2024). Full citations, including the Kraljic and power-asymmetry theoretical foundation, are on the methodology page above.