Supplier power and automated negotiation: the counterparty decides whether a bot can work

By EXOS Research Team

Automated negotiation is documented as effective in one setting: high-volume, low-value categories with many substitutable suppliers. Where supplier power is high — Strategic and Bottleneck categories — no cited success case operates, measured trust in AI-mediated negotiation falls as automation displaces relational cues without explainability or oversight, and automating both sides compounds rather than cancels a pre-existing asymmetry. Supplier power is the variable that decides which side of the automation boundary a category falls on.

Applies to: Any category where automated negotiation is being considered
Unit of analysis: One category, characterised by relative supplier power and substitutability
Position in the lifecycle: Before a negotiation tool is selected for the category

Autonomous negotiation tools are consistently reported as effective in one specific setting: high-volume, low-value, substitutable-supplier categories, where the same terms can be offered to any of many interchangeable counterparties. The evidence for that setting is real and quantified. It is also, on its own account, evidence for a boundary — every source describing where automated negotiation succeeds describes the same condition, and every recent study measuring how procurement professionals actually trust AI-mediated negotiation finds that trust falls, not rises, as automation intensity increases without human oversight and explanation. Supplier power is not incidental to this pattern; it is the variable that decides which side of the boundary a category falls on, and it is measurable independently of any specific technology.

Automated negotiation succeeds exactly where vendors themselves say it does

Pactum’s deployment with Walmart, reported by Supply Chain Management Review and Harvard Business School Online’s Negotiation Mastery course, is the most-cited working example of AI-run supplier negotiation: agreement reached with 68% of approached suppliers against a reported 20% benchmark, alongside a 1.5% cost saving. Both sources describe the same scope condition — this is tail-spend negotiation, run against thousands of small suppliers simultaneously, not a small number of strategic counterparties. A separate industry account of the same tool class puts it plainly: these systems address "the 80% of suppliers that buyers rarely have time to contact." INSEAD Knowledge, in a December 2024 analysis of AI in negotiation, states the boundary directly: "AI-automated negotiations are currently limited to small-value, few-issue, repetitive and long-tail negotiations." None of these sources report a comparable deployment against a concentrated, high-power supplier base — not because the case is untested and unreported, but because, on the evidence available, that is not where the tool class currently operates.

Automation raises efficiency and lowers trust in the same study

A 2026 mixed-methods study in Frontiers in Artificial Intelligence — six case studies across the EU and ASEAN (36 interviews), followed by a structural-equation-modelled survey of 238 procurement, legal and digital-transformation professionals — measured this directly rather than inferring it. Higher AI use was associated with higher perceived negotiation efficiency, and, in the same model, with lower trust in the AI system itself, where automation displaced relational cues and explanations were weak. The relationship was not linear: trust rose with AI use up to a moderate level and fell beyond it, an inverse-U pattern consistent with a specific mechanism rather than a blanket resistance to automation. Explainability was the strongest positive correlate of trust; human-in-the-loop design and visible governance were both independently positive. The qualitative phase named the mechanism the survey later quantified: participants, concentrated in cross-border and public-tender negotiations, described "ethical tensions and perceived power asymmetries when AI access is unequal." The study measures trust in the AI system, not a supplier’s willingness to negotiate with one — a boundary the authors state explicitly, and one this page keeps as well (see "What this does not argue," below).

Supplier power is measured, and it varies by Kraljic quadrant — this part is not new

The finding that power and dependence differ systematically across the Kraljic portfolio is not an AI-era discovery. Caniëls and Gelderman’s 2005 study, a survey of Dutch purchasing professionals published in the Journal of Purchasing and Supply Management, quantified "relative power" and "total interdependence" separately for each quadrant of the matrix and found the results consistent with the model’s own theoretical expectations: buyers hold comparatively more relative power in Leverage and Routine categories, and comparatively less in Strategic and Bottleneck ones. A related empirical strand from the same period, summarised in a review of electronic reverse auction research, reports that supplier resistance to automated bidding tools increases as supply markets tighten. Neither study says anything about artificial intelligence, and neither is used here as evidence about attitudes toward AI — they establish that the underlying economic condition this page depends on (power varies by category, and automated tools meet more resistance where it is high) predates AI negotiation tools and is not a claim invented for this argument.

When both sides automate, the asymmetry does not disappear

A further, more recent strand addresses a related question directly: what happens when both the buyer and the supplier use an automated negotiation agent. A large-scale autonomous-negotiation competition analysed by MIT and Cornell researchers in 2025 found that outcomes depend materially on the relative capability of the two agents rather than converging on a fair split — the stronger agent wins more, consistently. A separate 2025 benchmark of agent-to-agent negotiation in consumer markets found the choice of the stronger party’s model produces a materially larger swing in outcome than the choice of the weaker party’s — in one comparison, a gap of up to 14.9 percentage points in realised value depending only on which model played the stronger role. Automating both sides of a negotiation does not neutralise a pre-existing power asymmetry; on this evidence, it gives the more capable or better-resourced party a second, compounding advantage on top of its existing one.

Where each Kraljic quadrant sits, and why

Supplier power and automated-negotiation evidence by Kraljic quadrant
Kraljic quadrant	Typical relative supplier power	Automated-negotiation evidence	Appropriate AI mode
Routine	Low (Caniëls & Gelderman, 2005)	Matches the tail-spend condition every cited success case describes	Automate
Leverage	Low (Caniëls & Gelderman, 2005)	Matches the tail-spend condition every cited success case describes	Automate
Bottleneck	High (Caniëls & Gelderman, 2005)	No cited success case operates in this condition	Augment
Strategic	High (Caniëls & Gelderman, 2005)	No cited success case operates in this condition	Augment

This table is the supplier-power half of the automation boundary set out in full, alongside the process-length half, at The automation boundary: where accumulated error and supplier power rule out autonomous AI. The two pages describe the same boundary from two independent variables; neither is derived from the other, and both point to the same quadrants.

What this does not argue

This is not evidence that suppliers with market power refuse automated negotiation out of irritation or perceived disrespect — no source cited here tests that claim, and it is not made here as one. What the evidence supports is narrower and better established: automated negotiation’s documented successes sit specifically in low-power, high-substitutability categories; trust in AI-mediated negotiation falls, in a measured study, when the interaction displaces relational cues without offsetting explainability or human oversight; and capability asymmetry between automated agents does not cancel out a pre-existing power asymmetry when both sides automate. The 2026 trust study’s own stated limits apply here too: its sample is a non-probability, purposively recruited group of AI-exposed professionals, its findings describe trust in the AI system rather than buyer-supplier relationship outcomes directly, and its fieldwork covers the EU and ASEAN specifically rather than a global population. None of this argues against automation in the categories where the evidence shows it working. It argues that the categories where it is documented to work and the categories this site addresses are, on current evidence, different ones.

Where this runs in EXOS

Negotiation Preparation (BATNA/ZOPA) is the scenario built for the Strategic and Bottleneck condition this page describes: constructing a costed alternative and mapping the zone of possible agreement before a negotiation starts, rather than relying on counter-offer automation once it has. See also the process-length half of this boundary and where EXOS applies on the Kraljic matrix.

Sources

Evidence on AI-specific negotiation trust and behaviour (2024–2026):

Mejri, R. and Skhiri, S., "Trust by design in AI-augmented procurement systems: the roles of explainability, governance, and human oversight," Frontiers in Artificial Intelligence, 9:1752349 (2026).
Vaccaro, M., Caosun, M., Ju, H., Aral, S. and Curhan, J.R., "Advancing AI Negotiations: New Theory and Evidence from a Large-Scale Autonomous Negotiation Competition," arXiv:2503.06416 (2025).
Zhu, S., Sun, J., Nian, Y., South, T., Pentland, A. and Pei, J., "The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets," arXiv:2506.00073 (2025).
"AI in Procurement Automation: Use Cases for Negotiation," Harvard Business School Online, citing R. van Hoek, Negotiation Mastery (2026) — the Pactum/Walmart figures.
"Cognitive procurement with chatbots: Pros and cons," Supply Chain Management Review (2025) — same Pactum/Walmart figures, independent citation.
"The Power of AI to Shape Negotiations," INSEAD Knowledge (Dec 2024).

Theoretical foundation — power varies by category; predates and is independent of AI (not evidence about AI attitudes):

Caniëls, M.C.J. and Gelderman, C.J., "Purchasing strategies in the Kraljic matrix — a power and dependence perspective," Journal of Purchasing and Supply Management, 11, 141–155 (2005).
Beall, S. et al. (2003), as discussed in "Do all suppliers dislike electronic reverse auctions?", Journal of Purchasing and Supply Management — supplier resistance to automated bidding tools rising with market concentration.
Kraljic, P., "Purchasing Must Become Supply Management," Harvard Business Review (1983).

The Pactum/Walmart figures are reported via two independent secondary sources rather than a primary company disclosure located directly; both cited sources describe the same scope (tail-spend, high-volume) and the same figures, which is noted rather than treated as independent confirmation of the underlying deployment.