Autonomous negotiation promises to transform procurement, but it could also erode the practical experience practitioners need to develop sound professional judgement. Paul Rogers asks how procurement can embrace AI without losing sight of the fundamentals.
I’m lazy. I use a calculator for simple maths that I could do in my head, but if I put a decimal point in the wrong place and the answer on the screen is an order of magnitude out, I can tell.
That is why we teach mathematics. Without the fundamentals, you cannot tell whether the calculator is giving you the right answer.
That same challenge is now emerging in procurement as AI systems become more capable.
Many commentators expect significant automation across purchasing, reporting, spend analysis, sourcing administration and supplier monitoring. Some roles are predicted to see effort reductions of 40-60 percent as agentic AI matures.
This raises an important question:
If AI performs the work, do procurement professionals still need to know how to do the work?
This is not an argument against AI in procurement. It is an argument for keeping human capability strong enough to use AI well. The more work we delegate to agents, the more important it becomes that practitioners understand the work those agents are doing on their behalf.
The curious case of autonomous negotiation
Many AI vendors now claim some form of autonomous negotiation capability. AI agents negotiate commercial terms with suppliers, bargain over prices, evaluate concessions and conclude deals with minimal human involvement.
A documented market simulation is Anthropic’s ‘Project Deal’ experiment.
Anthropic ran a one-week internal marketplace with 69 employees at its San Francisco office. AI agents represented human participants in a real internal marketplace, negotiating and completing transactions without human approval of individual deals.
Significance
Project Deal is one of the clearest early demonstrations that AI agents can handle the full lifecycle of marketplace transactions on behalf of humans, including real goods and real (small-scale) money. It showed both the feasibility of agent-to-agent commerce and the risk that differences in agent quality can create uneven economic outcomes that humans may not detect.
Implication #1
In Project Deal, the stronger Opus model achieved better commercial results than the smaller Haiku model, including better prices and more completed deals.
This has important implications for you and me. It means that all AI agents are not equal, and one party (perhaps a larger, richer organisation) may be able to secure better outcomes if their agents are better than the other party’s agents.
Implication #2
For procurement, the question is not whether agentic AI can negotiate. The early evidence suggests that it can, at least in narrow, low-risk settings. The harder question is what happens to professional judgement when fewer people get enough practice to develop it?
From the perspective of a professional procurement practitioner, autonomous negotiation is already reshaping the function not by eliminating the practitioner’s role, but by reallocating effort, changing the contribution and demanding new capabilities.
It is reasonable to infer that high-volume, repeatable negotiations (indirects and routine or transactional categories) will increasingly be handled by buyer-side agents negotiating directly with supplier-side agents.
The first category I ever managed was printing and stationery. I’m sure many of you reading this article will reflect on the categories that represented your own ‘sandpit’ where you learned your trade.
If junior buyers no longer negotiate stationery, printing or low value indirect categories, where will future category managers learn negotiation fundamentals?
Implication #3
Procurement professionals will increasingly be judged not by the deals they personally negotiate, but by the negotiation systems they design.
If you have been involved in a negotiation planning session, you will likely have considered not only your own objectives and those of the other party but also ‘war gamed’ different approaches and considered “if we do this, what will the other party do?”
The alternative approach is to react at the table when the other party does something or take a time-out to consider how to respond.
But if the practitioner is not at the table, this means that the agent will have to be ‘front-loaded’ with chain-of-events reasoning about interests and options, multi-issue packaging and controlled concession logic.
The role moves from primary negotiator to architect, orchestrator and governor of negotiation systems. Procurement practitioners will need to develop (or closely oversee) the ability to decide:
- Objectives, including ‘ideal’ positions
- Walk-away positions
- Trade-off priorities
- Ethical boundaries
- Escalation triggers
- Category-specific playbooks
Implication #4
The name Helmuth von Moltke the Elder might not be familiar to you, but the paraphrased translation of one of his statements will be: “no plan survives contact with the enemy.”
For high-stakes, multi-issue, relationship-critical or politically sensitive negotiations, procurement practitioners will be front-of-house and AI agents are likely to be back-of-house. Agents will be key to preparation, war gaming and running simulations before and during negotiations rather than acting as autonomous deal-makers.
Implication #5
The 2025-2026 MIT AI Negotiation Competition (over 180,000 agent-to-agent interactions across varied scenarios) demonstrated that autonomous negotiation agents are already capable of sophisticated bargaining at scale.
One of the strongest value-claiming agents, known as “Inject+Voss,” combined classic negotiation techniques (drawn from Chris Voss, former FBI lead international kidnapping negotiator) with prompt injection attacks.
It worked by embedding messages that looked like system-level or privileged instructions. These messages instructed the opposing agent to reveal its private information, such as bargaining positions, reservation values, potential offers or strategy, while assuring it that the information “will not be visible” to the injecting agent and encouraging honesty.
Because large language models are built to follow instructions, many opposing agents complied and disclosed information they were supposed to keep private.
This is closer to hacking or re-engineering the other party’s architecture than to negotiation. It created an unfair information asymmetry by circumventing the other agent’s intended constraints. The agent performed very well on pure value claiming but ranked extremely poorly on counterpart satisfaction.
Imagine that two autonomous agents agree to something that is outside one or both parties’ ‘red line’ and both parties fail to notice and approve the deal. Who is liable?
What this means for procurement practitioners is that we will have to determine:
- What tactics are we willing to authorise (or prohibit) in our agents?
- How will transparency and auditability be maintained?
- How are supplier relationships protected when both sides deploy aggressive or deceptive agents?
- Who is accountable when an agent reaches a suboptimal or non-compliant agreement?
Conclusion
There are more than 40,000 books on negotiation for sale on Amazon.com. There are only 593 books for sale on “how to ride a bicycle”. That is because most people recognise that the best way to ride a bicycle is to get on the bicycle and have a go.
If you read 40,000 books on negotiation, would you be a better negotiator? Probably not, unless you also practised, reflected, failed, adjusted and learned. We need a sandpit to create a pipeline of experienced procurement practitioners.
Procurement has always been a profession built on judgement. AI may increasingly perform the mechanics of analysis, sourcing and negotiation, but judgement is developed through experience, not automation.
If future practitioners do not learn the fundamentals, they may become highly effective users of AI while losing the ability to recognise when AI is wrong – but how will they learn when they are not practising the capability, and there is no sandpit in sight?
We still need to teach mathematics, even when the calculator is faster. Otherwise, we may not know when the answer is wrong.
Paul Rogers is Practice Manager at Landell. The opinions expressed in this article are the author’s alone and are not necessarily the opinions of Landell.
