Few subjects get as much attention as AI in negotiation. As most of the discussion seems to be about models and capabilities, this article looks elsewhere. It sets out three organisational dimensions that matter just as much: governance (the system), design (the scope) and organisation (the people).

Each takes into account the everyday negotiation complexity of organisations, and together they point to the conditions under which autonomous negotiation could actually work.

Governance (The System)

Let us start with data protection and system architecture. Suppose two companies run an entire negotiation through AI. The first question is not which model each side uses, but on which infrastructure the negotiation actually happens. Two language models negotiating directly with each other? A bot tied into a company's systems? A third-party or open-source platform? Or, in footballing terms: in which stadium is the match being played, who owns the stadium, and which party sets the rules? The system question is relevant not just from a data protection perspective, but also relates to governance, liability, auditability and resistance to manipulation.

Even if you set aside the EU AI Act, the GDPR and other data-protection regulations, this would require both sides to trust the same platform enough to enter proprietary company data before the actual negotiation.

And even if such a system existed, the questions follow immediately. Which data actually goes in, at what quality and to which standard? What happens if one side is unhappy with the result? That alone could, ironically, lead to a substantial negotiation in its own right. AI is meant to take the negotiating off your hands, yet using it may first require another negotiation. Given how sensitive that data is, the risk of a leak also looms over the negotiation process.

While debating autonomous negotiating agents, a significant share of companies still do not have an AI policy in place. If there is one, the question is whether it is actually holding up. While some organisations allow only tightly controlled enterprise tools; others built a constrained in-house system early on. The result – at worst – may be the same: people fall back on shadow AI, because they already have a better tool in their pocket. In such a scenario, autonomous negotiation is still a long way off.

Design (The Scope)

Suppose the system question is settled. The next one, however, is just as fundamental: how do both sides pin down what is on the negotiating table and what is not? In other words, how is the scope set? For an autonomous negotiation to work, even a seemingly “simple price conversation”, the scope has to be clear. Yet experienced negotiators know that the agreements that hold often come about precisely when something enters the conversation that no one had on the table to begin with. What starts as a five per cent discount can turn into value through payment terms, volumes, exclusivity, service levels or joint investment. That is value creation in its purest form, and it depends on uncovering the real interests and motives behind the stated positions.

Drawing on Michael Wheeler’s work, one of the central dilemmas in negotiation is this: reveal nothing and you cannot be exploited; reveal nothing and you close no deal. In the context of AI, that is a paradox. An autonomous negotiation needs a scope that is fixed and unambiguous, yet in demanding negotiations the decisive move is often to shift that very scope by creating something new entirely.

Recent studies of AI negotiation point to exactly this weakness. Models struggle to manage the tension between cooperation and competition, between creating value and claiming value. Whether future AI will handle that dilemma better is worth watching closely, particularly in complex, protracted negotiations. For now, it still makes too many mistakes. That said, the weaknesses of today's models are no argument for humans either. Human negotiators show many of the same mistakes.

Even so, autonomous negotiation has been gaining ground for some years and looks set to grow further. At Coupa Inspire 2026, the conference run by the software provider Coupa, AstraZeneca showed how agentic AI can work within a defined governance framework without constant human oversight.

Procurement is where there seems to be most movement, and a considerable chunk of it happens in the long tail. In that balance of power, a tail supplier will accept a large buyer's rules. Interestingly, agentic AI can not only replace negotiations people used to run themselves; it also automates the ones that were never worth running, because the single transaction was simply too small. Hence, it also comes down to transaction volume against transaction value.

Organisation (The People)

The third dimension is the human one. Any technology is only as good as the person using it. In that sense, AI is an amplifier. Anyone who has negotiated for years without a clear process will, with AI, simply negotiate without a process faster.

Hypothetically, even with the system and scope questions settled, the question remains: what is left for the human to do? Perhaps the role shifts from negotiating to structuring and overseeing the entire process. AI does the running, but what gets negotiated, to what end and within which limits looks set to stay human work for the foreseeable future. A prompt is quickly put together. An understanding of the negotiation process is not.

Moreover, there is another factor that many AI discussions underrate: not every workforce comes along for the ride. Resistance to AI roll-outs is a reality in plenty of organisations.

And even if every one of these hurdles were to be cleared, a last, almost philosophical question remains. Is a good negotiated outcome only an outcome, or also a deeply human social process, shaped by friction, by coming closer, and ultimately by mutual respect?

Conclusion

The debate about autonomous negotiation tends to start with the models. In practice it is decided by governance, design and organisation. Today's models can already process vast amounts of information and generate plausible options. Their limits emerge where uncertainty is not merely informational, but social and strategic: where trust, organisational politics and unspoken interests shape the outcome. And that is precisely where complex negotiations are often decided. This is why agentic AI will most likely take hold first where transaction volumes are high, individual transaction value is comparatively low, and the room for negotiation is clearly defined.