# What AI Cannot Negotiate
> AI can price, source, score, and draft contracts. It cannot underwrite a relationship, hold an unspoken promise, or absorb reputational risk. A taxonomy from inside the function being automated.
**Author:** Maximilian Groh
**Published:** 2026-05-25
**Updated:** 2026-05-25
**Canonical URL:** https://drgroh.com/blog/what-ai-cannot-negotiate/
**Tags:** Procurement, AI, Negotiation, Leadership
---Last quarter I watched two algorithms set a price between them in under 200 milliseconds.

Neither of them paused. Neither of them asked what the relationship was worth. Neither of them noticed that the supplier on the other end was twelve months from insolvency — which would have cost us more in a year than we saved in a second.

The machines were faster than us. They were not wiser. And somewhere in the gap between the two, there is a job that does not get automated, no matter how much the consultants insist it will.

## Three things the machine did not see

The standard conversation about AI in procurement is a conversation about what gets automated. It is a useful conversation. It is also the easier half of the question.

The harder half is what does not. Not as a comforting story about human value, but as a precise inventory of work that has features algorithms are structurally not built to handle.

That 200-millisecond price missed three things at once.

It missed **counterparty fragility**. The supplier's financials looked fine on a quarterly lag. The signal that mattered was earlier and softer: a pattern of late payments to their own suppliers, a quiet departure of two senior engineers, a tone shift in their last quarterly call. A human reading those signals together would have flagged the deal. The algorithm read each one in isolation, scored none of them as critical, and cleared the trade.

It missed **relational debt**. We owed this supplier something — not contractually, but in the way commercial relationships actually work. They had absorbed a quality issue on our side eighteen months earlier without escalation. The price the algorithms agreed on was technically fair and relationally cold. The model had no field for that.

It missed **reputational exposure**. The supplier sat at the edge of a category that had been under regulatory scrutiny in two of our markets. A failure here would have produced a story we did not want to be in. The algorithm did not know what stories we did not want to be in. There is no API for that.

These were not failures of the model. They were features of the work the model was not built to do.

## This is not nostalgia

The argument here is not that humans are better than algorithms. In a great many procurement decisions, they are not. The price models are sharper than most buyers. The contract reviewers catch clauses tired humans miss. The spend-analysis engines see patterns no category manager will ever see manually.

[That is a large and growing share of the work](https://drgroh.com/blog/the-70-20-10-rule-of-procurement-ai/), and pretending otherwise would be dishonest.

The argument is narrower and more interesting. Certain decisions have features algorithms structurally cannot weigh — not because the technology is immature, but because the features themselves resist the kind of representation algorithms need to operate. They are not slow human decisions waiting to be sped up. They are different decisions entirely.

Saying this is not a defense of the function. It is a diagnosis.

## A taxonomy of un-automatable procurement work

Four categories. Each one is concrete, each one is demonstrable inside any reasonably complex procurement portfolio.

### Relational underwriting

The work of taking a position on a counterparty that goes beyond what their financials say. Reading a supplier the way a credit officer reads a borrower — not as a snapshot, but as a trajectory. Noticing when a relationship is healthier or sicker than the data shows.

[Counterparty fragility](https://drgroh.com/blog/seeing-risk-before-it-breaks-you/) is the canonical example. So is the inverse: identifying a supplier whose balance sheet looks weak but whose operational competence and management quality are stronger than the numbers suggest. Algorithms run scorecards. Humans underwrite.

### Narrative judgment

The ability to know what story a decision will become. Every significant procurement decision is, retrospectively, a story — told in audit committees, in supplier reviews, in tribunals when something goes wrong, in onboarding materials for the next category manager.

Senior buyers learn to think two moves ahead of the narrative. They make decisions that will hold up not only in the spreadsheet, but in the room where the decision gets retold. Algorithms optimize for the spreadsheet. They have no concept of the room.

### Ambiguity arbitrage

The work that creates value precisely because the situation is under-specified. A supplier asking for a price adjustment they cannot formally justify but have a real basis for. A contract clause that two parties read differently and both prefer to keep ambiguous. A category strategy that depends on a market signal no one has yet named.

Algorithms hate ambiguity. They require the situation to be reducible to features and weights. Some of the most valuable work in procurement happens in moments where reducing the situation that way would destroy the value. A buyer who can sit inside ambiguity longer than the counterparty can extracts terms a model would have closed prematurely.

### Ethical authorship

The work of deciding, on behalf of a firm, what it will and will not do. The supplier you do not onboard even though the price is right. The clause you insist on even though the legal risk is contained. The country you exit before regulation arrives. The contract you sign with a supplier whose conditions do not yet meet your standard, because walking away would be worse for the workers than staying engaged.

These are not optimization problems. They are authorship problems. They require someone to put their name on a position the firm will own afterward. Algorithms do not put their names on anything. That is the point.

## What this means for procurement leaders in 2026

If this taxonomy is roughly right, two things follow for how a serious procurement organization invests in its people over the next eighteen months.

**Hire for the un-automatable.** The category manager who was best at the work AI now does is, structurally, the wrong profile for the next decade. The category manager who is best at relational underwriting, narrative judgment, ambiguity arbitrage, and ethical authorship is the right one. These are not entry-level skills, and they are not trained in a week. They are [developed over years](https://drgroh.com/blog/the-procurement-skills-roadmap-for-the-ai-era-2026-2030/), often by working alongside someone who already has them.

**Stop measuring what the machine already does.** A procurement function that still spends most of its performance reviews on cycle time, savings against benchmark, and contract turnaround is grading its people on the parts of their work the machine is about to take. The metrics need to migrate upward, toward the work that will define the function five years from now. This is hard. The new metrics are softer. They are also more honest.

There is a third implication, and it is the one most leaders avoid. A procurement organization built around the un-automatable looks different. It is smaller in the operational layers, but more senior on average. It pays more per head, not less. It hires later from adjacent disciplines — credit risk, journalism, law, ethics — and trains the procurement craft into them rather than the reverse.

That is a different operating model. Most functions are not ready to admit it. The ones that will lead in 2026 are the ones that already are.

## Key Takeaways

- AI can price, source, score, and draft contracts at high quality. It cannot perform four categories of work that often determine whether a procurement decision is a good one: relational underwriting, narrative judgment, ambiguity arbitrage, and ethical authorship.
- **Relational underwriting** is reading a counterparty as a trajectory rather than a snapshot — the kind of pattern recognition algorithms run feature-by-feature and miss in combination.
- **Narrative judgment** is the ability to anticipate the story a decision will become, in audit committees, supplier reviews, and the room where it gets retold.
- **Ambiguity arbitrage** is the work that creates value precisely because the situation resists clean specification.
- **Ethical authorship** is the act of putting a name on a position the firm will own afterward. Algorithms do not put their names on anything.
- The risk of fully algorithmic procurement is not that the machine makes wrong decisions. It is that it makes decisions whose costs only show up later, in dimensions the model was not built to weigh.
- The procurement leaders who win the next decade will hire for the un-automatable, retire metrics that grade people on work the machine now does, and build smaller, more senior, more cross-disciplinary teams.

## The negotiation that never happened

A different deal, six months later.

The algorithms had cleared a sourcing decision. The numbers were good. The supplier scorecard was clean. The contract was ready to sign.

A senior buyer asked for twenty-four hours. Not because she could articulate, in the moment, what was wrong — but because something in the supplier's recent communication had shifted in a way she could not yet name. The next morning, the supplier withdrew. A larger competitor had acquired them overnight, and the terms we had been about to sign would have been honored for a quarter and then quietly renegotiated against us for years.

The algorithms could not have known. They were not slow. They were unwitnessed.

The negotiation that never happened was the most valuable negotiation of her year. There will be no line for it in her performance review.

That is exactly the work that does not get automated. And in 2026, [recognizing it](https://drgroh.com/blog/modern-procurement-leadership/) — naming it, protecting it, paying for it — is the difference between a procurement organization that gets faster and one that gets wiser.

The two are not the same.