Let the Agent Negotiate. Don't Let It Set the "Walk-Away"

There is a pitch doing the rounds that goes like this. Procurement has always been labor-intensive. AI agents can now find suppliers, talk to them, negotiate with them, and over time learn what cost and quality thresholds are acceptable. Humans set the parameters and provide oversight. The procurement team does more with the same people.

Most of that I agree with, and I have lived the reason it matters. For the last five years I built procurement functions from scratch in regulated fintech, with spend and supplier complexity growing every quarter and no extra headcount coming. The AI augmentation I use today started because I had no other way to keep up.

But one line in that pitch should make any procurement leader stop: the agent "ultimately decides" what the acceptable thresholds are, learning over time how far it can push.

That is not a task. That is the job.

What agents are genuinely good at

Be honest about where the labor actually goes in procurement. Very little of it is the negotiation itself. Most of it is:

• Finding and qualifying suppliers. Long lists, capability checks, first-pass screening, chasing documents.

• The tail spend. Hundreds of low-value, low-risk purchases that nobody has time to negotiate, so nobody does.

• Preparation. Pulling past prices, previous terms, performance history and market benchmarks before a conversation.

• Routine back-and-forth. Clarifying specs, confirming delivery windows, requesting revised quotes.

• Monitoring. Certificates expiring, price indices moving, renewal dates approaching.

Agents can do all of this well, today. Tail spend in particular is where the case is strongest: it is spend where the alternative is not a human negotiator, it is no negotiation at all. Even a mediocre agent beats that.

Where the pitch goes wrong

The problem is the idea that an agent learns acceptable thresholds by pushing boundaries and seeing what happens.

Thresholds are business decisions, not patterns in the data. What price is acceptable for a critical supplier depends on switching cost, supply risk, the relationship, what else you are negotiating with them this year, and what the business can actually absorb. None of that is visible to an agent looking at past deals.

Learning by pushing means learning on your suppliers. Every boundary the agent tests is a real conversation with a real counterparty. A supplier that gets squeezed by software on a low-value order remembers it when you need them on a high-value one. Relationship damage does not show up in the savings report.

Past outcomes are a biased teacher. If the agent learns from what was "accepted" before, it learns your historical mistakes along with your wins. Maverick spend, deals done under time pressure, the contract everyone knew was overpriced but nobody reopened. Garbage in the history becomes policy in the agent.

Quality thresholds are rarely written down. In most organisations, "acceptable quality" lives in the heads of the requester and the category manager. An agent inferring it from outcomes will get it wrong in the cases that matter most: the new supplier, the changed spec, the safety-relevant part.

Accountability does not delegate. In a regulated environment, someone signs off on supplier selection and commercial terms, and a regulator or auditor can ask them why. "The agent had learned that was acceptable" is not an answer I would want to give.

The model that works: humans set the corridor and agents work inside it

The useful framing is not "human oversight" as a vague safety net. It is a clear split of authority.

Humans defines the category:

• Target price, walk-away price, and the conditions for escalation

• Minimum quality and compliance requirements, written down

• Which suppliers the agent can approach and which relationships it must not touch

• What the agent can commit to on its own, by value and risk

  
  

Agents execute within that corridor:

• Sourcing, screening and shortlisting

• Prepare the negotiation pack from history and benchmarks

• Run the routine rounds on low-risk spend

• Escalate the moment a deal approaches the edge of the corridor, rather than testing it

And the corridor moves only by human decision. The agent can recommend: "I have seen three suppliers accept 4% below target this quarter, consider tightening." A person decides whether to act on it. Learning feeds the human, not the threshold.

The knowledge question, again

The pitch mentions agents drawing on past results and what you know about your suppliers. That is right, and it is also where most deployments will quietly fail.

That knowledge sits in contracts, supplier correspondence, performance reviews and the heads of experienced category managers. If it is scattered, the agent negotiates blind. If it sits inside a vendor's platform, your negotiating memory belongs to someone else, and your counterparties may well be using the same vendor.

This is the same argument I keep making. The agent is only as good as the context it reasons over, and that context should be yours: owned, in one place, in your environment.

Before you deploy a procurement agent

1. Start with tail spend. Low risk, high volume, and the baseline is no negotiation at all.

2. Write the corridor down per category. If you cannot state the walk-away price and minimum quality, you are not ready to delegate the conversation.

3. Set commitment limits in the system, not the prompt. The agent should be technically unable to sign above its limit, not just told not to.

4. Tell suppliers when they are talking to an agent. It protects the relationship and, increasingly, it is expected.

5. Log every round. Offers, counter-offers, reasoning. Audit will ask, and so will you when a deal goes wrong.

6. Review what the agent learned, monthly. Its recommendations to move the corridor are some of the best market intelligence you will get. Treat them as input, not instruction.

AI agents will make procurement teams far more capable. I have seen it in my own work. But the value of a procurement professional was never in sending the emails. It was in knowing what a good deal is for this business, with this supplier, right now. Keep that part.