Four things, and they are the same four in every serious governance framework: award business, commit money or contractual terms, release drawings or other controlled information, and resolve an exception it was never scoped for. Each is irreversible and visible outside the company. An agent can prepare the recommendation, the draft commitment or the redacted package; the act itself stays with a person.
Why This Matters
Agentic AI is the first layer of procurement software that can be wrong in public. A generative model that drafts a poor request wastes a buyer's time; an agent that sends it wastes a supplier's, and suppliers remember. The four actions above are the ones whose consequences cannot be recalled by a follow-up email, which is why the line sits there rather than at some measure of how capable the agent has become.
How It Works
Award. The comparison ends in a ranked recommendation with reasoning. A person makes the award, and for public-sector and regulated buyers the audit trail has to show that person doing so.
Commitment. Prices, quantities, lead times and payment terms bind the company. An agent may draft the confirmation; a person sends it.
Controlled information. Drawings, specifications, price lists and anything covered by an NDA or export control leave your control the moment they are sent. Deciding what goes out, and to whom, is a human decision every time. The redaction choices are covered in when to transform and when to redact an RFQ.
Unscoped exceptions. An agent meeting a situation nobody anticipated is operating without a specification. The correct behaviour is to stop and escalate with the context attached, and a pilot should test that it does before its scope is widened.
Everything else in the RFQ cycle, from supplier research through drafting, sending under a template, follow-up, reply triage, quote validation and normalization, is reversible or internal and can be handed over in stages. The full mapping of actions to gates is in what to let an AI agent do in procurement.
FAQ
Can an AI agent negotiate with a supplier?
It can draft counter-proposals and prepare a negotiation brief, and some products run autonomous negotiation on low-stakes, high-volume terms. Committing to the outcome is a contractual act and stays with a person. Treat autonomous negotiation as ask-before-send at most until it has a measured record.
Can an agent reply to a supplier's clarification question?
Yes, when the answer is in the original brief. That is reversible in substance and confined to one thread. If the question requires a decision the brief does not contain, the agent should escalate rather than improvise.
Does this list change as agents improve?
The reversible-work boundary moves as measured performance improves; these four do not, because the reason they stay human is the consequence of error, which is unaffected by how rarely the error occurs.
People also search for: