Guide

AI agents in procurement: what they do, what to keep human, and how to evaluate them

By the EvaluationsHub team. Last reviewed 8 November 2026.

What an AI agent is, and what it is not

Definitions differ, so it helps to use the plainest ones. Anthropic separates workflows, where code paths are fixed in advance, from agents, where the model decides its own steps and tool use, and advises starting with the simplest thing that works, which may not be an agent at all. IBM distinguishes assistants, which respond when asked, from agents, which plan and act toward a goal, and notes that human oversight still matters. Google makes a similar split between agents, assistants and rule-following bots.

For buying purposes the label matters less than three questions. What can it do without asking? What does it read? What does it write to? A system that drafts a brief and waits for approval carries a very different risk from one that sends messages to suppliers by itself.

Where agents fit in procurement

Autonomy is a dial, not a switch

This is our own framing, not a standard. Think of five settings, and choose one per task, not one per product.

1. Suggest. The agent shows a finding. A person does everything else.
2. Draft. The agent produces the document, message or brief. A person edits and sends.
3. Act with approval. The agent prepares the action and executes it once a person approves.
4. Act and report. The agent acts within limits and reports afterwards. Suitable for low-risk, reversible steps such as reminders.
5. Act on its own. No review. We would not put anything with commercial or compliance consequences here.

Move a task up a setting only when the record shows the drafts at the current setting were right, and keep a log you can audit.

What the evidence says about adoption

We found no independent study with a named sample that measures AI agents in procurement specifically, so treat any vendor figure on that as marketing. What exists is broader, and it points the same way.

• McKinsey’s State of AI 2025 (1,993 respondents in 105 countries) found 62% of organisations at least experimenting with AI agents, and no more than 10% scaling them in any single function.
• A Gartner survey of 101 chief procurement officers, published 19 May 2026, found 36% very confident in their ability to redesign roles and processes around AI.
• A later Gartner release, 8 October 2026, reported 25% of senior procurement leaders had redesigned jobs as a direct result of AI. We read this through trade press, because Gartner’s page would not load.
• The Hackett Group’s 2026 Procurement Key Issues Study reports 43% of organisations actively pursuing AI deployment and 12% at large scale. Hackett does not state the sample on the release, and it sells procurement consulting.
• Gartner has forecast that more than 40% of agentic AI projects could be cancelled by the end of 2027 on cost, unclear value or weak risk controls. That is a prediction, based on a webinar poll, not a measurement.

Risks, and the controls that matter

OWASP’s 2025 list of risks for applications built on large language models includes three that matter for procurement agents. Prompt injection means content the agent reads, such as a supplier document, can change what it does. Excessive agency means the agent has more functions, permissions or autonomy than the task needs. Misinformation, including overreliance, means people trust a fluent answer that is wrong. OWASP’s mitigations are plain: least-privilege access, validation of outputs, separating untrusted content, and human approval for high-impact actions. OWASP also published a separate top ten for agentic applications on 9 December 2025.

Practical controls for a buyer: every claim the agent makes cites the document it came from; the agent can read what it needs and write only where it must; actions that reach a supplier or change a record need approval; and every decision is logged with who approved it.

The EU AI Act in brief

We keep this short, because we already publish a complete guide to the EU AI Act for procurement. This is not legal advice, and the position below is as of 8 November 2026.

• The AI Omnibus, Regulation (EU) 2026/1744, is adopted. It entered into force on 27 July 2026 and moved the high-risk application dates to 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for product-embedded ones.
• The transparency rules in Article 50 apply from 2 August 2026. The Commission’s FAQ lists AI agents among the systems that must tell people they are dealing with an AI when they interact with them directly. Systems working only in the background fall outside that duty.
• The AI literacy duty in Article 4 has applied since 2 February 2025, and the Omnibus softened its wording to require measures to support literacy, not a guaranteed level.
• Whether a procurement agent is high-risk: not stated in any primary source we read. Annex III as published does not list procurement or supplier management. An agent that scores or monitors the performance of people could touch the employment area, so check that case with counsel.

Governance frameworks worth knowing

The NIST AI Risk Management Framework 1.0 is voluntary and organised around four functions: govern, map, measure and manage. It asks you to document human oversight and who holds it. NIST says a revision is under way, and we did not find a published one. ISO/IEC 42001:2023 sets requirements for an AI management system inside an organisation that builds or uses AI. Neither is specific to procurement, and neither is a substitute for deciding what each agent may do.

How to evaluate an agent in a demo

• Ask what it can do without approval, and have them show you the setting that controls it.
• Ask to see the source behind a finding, not just the finding.
• Ask what happens when a supplier document contains instructions.
• Ask for the audit trail of one decision, with the approver.
• Ask what is live today and what is on a roadmap, with a date.

For feature-level questions, see how to compare AI features in SRM software.

What Eva does in EvaluationsHub today

Eva AI works in the background and finds things on its own. It reads each contract, checks the obligations against live supplier performance and risk scores, and flags drift without being asked. Ahead of a supplier renewal it prepares a negotiation brief from the performance record, again without being asked. The guardrails are human. Eva recommends, and a person approves, edits or dismisses each action before anything is executed or reaches a supplier, with the reasoning and an audit log kept for every decision. On the dial above, Eva finds and prepares on its own and executes only at “act with approval”. It does not negotiate on your behalf. Read more on the Eva AI page.

What we cannot tell you

• How any other vendor’s agents behave behind the login. Ask for the demo.
• Your legal position under the AI Act. That depends on what the agent does and who it affects.
• A reliable adoption rate for agents in procurement. We could not find an independent one.