For a while, the big pitch around AI in procurement was: "You can ask it questions." Need a quick market summary? Ask the bot. Want to pull together supplier data faster? Ask the bot. It was useful, in the way a well-organised search engine is useful. But it was still completely dependent on a human deciding what to ask, and when. The AI didn't care whether you showed up or not.
Agentic AI in procurement is the part most people in this space are still underestimating. The shift is about initiative more than capability.
AI agents don't wait to be asked. You assign a task, and they go. They make decisions, take actions, and return results, often without a human touching the process at all. This is a different model, not an incremental upgrade on what came before.
The Difference Between a Chatbot and an Agent (And Why It Matters)
The chatbot framing made a lot of sense in 2022 and 2023. You had a question, the AI had an answer. Interaction was the whole point. Think of it as a very fast, very well-read colleague who would only speak when spoken to.
Agentic AI flips that dynamic. Instead of responding to inputs, agents operate on objectives. You give them a goal (find me three qualified suppliers for this category, ESG-compliant, sub-M annual revenue) and they go figure out how to accomplish it. They don't sit in a chat window waiting. They work.
The implications for how procurement teams are structured are significant. If your value-add as a category manager is knowing which questions to ask and when, that's an increasingly narrow competitive position. The higher ground is in knowing what the right answer looks like, which is a judgment call agents still can't make reliably. More on that in a moment.
What Autonomous Sourcing Actually Looks Like Right Now
I want to get specific here, because "autonomous AI" is the kind of phrase that gets thrown around until it means nothing.
- Supplier discovery without a human in the loop. This is already happening. The agent identifies the category need, goes out and scans potential suppliers across markets (often tens of thousands of company profiles), and returns a shortlist with reasoning attached. No one kicked off the search. The system identified the need from spend data or contract triggers and acted on it.
- ESG vetting at scale. Agents are now capable of running that vetting autonomously: environmental compliance, labour practice flags, governance concerns. The outputs still need review for high-stakes decisions, but the speed and coverage are beyond what humans can match.
- Initiating RFPs without a human prompt. The agent identifies a sourcing need (based on contract expiry, demand signals, price variance), qualifies a supplier pool, and fires off the first round of an RFP. The category manager steps in when what's needed is judgment, negotiation, and relationship management.
Why This Fixes Something Real
The procurement function has always had a resourcing problem. The ratio of spend under management to headcount is brutal in most organisations, and the high-value strategic work (category strategy, supplier relationship development, risk management) gets crowded out by operational work that should have been automated years ago.
The version of procurement that agents enable is the same people doing less of the low-value work so they can do more of the high-value work, not fewer people doing more of the same thing. That sounds like a corporate talking point, but the underlying logic is sound. If your senior category managers are spending 40% of their time on administrative sourcing tasks, that's a process problem, not a people problem. Agents fix the process.
The 2026 Art of Procurement industry survey found that nine out of ten procurement leaders are either implementing AI agents or actively planning to. That's a remarkable number, and it reflects genuine momentum rather than just interest.
The Adoption Challenges Nobody Talks About Honestly
The vendor pitch tends to skip the hard parts. Here's what it usually leaves out.
Data readiness is a real blocker, not a footnote. Agents are only as good as the supplier data, contract repositories, and category taxonomies they're working from. A lot of organisations are sitting on years of inconsistent, siloed, poorly classified data that would make any autonomous process unreliable. You can't automate your way out of bad data. Before you can get agents to work well, you often need a data cleanup project that's unglamorous and time-consuming. That's not a reason to avoid starting. It's a reason to be realistic about the timeline and sequence.
The governance question is genuinely unresolved. When an agent initiates an RFP, or shortlists a supplier, who is accountable for that decision? This isn't rhetorical. Legal, compliance, and procurement leadership are going to have to work through it together, and in most organisations they haven't yet. The organisations deploying autonomous procurement well aren't turning agents loose. They're deploying autonomy in narrow, well-structured areas where the rules are clear, the data is clean, and the outcomes are measurable.
Change management is the underrated problem. Procurement professionals who've built careers on supplier relationships and deep category expertise aren't going to hand that over to an agent, nor should they entirely. The case for what agents are for (freeing up strategic capacity) needs to be made clearly and consistently, from leadership down.
What the Human-Agent Split Actually Looks Like
This is the question I find most interesting, and the one that has the least consensus right now.
Based on what's being built and deployed, the emerging model looks something like this: agents own the transactional layer (discovery, qualification, compliance checks, standard event management, bid collection). Humans own the strategic layer (category positioning, supplier relationship investment, negotiation, risk judgment on non-standard situations).
What doesn't fit cleanly into either bucket is the interpretive middle. When an agent returns a shortlist and the category manager looks at it and thinks "that doesn't feel right," what happens then? That intuition is often based on relationship context, market knowledge, or pattern recognition that isn't in the data the agent used. Building workflows that capture that human override gracefully, without just making agents pointless, is genuinely hard. The organisations getting this right are the ones treating it as a workflow design problem, not a technology problem.
The organisations that are struggling are the ones that either (a) treat agents as a replacement for human judgment wholesale, or (b) add agents to their process without redesigning the process around them, so the agents just become an extra step nobody trusts.
Where This Is Going
Agentic AI in procurement is early-stage in most organisations and moving fast. The capability curve is ahead of the adoption curve, which means there's a real window right now for procurement teams that move thoughtfully: not recklessly, but not cautiously either.
The better question isn't "should we use agents?" It's: what's the right first task to hand off, given where our data is, what our governance model looks like, and where our team's time is most wasted right now? That's a procurement problem more than a technology problem. And it turns out procurement people are pretty good at those.
Frequently Asked Questions
What's your experience with this? If you're already running agents in your procurement function, or you've hit a wall trying to get there, we'd genuinely like to hear what that's looked like.