Why most procurement teams haven't started with AI agents
If you spend enough time on LinkedIn, you could believe every procurement team already has a few AI agents on the payroll. My feed is full of agent demos, agent frameworks and people describing what their agents did this week. It's easy to forget that this is a bubble. Most of the people in it build, sell or teach AI for a living, and that includes me.
The bubble bursts quickly when we leave it. When we sit down with procurement leaders who are prospects, I'm reminded how far most organisations are from what my feed suggests. For many of them, AI still means a chat window someone uses to tidy up an email, and "agents" sounds like either science fiction or the next vendor pitch. I don't say that as a criticism. These people are running procurement functions, and nobody has explained to them, in procurement terms, what an agent would actually do on an ordinary Tuesday.
This article is my attempt to write down what I try to explain in those meetings. My view is simple. Procurement needs AI agents because most of its week is follow-through work that previous generations of software could not do, and agents can now do a useful share of that work under supervision.
Where a procurement team's week actually goes
Ask someone who has worked in procurement for a few years what takes up their week, and they rarely say negotiating. They say chasing approvals. Matching an invoice that's off by a few hundred because the supplier billed for twelve units and the warehouse logged ten. Emailing a vendor, again, because their insurance certificate expired in March. Replying to the fourth "any update on my PO?" message before lunch.
Procurement teams are hired to help the company spend money well, yet most days go on making sure money gets spent at all. Much of that time is follow-through: reading a document, checking it against another one, making a small decision and nudging the next person. Software has always been poor at this kind of work. In my experience it has only recently become good enough to trust with some of it.
Why e-procurement suites and shared services didn't shrink the workload
Procurement headcount almost never grows at the pace of spend. The company opens a new plant, enters two new markets and picks up a few hundred suppliers along the way, and the team gets one extra analyst if it argues well at budget time.
The tools that were supposed to absorb that growth mostly moved paper onto screens. E-procurement suites and purchase-to-pay platforms gave us cleaner forms, approval routing and dashboards. Those were real improvements. But a workflow tool routes a task to a person; it doesn't do the task. Someone still has to open the request, work out whether there's an existing contract, notice that the cost centre is wrong and write to the requester about it.
I have also spent a good part of my career on the other big lever, moving procurement operations into global business services centres. That lowered the cost of each transaction. It didn't reduce the number of times a person had to open a request and decide what to do with it. Between the software and the shared services model, the number of small decisions per request barely moved.
What an AI agent does that a workflow tool can't
An AI agent is software you give a goal to, rather than a fixed script. It can look things up across systems, take several steps and come back with either a finished piece of work or a specific question.
Here is an illustration I often use with prospects. Someone in sales wants 40 laptops for new joiners. A workflow tool creates a ticket and waits. An agent can check the catalogue, find that there's already a framework agreement with a hardware supplier, confirm the budget line has room and draft the PO at contracted pricing for the buyer to approve. Along the way it might notice the delivery address is an office due to close next quarter and ask the requester whether that's still right before anything goes out.
None of those steps is hard. A good buyer would do all of them. The difference is that the buyer has 60 other requests in the queue, while the agent can prepare this one in a few minutes and move on to the next.
People in these meetings often ask whether this is just a procurement chatbot. It isn't. A chatbot answers a question about your laptop policy, while an agent prepares the purchase within that policy and tells you what it did. I wrote about that difference in more detail in From Chatbots to Agents: What Agentic AI in Procurement Actually Changes.
Which procurement tasks to give AI agents first
My rule for choosing a starting point is to look for work that is high in volume, follows written rules, is already neglected and produces mistakes that are cheap to catch. Procurement teams usually know exactly where that work sits, because it's the work nobody has time to do properly.
Tail spend. The usual rule of thumb is that a small share of spend is spread across the large majority of suppliers. Nobody runs a sourcing event for a one-off purchase of safety signage, so it gets bought from whoever the requester found on Google. An agent can collect three quotes, check the vendors against your approved list and recommend one for the buyer to confirm.
Supplier onboarding and compliance. Collecting tax forms, bank letters, certifications and insurance documents is slow and repetitive. An agent can request them, check expiry dates, spot a mismatch between the company name on the bank letter and the one on the registration certificate, and follow up until the file is complete. We cover this workflow in more depth in AI for supplier onboarding.
Invoice exceptions. When a three-way match fails, someone has to work out why. Was it a partial delivery, a price change nobody updated in the system or a duplicate? An agent can pull the PO, the goods receipt and the invoice, identify the likely cause and propose the fix for accounts payable to approve.
Contract obligations. Most contracts contain terms the company should be acting on and isn't: price escalation caps, volume rebates and auto-renewal notice periods. They sit in PDFs. An agent can extract them, keep track of them and warn you 90 days before a renewal you meant to renegotiate. Be careful here, because extraction from scanned or complex PDFs is still inconsistent. Check the extracted terms against the source document before anyone relies on them.
Supplier risk monitoring. Watching news, filings and delivery performance for a few thousand suppliers isn't something a team can do continuously. An agent can, and it only needs to raise an alert when something changes.
What should stay with procurement people
Agents are poor at the parts of procurement that depend on relationships and on context nobody wrote down. They won't know that a key supplier is having a rough quarter and that pushing for a discount now would cost you goodwill you'll need next year. They won't know the CFO has quietly decided to consolidate logistics vendors. Strategy, any negotiation that matters, judgement calls about people and every approval decision remain the team's job.
The sensible way to bring agents in is to treat them like a capable new hire on probation:
Stage 1: Observe. The agent gets read access and reports what it finds, without sending or changing anything.
Stage 2: Draft. The agent prepares the PO, the supplier email or the exception fix, and a person approves every item.
Stage 3: Act within limits. The agent handles low-risk, reversible actions on its own, such as chasing a missing certificate or updating a tracker, within a defined set of categories and a spend threshold. Anything that commits spend, approves a supplier or accepts terms still goes to a person.
Decide early what the recovered hours are for. If you don't, they disappear into more of the same queue. The supplier reviews, category plans and negotiation preparation that currently get squeezed are the obvious candidates.
Why AI agent projects in procurement fail
In my opinion, most failed agent projects in procurement will fail for unglamorous reasons.
The first is data. If your vendor master lists the same supplier four times under slightly different names, an agent working from it will make mistakes faster and with more confidence than a person would. Cleaning that up is usually the real first step, and we've written about how to approach the procurement data problem.
The second is policy that lives in someone's head. If your purchasing rules have fourteen exceptions and only one person in accounts payable knows them all, the agent won't follow them. Writing the rules down properly helps the agent, and it also helps the next person you hire.
The third is audit. Procurement decisions get questioned, sometimes years later. Every action an agent takes needs a record of what it saw, what it decided and why, in a form an auditor can read. Build this into your procurement AI governance framework before the first pilot starts.
The fix for all three is the same: start narrow. Pick one process, one category or one business unit. Record today's baseline, such as cycle time, the number of open exceptions or the share of tail spend going to approved suppliers, so you can measure the change honestly before you widen the scope.
Where to start with AI agents in procurement
For twenty years procurement leaders have argued that the function should be strategic: shaping supplier relationships, managing risk and finding value the business didn't know was there. Most of them meant it. Few teams ever had the time, because operational work filled every available hour.
Agents are the first tool I've seen that can take work off that pile instead of reorganising it. That only holds if the data is usable, the rules are written down and the agent is supervised while it earns trust. If I were a CPO reading this, I would pick one process from the list above, run the agent in observe mode for a month and look at what it finds before deciding anything else.
If you're trying to decide which process to start with, we can help you choose the workflow and design the pilot. If your team first needs to get comfortable working with AI tools day to day, our hands-on procurement AI training is built around their own procurement work.
Deepak Chander is Co-Founder of MoleculeOne.ai, an AI-native procurement consultancy that trains and builds alongside procurement and finance teams turning AI adoption into decisions they can trust.