Key takeaways

‍

• The purchase order process is the strongest first project for AI in procurement, because its work repeats, follows rules, and spans several systems.

• Turning on the AI built into your existing software is adoption, but doesn’t necessarily improve outcomes on its own. Correctly choosing which steps an agent should take on is what leads to improvement.

• Step-level data on time, waits, handoffs, and rework shows which purchase order steps an agent can take on.

• Approvals, supplier decisions, and exceptions should have a larger degree of human-in-the-loop, with approval limits and segregation of duties dependent on each workflow.

‍

AI in Procurement: Why the Purchase Order Is the Place to Start

‍

Procurement workloads are projected to rise 8% in 2026, while headcount will dwindle by .09%, according to The Hackett Group, via Business Wire. That means more requests to clarify and more approvals to chase, with fewer people to do it.

‍

Many teams are trying to stay balanced through the use of AI in procurement. According to the same study, 43% of procurement organizations are actively pursuing AI deployment, nearly double the year before. Even so, only 12% report large-scale implementation.

‍

Practitioners are justifiably skeptical about having AI make decisions about the acquisition of goods and services, and especially decisions about people. AI’s judgment is sometimes suspect, as anyone who has experienced AI hallucinations can attest, so automating key purchases or supplier negotiations makes businesses wary.
‍

Fraud is another worry, since an agent needs to detect suspicious billing patterns, irregular contractor behavior, and shoddy tools among other scamming. AI fraud detection is in a cat-and-mouse battle with AI scamming, so any AI procurement planning needs air-tight governance.
‍

AI in procurement is most efficient when working off your company’s own data as a baseline. It does best in purchase orders or other areas with well-defined rules. But even with well-defined rules, it needs as much relevant context as possible to deliver reliable outcomes, with human overview to fill in the gaps.

‍

What is the first procurement process to replace with an AI agent?

‍

Start with the purchase order process. Turning requests into POs, matching POs to receipts and invoices, and routing approvals is high-volume, rules-based work spread across email, the ERP, and supplier portals. An AI agent can handle such repetitive busywork easily while also cutting down on the cumulative time drag of context switching, navigating, and logging into multiple systems. 

‍

Four traits make purchase order automation a strong first project: 

‍

• High volume. The same work repeats hundreds or thousands of times a year, so a small efficiency gain on each PO adds up to massive savings in time, money, and operational errors.

• Easily controlled. Clear chain of custody sets out who buys, handles, and approves the PO at each step. This gives an agent something concrete to base its model on, and a reviewer something concrete to verify.

• Cuts down repetitive manual work. PO work crosses several systems, with details getting re-keyed by hand, multiple apps logged into, and approvals getting chased up by email. Once fine-tuned, AI can shave off a substantial amount of rote work with little oversight.

• Built-in approvals. Approval steps already exist, so people stay in charge: an approver decides, and an agent routes the request and sends reminders.

‍

Procurement leaders see the opportunity. The Hackett Group’s 2025 Study found purchase order processing among the most transformative areas for implementing AI, alongside spend analytics and e-procurement.

‍

Strategic sourcing, negotiations, and supplier relationship decisions are poor first picks for agentic process automation. They are more ill-defined areas of decision making involving human relationships, so an agent rarely has clear policy to follow, or enough context.
‍

Always check each step against your own process before implementing. Pick the exact steps and sequence from measured data on how your team actually works.

‍

Why most procurement teams pick the wrong first step

‍

Two neologisms get mixed up: Human + AI and AI + Human.
‍

Human + AI means staff using AI tools in their daily work, which you could call Human + AI, or basic AI adoption. A buyer asking a chat assistant to tidy up a supplier email is an example of Human + AI adoption.
‍

AI + Human reverses means handing an existing process to AI while people review and approve, which you could call AI + Human, or implementation. An example is an agent drafting a requisition from an incoming request, and a buyer checking it before sending it forward.
‍

One trap is simply switching AI on and assuming it drives results. In Hackett’s 2026 study, 69% of organizations access AI through features built into their existing procurement platforms. That’s AI adoption, but it isn’t the same as making informed decisions about where AI can change work. An AI adoption report can show you the difference.
‍

A built-in feature might summarize supplier emails, while your team loses most of its time somewhere else entirely. Choosing well starts from your own process and drilling down on the repetitive, rule-based work inside it.

‍

A second trap is measuring how long buyers spend in the ERP. That number shows a system is in use, but it says nothing about which steps are ready for an agent. Hours in the ERP could be repetitive keying or a slow approver, and misled you about productive time spent. It also misses the email threads and supplier portals where the rest of the PO work happens.

When determining which AI is best for procurement, ask: which process, and which steps in it, should an agent take on? Step-level data answers that.

‍

What the purchase order process looks like, step by step

‍

Say a department manager asks your team for “the usual order.” In a typical purchase order process, the request goes through nine steps before anyone is paid:

‍

1. Request. It arrives by email or chat late on a Friday, with no account number or clear quantity.

‍

2. Clarification. A buyer writes back for the quantity and cost center, and the request waits for the reply.

‍‍

3. Requisition. Someone keys the request details in by hand.

‍

4. Budget and contract. Someone looks up the budget separately and checks the contract price in files and spreadsheets.

‍

5. Approvals. Someone chases each approver for a signature, and the requisition waits again.

‍

6. PO sent. The purchase order goes to the supplier.

‍

7. Confirmation. The supplier confirms, or the buyer chases the confirmation by email.

‍

8. Delivery. The order arrives and someone records the receipt.

‍

9. Three-way match. The invoice is compared with the PO and the receipt, line by line, and any mismatch turns into more emails while the order waits further.

What that costs varies widely. According to APQC, organizations spend from about $14 to more than $54 to process a single purchase order. The difference is driven largely by how procurement work is structured and done, and the industry involved. Procurement cost reduction in banking needs a different lens, for example.

‍

APQC also found that top performers issue 4,000 orders per FTE and take 1 day to issue a PO, while bottom performers issue 1,619 per FTE in 2.5 days.

‍

Those figures show how a team compares with its peers. They don’t show where each order waited along the way, which is only fixable when you measure each step. Many of the waits are handoffs between people and systems, which is the kind of work an agent can pick up.
‍

Your own process will have specific snags a general list won’t show, which will only surface with context-sensitive data capture and processing.

The Process Savings Matrix for the purchase order process

‍

A matrix for procurement process automation puts two views of the PO process side by side. The left side is what surface-level tool data shows: where the hours went. The right side is what step-level process data shows: which step used them.

‍

‍

The last two columns turn that detail into a decision. High agent fit means an agent can handle most of the step, and Medium means the agent suggests while a person decides. Three rows show the range:

‍

• Issue the PO and confirm with the supplier: high. Each PO is sent and followed up by hand, and neither action needs much judgment, which is why automated purchase orders work well here with a person spot checking.

• Pick the supplier and check the contract price: medium. Looking up the contracted supplier and price is mechanical, but choosing between options isn’t, so the agent saves the lookup and the buyer keeps the decision.

• Match PO, receipt, and invoice: high for matching, medium for exceptions. 3 way match automation suits the comparison itself, since checking three documents follows fixed rules. An exception such as a short delivery still needs a person to work out what happened.

‍

Read the ratings as a shortlist of steps to examine. What an agent saves depends on how often each step runs and how often it goes wrong, and only measurement shows that. That measurement is where workflow optimization starts. Task mining explains how capturing activity across apps works, and process mining vs. task mining compares it with reconstructing a process from system records.

‍

Measure the process before you automate it

‍

The old way to measure a process is to bring in a consultant, who sits behind your staff for two to three weeks and documents the work by hand. It misses steps for two reasons. A few weeks covers only the orders that happen to arrive in that window, so rare exceptions may never appear, and one person can’t follow every order through every application.

‍

Five measures show which steps are ready for an agent:

‍

• Time per step, which shows the slow steps.

• App switches, where a high count suggests details are being copied by hand between systems.

• Waiting time, which separates slow work from work stuck with an approver or supplier.

• Handoffs, which show where work stalls between people, teams, or systems.

• Rework, which shows where a step fails the first time and has to be redone.

‍

Read together, they point to the right fix. If requisitions keep waiting at the same approval step, agent-routed reminders may help. If matches keep coming back as exceptions for the same reason, the problem may sit in the data feeding the match.

‍

Where people stay in control of AI in procurement

‍

Drafting a requisition, suggesting the contracted supplier, routing an approval, and matching documents are rule-based and easy to reverse. Approving spend, choosing a supplier, and resolving exceptions, or any other decision that commits money or changes relationships should remain human.

‍

Handing spend-related steps to an agent raises a fair question about fraud: an agent that could approve its own spend, or change who gets paid, would be a serious problem. Agentic AI procurement needs the same controls a well-run team already uses:

‍

• Approval limits. Thresholds by amount and role decide which requests need a named approver.

• Segregation of duties. The person who raises a request never approves it.

• Supplier master review. A changed supplier record is a common way payments get redirected, so every change gets a second look.

• An action log. Every agent action is recorded, so anything it touched can be checked afterward.

‍

People set the limits and the workflow enforces them, so anything outside a limit goes to a person. That frees buyers for the work that needs them, like comparing suppliers or renegotiating a contract, instead of chasing signatures.

‍

Start with a 14-day audit

‍

The most reliable way to ground AI for procurement in real data is to record the work itself. Work Intelligence captures how the work actually happens across apps, step by step, across the whole team. For the PO process, that shows how long a requisition waits for an approver and how often a step has to be redone. 

‍

Because it draws on recorded work across the full period, it picks up exceptions a consultant would miss. It’s a low-risk first step for AI in procurement, with nothing automated until the data confirms it. Request your 14-day audit to see which procurement steps are ready for automation.

‍

Frequently asked questions

‍

Can procurement be replaced by AI?

No. AI takes over repetitive purchase order work, and people keep sourcing, negotiation and supplier decisions, which rest on judgment and context. Agents handle steps such as drafting requisitions and chasing approvals, so buyers have more time for that work.

‍

Which part of procurement should I automate first?

The purchase order process. It’s high-volume work with clear rules and approval steps that keep people in charge. Measured data then shows which steps inside it to automate first, and choosing a tool comes after that.

‍

What is purchase order automation?

Purchase order automation is software and agents that create, route, send, and match purchase orders with less manual work. It covers requisitions, approval routing, sending the PO to the supplier and matching it to the receipt and invoice, while the people who approve spend stay in the process.

‍

What is procure-to-pay automation?

Procure to pay automation covers the full flow from a purchase request to supplier payment, and P2P automation is the shorter name for it. The PO sits in the middle, after the request is approved and before the invoice is matched and paid, which is why it touches both approvals and matching.

‍

How much does it cost to process a purchase order?

According to APQC, organizations spend from about $14 to more than $54 to process a single purchase order. The difference comes largely from how the work is structured and done, so measuring your own process shows where your cost sits in that range.

‍

Can AI agents approve purchase orders?

Only inside limits your team sets. An agent can draft a PO, route it for approval, and flag a mismatch, while spend approvals and exceptions stay with people. Approval limits and segregation of duties keep it that way.

‍

How do I measure my PO process before automating it?

Measure time per step, app switches, time waiting on approvers and suppliers, handoffs, and rework. Take the numbers from real work instead of interviews, because interviews describe the intended process, and collect them across the whole team so no single person’s habits skew the picture.

‍