AI in Procurement: What Works, What Fails, and What to Do First

Start with the two numbers that frame everything else.

In the Hackett Group's 2026 research, 43% of procurement organisations are actively working on AI. Only 12% have it running at large scale.

That gap is the subject of this article. Not whether AI works in procurement. It does, in specific places. The question is why so many attempts produce nothing, and what the ones that work do differently.

The short answer, which the rest of this article is evidence for: the software is rarely the problem.

#First, what AI actually is in a procurement context

Three different things get called AI, and they fail in different ways. Worth separating before going further.

Automation with rules

If the invoice matches the purchase order within tolerance, post it. No intelligence involved, and none needed. This is the oldest and most reliable category, and a lot of what gets sold as AI is this with a new label.

Prediction and classification

Sorting spend into categories, matching duplicate supplier records, flagging an invoice that looks unusual, scoring a supplier's risk. The machine learns patterns from your data. Genuinely useful, and completely dependent on the quality of what you feed it.

Generative AI

The large language models. Good at reading and writing: summarising a contract, drafting an RFP, answering a question about a policy, pulling key terms out of 400 agreements. Confident even when wrong, which is the whole problem.

A fourth term, agentic AI, has arrived recently. It means software that chains several steps together and acts, rather than just answering. More on that below, because it deserves care.

#Where AI is genuinely working today

Deloitte's 2025 survey of more than 250 chief procurement officers across 40 countries asked where they are actually using generative AI. The answers are narrower and duller than the marketing suggests, which is a good sign rather than a bad one.

Look at what these have in common. Every one is reading, sorting, or drafting. Not one of them is deciding.

That is the honest shape of AI in procurement in 2026. It reads faster than you and writes a passable first draft. It does not choose your supplier.

When Deloitte asked where the value actually came from, the top answer was better analysis and decision-making at 67.7%, with productivity second at 49.4%. Note the order: the gain is mostly in seeing your own data properly, not in doing tasks faster.

#Why most procurement AI projects produce nothing

The failure rate is not a secret and it is not small.

• MIT's Project NANDA studied 300 public AI deployments and found 95% of pilots produced no measurable effect on profit or loss.
• Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls.
• The 2026 ProcureCon CPO report found only 11% of procurement leaders using AI with measurable impact, against 89% who rate themselves ready for it.
• The year before, Hackett found 49% of procurement teams had piloted generative AI and only 4% had deployed it at scale.

Four different research firms, four different methods, the same shape of answer. So what is going wrong?

Reason one: the data underneath is not usable

This is the big one, and it is not close.

Think about what the top use cases require. Spend analysis needs spend that has been sorted into categories. Supplier risk scoring needs a supplier list where each company appears once, not four times under slightly different names. Contract analysis needs contracts stored somewhere a machine can read them, rather than in an email folder and a filing cabinet.

Point a good model at a bad record and it produces confident nonsense, faster than a person could produce careful nonsense. That is worse than no answer, because it looks like an answer.

Reason two: the barriers are human, and known in advance

The 2026 ProcureCon report asked what was blocking progress. Data privacy, security and compliance concerns came first at 67%. Data quality and getting systems to talk to each other came second at 54%. Resistance to change, including discomfort with AI replacing human judgement, came third at 51%.

This matches what people say when they are not being surveyed. Saurabh Gupta, Chief Strategy Officer at HFS Research, described what came back when his firm asked around a hundred senior executives what was actually stopping them:

The two biggest barriers were culture, and two was mindset. Nobody said anything else. Probably the third was talent, lack of talent. None of these has to do anything with technology.

Saurabh Gupta, Chief Strategy Officer, HFS Research, on the Art of Procurement podcast

Reason three: automating a broken process just breaks it faster

In the same conversation, the interviewer put the point more simply: there is no good in making something that does not work more efficient.

If your approval chain has five steps that nobody can justify, adding AI gives you five unjustifiable steps that execute quickly. The consultancy Comprara describes the same failure as a sequencing error — technology bought before the process is fixed, and operating models redesigned without a baseline. Their line is the right one: a hospital does not design a treatment plan before running the diagnostic.

#The returns, and why they vary so much

Deloitte's respondents estimated roughly 2x return on their generative AI investment on average. Split by how mature the organisation was, the leading group estimated 3.2x and the rest 1.6x.

Worth flagging honestly: the same report's summary page quotes the leaders at 2.8x rather than 3.2x. Two numbers for the same thing in one document, so treat the exact multiple with caution. These are self-reported estimates, not audited results.

The pattern is the durable part. Roughly double the return, from comparable tools, in organisations that had already done the unglamorous work underneath. That is the whole argument of this article expressed as a ratio.

The same survey shows how much further ahead the leading group is. They deploy generative AI four times more often than the rest, 62% against 15%. They use flexible automation more than three times as often, 60% against 18%. They are not using different software. They got the foundations right, then went faster.

#Agentic AI: what it means and how much to believe

Agentic AI is the current headline. The idea is that the software carries out a sequence rather than answering a single question. Read the requisition. Check it against policy. Find suppliers, request quotes, compare them, route for approval.

Some of this is real. Tail spend, routine sourcing events and continuous supplier monitoring all suit it well. The rules are stable, and the value of any single decision is low.

Two cautions, both from the same analyst house that promotes the category.

First, Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Second, and more useful when you are being sold to: Gartner estimates that of the thousands of vendors describing themselves as agentic, only around 130 genuinely are. The rest is what they call agent washing — existing products rebranded.

#What AI cannot do, and probably will not soon

This matters as much as the use cases, because most disappointment comes from expecting the wrong thing.

• It cannot decide which categories matter to your business. That depends on strategy, not patterns.
• It cannot absorb the political cost of telling a business unit that its preferred supplier is failing.
• It cannot hold a relationship together through a shortage, when the question is who gets the limited stock and why.
• It cannot carry an accountability. A regulated approval, an airworthiness determination, a clinical supply decision — a person signs those, and no model changes that.
• It cannot fix data it was not given. If a cost sits in a spreadsheet on someone's laptop, no tool will find it.

Roland Berger put the division well in their research on procurement in 2030: AI synthesises and analyses, while humans judge, influence and decide.

#The rules are arriving, and the dates moved

If you operate in or sell into the European Union, the AI Act affects what you deploy. The timeline has shifted, so plans written in 2024 are now wrong.

For most procurement uses this is lighter than it sounds. Summarising a contract or classifying spend is not a high-risk use. Where it starts to matter is anything touching people — screening suppliers in ways that affect employment, or automated decisions with legal effect. The practical step is to write down which AI features you use, what each one decides, and who reviews it. That list is the beginning of every compliance answer you will be asked for.

#The order of work that decides the outcome

This is the part worth keeping. It is deliberately unexciting, and it is the difference between the 12% and everyone else.

Step 1: classify your spend

Every pound or dollar in a category your own team recognises, with an unclassified bucket that is reported rather than hidden. Nothing downstream works without this. Not analysis, not risk scoring, not supplier consolidation, not AI.

Step 2: clean the supplier list

One record per company. Most mid-sized companies have the same supplier several times over, under different spellings. Nobody can then see the real total spent with them. And no risk score built on that list is trustworthy.

Step 3: put the contracts somewhere readable

One repository, with end dates, notice periods and auto-renewal status recorded. This alone pays for itself through renewals you stop missing, before any AI touches it.

Step 4: pick one use case with a number attached

Not a strategy. One task, with a measurement taken before you start. Contract summarisation is a good first choice: the input is messy text, the output is checkable by a human in seconds, and being wrong is cheap. Spend classification assistance is another.

Step 5: measure against the baseline you took

Hours spent, error rate, cycle time. If you did not measure before, you cannot show a result afterwards, and the project gets defended rather than evaluated. This is the single most common reason a pilot that worked gets cancelled anyway.

#Will AI replace procurement jobs?

The honest answer has two halves, and most articles only give one.

KPMG has estimated that generative AI could automate 50 to 80% of current procurement work. Take that as a statement about tasks rather than people, because tasks are what it measures. A great deal of procurement work is reading, copying, checking and chasing. That is exactly what this technology does well.

The other half is the resourcing picture. Procurement workload is rising about 8% in 2026 while headcount and budgets fall. Most teams are not sitting on spare capacity waiting to be removed. They are behind, and quietly dropping work nobody is watching.

So the realistic outcome for most mid-sized teams is not fewer people. It is the same people finally doing the work they were hired for, because the checking and chasing stopped consuming most of the week. The roles that are genuinely exposed are the ones that consist entirely of moving data between systems.

#What to take from all this

AI in procurement is real, useful, and considerably narrower than it is sold. It reads, sorts and drafts well. It does not decide, and it cannot repair the data it is given.

The organisations getting twice the return are not running different software. They spent the unglamorous months first.

If you do only one thing after reading this, do not evaluate a tool. Find out what share of your spend is classified, and how many times your largest supplier appears in your master list. Those two answers will tell you more about whether AI can help you than any demo will.