Where AI ROI in Procurement Is Real, and Where You Will Struggle to Prove It
Engineering finds it easy to put numbers on AI. Non-technical teams usually can't, and the fix starts before the rollout.
AI ROI in procurement is easy to prove where the outcome is specific and checkable, such as classifying PO lines, validating requisitions or matching invoices to POs. It is hard to prove for category strategy, negotiation strategy and annual planning, where people make the judgement calls and AI handles drafting, formatting and coordination. Capture a baseline before rollout, deploy where the hands-on work happens, and convert results with three formulas: efficiency, quality and capacity value.
I recently had a conversation with a Head of Procurement at one of our clients in the UAE, who told me where his AI rollout had stalled. The tools were in use and his team was getting through work. His problem was the question leadership kept asking: what is this returning? Engineering had no trouble answering it. They came back with exact numbers and clean data. His department, like other non-technical teams, was struggling to measure the return properly.
My read is that the cause is structural. Engineering work already produces its own measurements: tickets closed, builds shipped, cycle times. Procurement work mostly lives in inboxes, spreadsheets and approval chains, and nobody was counting it before AI arrived. The problem sits in measurement, and it has a practical fix.
Where AI ROI is easy to prove
AI returns are easiest to prove where the outcome is specific and checkable. A purchase order line reads "36 ergonomic task chairs". Is that furniture or office supplies? A model classifies lines like this well, and your team can verify the answer. You can check the answer, the task repeats at volume, and you know how long a person took to do it by hand, which gives you a baseline.
Compare that with the instruction "improve this RFP". There is no right answer, so nobody can say whether the output is better. The model may be doing excellent work and you still cannot show it to a CFO.
Start with the baseline: the M in MERIT
In our MERIT framework, M stands for Measurement: define success metrics and capture a performance baseline before any AI goes live. Most of what the Head of Procurement described traces back to this step. If nobody recorded how long PO validation took, or how many invoices failed matching, before the rollout, there is nothing to compare today's performance against. A vendor dashboard full of usage statistics will not fill that in. We wrote about this previously in our MERIT framework series, in the first article, How to Build an AI Measurement Framework for Procurement.
If you are already live and skipped the baseline, you can rebuild one. Pull 30 recent items from the workflow, have someone do them by hand with a stopwatch and an error log, and compare that against the AI-assisted version. The baseline will be rough, and that is fine as long as you can explain how you got it.
Where to deploy AI first in procurement
If you are struggling to show AI ROI, go to the core of the hands-on work. These are the workflows where we would start, with what to measure in each.
| Workflow | What AI does | What you can measure |
|---|---|---|
| PO and PR validation | Checks requisitions for missing fields, wrong cost centres and policy breaches before they reach an approver | Rejection rate at approval, minutes per requisition |
| Invoice and PO matching | Compares the PO, goods receipt and invoice, and flags differences in quantity, price and terms | Exception rate, hours spent resolving exceptions |
| Intake and ticket routing | Reads free-text requests and sends them to the right team with the right priority | Misroute rate, time to first response |
| Spend classification | Maps line items to categories | Share of spend correctly classified, analyst days per quarter |
| Data cleaning | Merges duplicate supplier records, standardises names and fixes inconsistent fields | Duplicate count before and after, supplier master error rate |
| First-draft documents from context | Drafts a scope of work, a summary or a vendor comparison from emails and notes | Time to first draft, number of review rounds |
| Combining multiple files | Pulls contracts, quotes and spreadsheets into one comparison view | Hours spent consolidating per sourcing event |
| Contract checks against a standard | Compares incoming terms with your playbook and flags deviations | Review cycle time, deviations caught before signature |
| Document verification | Checks onboarding packets, certificates or expense receipts against requirements | Onboarding errors, turnaround time |
In each of these the input is defined and someone can check the output, so you can put a number against it.
How to calculate AI ROI from your baseline
Once the baseline exists, the I in MERIT (Impact) is arithmetic. In Part 2 of the series, on reporting procurement AI ROI to leadership, we used three formulas that cover most of what leadership asks for:
Time saved multiplied by loaded hourly cost.
Error rate before minus error rate after, multiplied by cost per error.
More volume processed with the same headcount, framed as hours reallocated to higher-value work rather than headcount removed.
Here is an illustrative example for invoice matching. The numbers are assumptions for the sake of the arithmetic, so replace them with your own. Say your team processes 2,000 invoices a month and 30% hit an exception, which is 600 exceptions. After AI-assisted matching the exception rate falls to 12%, which is 240. That is 360 fewer exceptions a month. If each one took 30 minutes to resolve, you have recovered 180 hours a month, and you can multiply that by your loaded hourly cost for a figure leadership can read. To run the same arithmetic on your own volumes, use our AI Procurement ROI Calculator, or read the full guide to calculating, projecting and defending AI procurement ROI.
That is also where E (Evidence) comes in. The first question in the room is often where the data came from and who has access to it. Have a one-page answer on data security, compliance and auditability ready before you present the numbers.
Where AI ROI is hard to see in procurement
Some work will never give you a clean number.
- Category strategy. AI can pull market data and draft options for a three-year IT hardware strategy. The payoff shows up over 18 months or more, mixed in with market movements and supplier behaviour, so you cannot isolate what AI contributed.
- Negotiation strategy. Savings depend on what the supplier does in response. Two negotiations are never the same, so there is no like-for-like comparison.
- Annual planning. Nobody knows what the plan would have looked like without AI, so there is no counterfactual.
- "Make this better" tasks. With no standard for better, you have no way to score the output.
None of this means you should keep AI away from these tasks. Use it, and keep the judgement with people. In category strategy, negotiation and annual planning, the calls that matter belong to the category manager or the CPO. AI works best around those calls: drafting the strategy document or negotiation brief from inputs you have agreed, formatting it into your template or deck, and handling the coordination around it, such as collecting stakeholder inputs, consolidating meeting notes, tracking actions and drafting the follow-up emails. Those are real hours saved, and they are the parts you can time.
The issue is the business case. If your case for AI rests on the strategy work itself, you will struggle to defend it at the next budget review. Let's be honest about the stakes. AI helps in most cases, but that does not protect a programme. Leaders still have to find a way to show the ROI, and in today's economy any programme that cannot show an ROI it can defend will be cut, AI programmes included.
You can often split a fuzzy task into measurable parts. "Improve this RFP" becomes "check the RFP against our template for missing sections" (checkable) and "draft the evaluation criteria from the approved brief" (you can time the first draft). The strategic judgement stays with the category manager, and you now have two numbers.
This is the T in MERIT, Trust: the conditions that let an AI programme earn the right to scale, which we cover in Part 3, on moving from pilot to infrastructure. Early wins on measurable workflows give you the credibility and the data to take on the harder problems later.
A quick test for your first AI use case
Before deploying AI on a task, ask three questions.
- Is there a right answer? Can someone check the output?
- Is there volume? Does the task happen dozens or thousands of times a month?
- Do we have a baseline? Do we know what it costs today in time or errors?
Three yeses make a good first use case. Zero or one means it is probably not where to start. If you want a structured view of where your team stands, the AI Readiness Assessment takes about ten minutes.
What I would do in his position
Pick two workflows from the table above. Capture a baseline on each from a sample of 30 items. Name one person who owns the numbers. Report monthly using the three formulas, with a single line on where the data comes from. It takes a few hours, and you will have an answer the next time leadership asks what the programme is returning.
Frequently asked questions
How do you measure AI ROI in procurement?
Capture a baseline before AI goes live: time per task, error rates and volume on the specific workflows it will touch. Track the same metrics afterwards, then convert the difference with three formulas: time saved multiplied by loaded hourly cost, error-rate reduction multiplied by cost per error, and volume growth with flat headcount.
Which procurement tasks give the clearest AI ROI?
Tasks with a checkable output and high volume: PO and PR validation, invoice and PO matching, intake and ticket routing, spend classification, supplier data cleaning, first-draft documents from context, contract checks against a playbook and document verification. Each has a right answer, a volume and a baseline you can measure.
Why is AI ROI harder to measure for non-technical teams than for engineering?
Engineering work already produces its own measurements, such as tickets closed, builds shipped and cycle times. Procurement work mostly lives in inboxes, spreadsheets and approval chains, and nobody counted it before AI arrived. Without a baseline there is nothing to compare against, so the return cannot be shown even when it exists.
What if we rolled out AI without capturing a baseline?
You can rebuild one. Pull around 30 recent items from the workflow, have someone complete them by hand with a stopwatch and an error log, then compare against the AI-assisted version. It is a rough baseline, but a rough baseline you can defend is more useful than a precise one you do not have.
What happens to an AI programme that cannot show ROI?
In today's economy it is likely to be cut, however useful the team finds it. AI helping in most cases does not protect a programme on its own. Leaders need an ROI they can defend, which means a baseline captured before rollout, metrics tracked consistently and a financial translation of the results.
Can you measure AI ROI for category strategy or negotiation?
Rarely in a clean way. The payoff takes months to appear and depends on market movements and supplier behaviour, so you cannot isolate what AI contributed. That does not mean avoiding AI there. People make the judgement calls, and AI handles document drafting, formatting and coordination tasks such as consolidating inputs and tracking actions. Those parts you can time. Do not build your business case on the strategy work itself.
We can help you design the pilot and the measurement plan. If you are rolling AI out to a team, our AI procurement training is built around hands-on workflows like these. The MERIT Baseline Capture Template is a good place to start: define metrics, capture your baseline and set up tracking before deployment.
Get the Baseline Capture Template →