If you already own Coupa or SAP Ariba, the best AI procurement tool to test first is the one in your suite, because there is no new vendor to onboard. Coupa says AI usage is free until the end of 2026. SAP is less clear: its own documents bill procurement agents per action, so get the Ariba cost in writing before you test. Add Pactum or Fairmarkit only for tail spend the suite leaves untouched after that test. Without a suite, start with Zip for intake. Do not buy Ariba for its AI. Fairmarkit, Pactum and Zip each do a different job, so a shortlist that treats them as substitutes will compare the wrong things.
| Tool | Job it does | Buy it when | Skip it when | Published price (checked Nov 5, 2026) |
|---|---|---|---|---|
| Zip | Intake and purchase orchestration | No suite, or employees bypass the suite's front door | Your suite's intake already gets used | None. Vendr reports an $88,856 average across 77 purchases |
| Fairmarkit | Automated quotes on tail spend | You have many small, comparable buys and suppliers who will bid | You cannot supply a historical price benchmark | None found |
| Pactum | Supplier negotiation by bot | Many suppliers on terms nobody ever negotiated | Your spend is concentrated in a few strategic suppliers | "We do not publish list prices" |
| Coupa | Full source-to-pay suite | You need sourcing, contracts, procure-to-pay and AP in one system | You only need a better request form | None. Vendr reports $94,519 a year across 128 purchases; moving to outcome-based pricing |
| SAP Ariba | Full source-to-pay suite | You already run it, or SAP is a hard requirement | You are buying it for the AI | No buyer list price found; supplier fees are documented |
Which AI Procurement Tool Fits Which Job?
The category splits into four jobs, and most bad evaluations come from scoring one job's tool against another's. Intake orchestration decides how a purchase request enters the company and who approves it. Tail-spend sourcing collects competing quotes on small purchases that no buyer has time to run. Autonomous negotiation means software bargains with a supplier inside limits you set beforehand. A source-to-pay suite is the system of record from sourcing through invoice payment.
The vendors are also moving into each other's territory, which makes the labels less reliable. Pactum's pricing page now lists Intake and Tactical Sourcing digital workers beside negotiation. Zip added a price negotiation agent described as finding overpayments and supporting better deals, which is help for a human negotiator and differs from a bot closing the agreement. Spendflo's catalogue lists a Coupa Sourcing Optimization Negotiation Agent without saying whether it is generally available, so ask Coupa for its status in writing.
These products also stack. Sonoco runs Fairmarkit inside its Coupa workflow: requests start in Coupa, route to Fairmarkit for quotes, and the award flows back to the requisition. So a suite owner should ask which gap remains after the suite does its part.
Most buyers are starting from embedded AI anyway. The Hackett Group's March 2026 study found 69% of organizations access AI through capabilities embedded in existing procurement platforms, and only 12% report large-scale implementation.
What Walmart's Pactum Deployment Proved
Walmart proved that a bot can close terms with small suppliers of goods Walmart does not resell, at modest savings. It did not test negotiation with merchandise vendors or strategic suppliers. The program grew out of a plain capacity limit: Walmart has more than 100,000 suppliers, and around 20% are on cookie-cutter terms that often go unnegotiated.
The numbers differ between sources because they describe different moments, denominators and questions:
| Figure | What it measures | Source |
|---|---|---|
| 64%, about 1.5% savings, 11 days | Three-month Canadian pilot, early 2021, reported against 89 suppliers | Harvard Program on Negotiation |
| 64% of 100 invited | The same pilot with a different denominator, plus payment terms extended to an average of 35 days and a 20% target | Talking Logistics summary of HBR |
| 68%, 3% savings | Late 2022, after expansion to the U.S., Chile and South Africa | Harvard Program on Negotiation |
| 83% favorable | Only suppliers that closed deals, in post-pilot interviews | Harvard Program on Negotiation |
| 83% "easy to use" | Pactum's January 2023 release, a different question from preference | PYMNTS |
| About 75% prefer the bot | Walmart's internal findings, as reported by Bloomberg in April 2023 | PYMNTS |
| 82% prefer the bot | A Walmart payment-terms campaign, per Pactum's product page | Pactum |
| 85% prefer the bot | Pactum's blog, with no sample size or date given | Pactum |
Treat the preference figures as unverified until someone shows you the sample. The 83% favorable number excludes every supplier who walked away, and the rest are drawn from unstated populations. Our search found no independent audit of any of these numbers and no supplier-side study.
The scope matters more than the percentages. Reporting describes smaller contracts with suppliers of equipment Walmart uses, not goods it sells, and the pilot covered fleet services and store equipment. Pactum's own page aims its discount campaigns at mid-sized and long-tail suppliers and names MRO, industrial supplies and services as the best-fit categories. Pactum's CEO has also warned that these bots can "create harm" if fed inaccurate information, so bad supplier master data is a negotiation risk as well as a reporting one.
Zip, Fairmarkit and Pactum: Who Should Not Buy Each
Each specialist has a clear disqualifier, and it is usually what you already own.
Zip reports $355 billion in spend processed in 2025 across more than 7 million suppliers, a usage figure that says nothing about savings. Its comparison page against Coupa claims faster time to value and lower total cost of ownership, and it gives no timeline or number for either. Skip Zip if you run Coupa or next-generation Ariba and your employees already use its intake. Paying roughly $89,000 a year (Vendr's average, in a range of $27,379 to $222,030, and Vendr does not say whether that includes implementation fees) for a second front door is hard to defend when SAP has announced its own Ariba Intake Management agent. Buy it if you have no suite, or if you can show that requests are bypassing the suite you have.
Fairmarkit fits high-volume, comparable buys where several suppliers can quote. Its Sonoco case study reports 9% identified savings and 3.7x first-year ROI, both vendor-published. Skip it if your tail is mostly sole-source or specification-heavy, since an automated quote has nothing to compare. Read the savings section below before you accept its dashboard numbers.
Pactum has the most candid commercial terms of the five: a 12-month minimum, pricing by revenue band, and no setup fee, savings percentage, per-user or per-negotiation charge. Its pilot needs about 150 negotiations, ideally drawn from 250 to 300 suppliers, so skip it if you have fewer suppliers than that in scope. Also skip it if your supplier relationships are few and strategic, because a bot opening a negotiation with a sole-source partner is a poor trade for what you could lose.
Pactum's headline claims need arithmetic. It reports $21M+ in discounts across 50,000+ negotiations for 50+ enterprises, which works out to about $420 a negotiation if you divide those floors. That sum covers only discounts and early-payment discounts, so it leaves out the value of longer payment terms. Pactum's example is that 10 more days on $1B of spend frees about $27M. That is cash, which a CFO values at cost of capital, so at an illustrative 5% it is worth about $1.35M a year and never appears as a cost reduction.
Coupa or Ariba: When a Suite's Built-In AI Is Enough
A suite's built-in AI is enough when your gap is automation of work already inside the suite, such as drafting sourcing events, comparing bids or checking invoices. Coupa says more than 450 customers run its agents in production. The count of agents depends on who is counting: more than 50 released across early access, limited availability and general availability in August 2026, against 20 in production and a target of 65 by January 2027 reported after Coupa Inspire in May. Ask for your tenant's list by status.
Pricing changes set a deadline. Coupa's CEO told diginomica in July that AI usage would be effectively unlimited until year-end, then move from seats to outcome-based pricing built on a spend threshold agreed with each customer and reset at renewal. That article does not define who qualifies, and Incisiv's May recap of Coupa Inspire described a flat AI rate for new clients and upgrades, so confirm your contract is covered. SAP's position is harder to read. ERP Today reported in June that Joule Agent Runtime carries zero incremental cost for current Ariba and Fieldglass customers through December 31, 2026. SAP's October TechEd announcement limits the free runtime to custom agents, for select customers, and a June SAP commercial FAQ lists Spend Agents under AI Units, metered per action. Treat SAP's own Ariba agents as metered until your account team says otherwise in writing. You can measure Coupa before the new pricing starts, if your contract qualifies. On Ariba, ask what the test will cost. Coupa's CEO claims customers get $50 back for every $1 spent. That is a vendor estimate, and it matters more once pricing depends on value, because some analysts have pushed back on how measurable value-based pricing is.
Coupa is the wrong pick for a team whose only problem is request intake, or one without the staff to run a multi-module rollout. Coupa's partner CrossCountry's pricing guide lists licensing, implementation services and provider selection as separate cost levers, so the subscription is only part of what you pay.
The loser here is Ariba as an AI purchase. If you already run it, use its agents. If you are choosing from scratch, its AI story gives you no reason to choose it. SAP planned its Intake Management agent for June 2026 general availability on next-generation Ariba, itself generally available only since March 2026, and it plans to back-port at least 60% of agentic capabilities to the current generation, which leaves up to 40% that may not arrive there. We found no SAP release note confirming the June date was met, so ask. As of October 2025 SAP had not said whether customers would be migrated automatically, so confirm that today. Your suppliers also pay network charges of 0.155% or 0.35% of qualifying volume, capped at $20,000 a year per buyer relationship, per an undated SAP training page. Expect some suppliers to try to recover that in price.
How Long Does AI Procurement Implementation Take?
Published implementation times are almost entirely vendor-reported, and they measure different things. Compare them only after asking what "live" means.
- Fairmarkit's figure of about 40 days comes from Fairmarkit's integrations page, which says most customers go live in roughly 40 days and elsewhere that technical integrations can finish in 40 days. Those are two different claims. Fairmarkit's SAP partner listing says weeks, not months, and its Sonoco case study reports two weeks, called a record. We found no independent measurement.
- Pactum's 2 to 4 weeks is its own figure for a pilot from a CSV upload, with no connector. A pilot is a campaign, so it tells you little about a rollout across categories.
- Coupa runs 2 to 8 months in the sources we could trace. A Coupa case study reports e-procurement in two months, a narrow scope. A single PeerSpot reviewer from 2022 gives three to eight months and admits to not taking part in the deployment. Longer figures circulate on rival vendors' comparison pages, and we give those no weight.
- Zip gives no number. Its page says customers get up and running quickly. Ask for a dated reference.
Where Savings Claims Come From and How to Baseline Them
Most AI procurement savings figures measure a gap against a baseline the vendor's system helped define, so they should not be compared across vendors. Zip claims $6 billion in customer savings, and we found no published method behind it. Fairmarkit's documentation is the clearest example, and credit to it for publishing the method: identified savings are baseline minus lowest bid, and when no benchmark is provided the baseline is the average of supplier responses, with outliers excluded. A lowest bid sits below the average of bids by arithmetic. Prices could rise 10% from last year and that event would still show savings. Supply a real historical benchmark on every event, and report awarded savings instead of identified savings.
Compare that with the Walmart numbers on one workload. Assume $1B of addressable spend, $100M of it tail across 1,500 suppliers (illustrative figures, with agreement rates assumed to apply evenly across spend). The pilot's rates of 64% and 1.5% imply about $960,000 a year. The later rates of 68% and 3% imply about $2.04M. Multiply Sonoco's 9% by the same $100M and you get $9M, a number that means little because it uses identified savings on a different baseline. Set these rules before go-live, with finance in the room:
- Freeze the baseline. Use the last 12 months of actual spend, and do not use a budget that already assumes savings.
- Count only invoices. Negotiated savings are a leading indicator, and realized savings begin once new terms appear on purchase orders and invoices. Suplari's definition adds verification against the general ledger.
- Track leakage. Umbrex defines it as the gap between negotiated terms and what gets purchased. Set a compliance floor before anything counts as realized.
- Keep price reduction, cash from longer terms and cost avoidance in separate columns.
- Hold out a control group. Randomly exclude a share of tail suppliers for two quarters. Prices drift on their own, and this is the only way to tell what the tool caused.
What to Ask For in the RFP When No One Publishes Prices
No published list price exists for any of the four named vendors that we could find, so your RFP has to force comparable numbers. Vendor quotes are structured differently: Pactum uses revenue bands and digital workers, Zip uses a platform fee, seats and add-on modules according to Vendr, and Coupa's new model uses a spend threshold. Fairmarkit's tail-spend page offers a demo request and no pricing. We could not load a pricing page for Coupa, and Vendr puts it at $94,519 a year, in a range of $20,804 to $331,084. Vendr's data is a marketplace sample, useful as a sanity check and weak as a benchmark.
Ask every bidder to price the same defined workload, and to answer these in writing:
- Three-year total cost for $1B addressable spend, 1,500 tail suppliers and your actual ERP count, including implementation and integrations.
- What triggers a price change at renewal, and whether AI usage is metered after any promotional period.
- The exact savings formula, including every fallback when no benchmark exists.
- Which agents are generally available on your tenant today, listed by name.
- Two references live for more than 12 months, one of which has expanded beyond its original scope.
- What happens to your negotiation history and supplier data at exit.
What Predicts Regret on a Procurement AI Purchase
Three conditions change the recommendation more than any feature list does. The first is spend concentration: if your top 50 suppliers hold almost all of it, a tail tool has little to work on, and you should put the money into category strategy and let the suite handle the rest. The second is master data quality, which decides whether a bot negotiates from facts. The third is whether finance will sign the baseline. If finance will not, every vendor's number stays a marketing claim, whichever tool you pick.
The answer also changes with size. With fewer than about 150 suppliers in scope, Pactum has nothing to pilot on. With no suite and a single ERP, start with Zip and reopen the suite question when sourcing and contracts outgrow it.
This Week:
- Pull 12 months of invoice data and split spend into strategic, mid and tail, by supplier count and dollars.
- Get your finance lead to agree the baseline rules above, on paper.
This Month:
- Ask your Coupa or SAP account team for your tenant's list of generally available agents and what they will cost in 2027.
- Pick one tail category and run the suite's agents on it, recording cycle time and invoice price against the frozen baseline.
Before December 31, 2026:
- Get 2027 AI pricing in writing while the current terms and your renewal give you leverage.
- Issue the RFP above only if the suite test left a measured gap, and only for that job.
The Bottom Line
Buy the suite's AI first if you own a suite, Zip if you have no front door, Fairmarkit where suppliers can bid against a real benchmark, and Pactum where large groups of suppliers sit on terms nobody negotiated. The evidence base for all of it is thin: one well-documented retailer pilot from 2021, plus vendor case studies.
Have finance sign that baseline before the first supplier gets an email from a bot.
Continue Reading
- Docusign Acquires Suplari to Link Spend and Contracts
- Your ERP Decides Whether Stampli, BILL or Tipalti Fits
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
- Vendor AI Benchmarks Need Independent Runs With Task Logs
- Five AI Vendors Give Customers 0 to 90 Days to Get Data Out
- AI Vendor Security Review: 6 Questions That Change the Answer