Among 17 AI procurement products reviewed, Pactum has the clearest named production evidence for autonomous negotiation; Keelvar for repeatable sourcing; and Kavida/QAD for post-PO coordination. No vendor has publicly demonstrated a routine purchase running from need through supplier choice, commitment, payment, delivery verification, and outcome learning without regular human control.
That conclusion comes from a public-evidence review of current product documentation, launch material, and named vendor customer stories available through August 13, 2026. We did not test the products or audit private deployments. The comparison asks one narrow question: what procurement action can the software take under delegated authority, and what still requires a person?
The four levels of procurement autonomy
Recommendations, summaries, classifications, and chat answers can save real time. For this comparison, a product reaches autonomy only when it can take an external or system-of-record action: contact a supplier, launch an RFQ, negotiate terms, approve a low-risk request, create a PO, update an ERP, resolve an invoice exception, or route a standard invoice to payment.
| Level | Test |
|---|---|
| Assistive | Produces insights, drafts, classifications, or recommendations. A person acts. |
| Workflow automation | Moves data and tasks through predefined rules. A person makes the material commercial decision. |
| Bounded autonomy | Reasons and acts inside policies, thresholds, approved suppliers, or a defined playbook. People handle approvals, exceptions, and strategy. |
| Closed-loop autonomous purchasing | Owns the normal path from business need to verified outcome, including commitment and payment. People set policy and handle rare exceptions. |
Evidence grades measure the quality of public proof:
- A — named production evidence: a named customer plus a specific autonomous action or result.
- B — product and adoption evidence: detailed current documentation plus production adoption, without a named action-level case.
- C — vendor claim: a current product page, demo, or announcement without a named production example for that action.
- D — ambiguous: conflicting directory data, generic AI language, or inadequate action-level evidence.
AI procurement tools compared
| Product | Best-supported level | Publicly supported actions | Routine human boundary | Evidence |
|---|---|---|---|---|
| SAP Joule and procurement agents | Bounded autonomy in selected workflows | Intake routing, sourcing preparation, source assignment, goods receipt, invoice tasks | Governance, approvals, and staged product availability | A/B |
| IBM watsonx Orchestrate | Documented bounded-autonomy capability | Requisitions, approvals, POs, goods receipts, suppliers, contracts, invoices, RFP evaluation | User requests, deployment configuration, enterprise permissions | C |
| Zip | Bounded autonomy in selected workflows | Intake pre-clearance, supplier outreach, response steps, PO posting, payout scheduling, payment triggers | Configurable approvals, checkpoints, and exception escalation | A/C |
| Beam AI | Bounded-autonomy claim | Supplier identification, RFQs, quote comparison, PO generation, order tracking | Approval workflows and human monitoring | C |
| Blackbee AI | Bounded-to-closed-loop claim | Intake, low-risk approvals, POs, vendor communication, invoice matching, ERP posting, payments | Policies, thresholds, overrides, high-risk cases | C/D |
| Aerchain | Bounded-autonomy claim | Intake, supplier evaluation, negotiation, onboarding, contracts, invoicing | Approval matrices, configurable workflows, compliance guardrails | C |
| Keelvar | Bounded autonomy | Request-to-sourcing event, supplier engagement, bid analysis, award recommendation | Customer rules and optional award approval | A |
| Pactum | Bounded autonomy | Multi-round supplier negotiation and requisition alignment | Procurement-defined playbooks, authority limits, strategic exceptions | A |
| Procol | Workflow automation; bounded-autonomy claim | Supplier discovery, RFQs, auctions, PR-to-PO conversion, approvals, renewals | Supplier selection, negotiation judgment, final decisions | B/C |
| ProcurEngine | Workflow automation | RFx, auctions, comparisons, approval routing, supplier records | Buyers configure events and negotiation strategy; approvers decide | A/B |
| Qvalia | Assistive | Spend and invoice analysis, classification, compliance insights, forecasting | People interpret insights and act | B |
| Coupa | Bounded autonomy in selected workflows | Smart intake, agent orchestration, sourcing and supplier tasks; broader external-system integrations remain staged | Governance, role access, approvals, intervention | B |
| QAD ChampionAI / Procurement Champion, formerly Kavida | Bounded autonomy | Supplier follow-ups, confirmations, document requests, PO-change detection, confirmed ERP changes | Buyers manage exceptions and material decisions | A |
| PRMAI / PRM360 | Workflow automation proven; broader autonomy claimed | Claims source-to-payment orchestration | Agent authority and production behavior remain inadequately documented | C/D |
| Spendflo | Bounded autonomy, early production claim | Standard intake-to-PO, contract checks, invoice matching, payment routing | Policy approvals, exceptions, strategy, negotiations, risk calls | B/C |
| Tendios | Assistive to workflow automation | Tender discovery, document generation, validation, alerts, repetitive tender tasks | Public officials and bidders make and submit decisions | B |
| Vertice | Bounded-autonomy claim; deployment-dependent | Drafted negotiation emails plus claimed automatic countering or direct negotiation for supported renewals | Supported vendors and contract types, buyer guardrails, expert negotiators | C |
The qualitative levels are intentionally coarse. A higher level does not make a product safer, more accurate, or better for every category. It says only that the public record supports more delegated action.
The three strongest named production cases
Pactum: autonomous supplier negotiation
Pactum gives negotiation agents procurement-defined goals, acceptable ranges, and fallback rules, then lets them negotiate directly with suppliers. Its Honeywell case describes more than 2,500 autonomous negotiations across nearly $500 million of spend, including a $65,000 saving that required no buyer involvement. Pactum’s customer results also report Walmart outcomes and its SUEZ case describes requisition alignment inside Coupa.
This is the clearest evidence in the set that an agent performs an external commercial action at production scale. Procurement teams still define the playbook, authority range, acceptable outcomes, and escalation rules. The evidence covers negotiation and alignment rather than the entire source-to-pay and delivery loop.
Keelvar: repeatable sourcing events
Keelvar documents an autonomous sourcing flow that can accept an approved request, select a workflow, build and launch an event, engage suppliers, run bidding rounds, analyze bids, and produce an award recommendation. The product documentation describes request-to-award automation for repeatable tail, spot, logistics, and MRO spend. Its Nissan story shows automated bid analysis while buyers retain business judgment over the award.
The supported boundary is meaningful: the agent can run a sourcing event, while customer rules and approvals govern the commercial commitment. Public evidence does not extend through contracting, payment, delivery, and outcome verification.
QAD ChampionAI, formerly Kavida: supplier coordination after the PO
QAD’s Procurement Champion, formerly Kavida Agent PO reads supplier email, follows up for confirmation and status, requests missing documents, identifies changes, records confirmed ERP changes, and escalates exceptions. The current QAD Dyer Engineering case describes automated supplier follow-ups, document handling, and confirmed ERP changes. QAD acquired Kavida in 2025. A later QAD executive post reports more than $1 billion in AI-managed orders; that is unaudited vendor-reported processed order value and does not mean the agent selected or committed that spend.
That is strong execution evidence within a defined stage. The agent begins after a buyer or ERP has created the PO, so people have already exercised much of the purchasing authority.
Real agent infrastructure with a narrower public record
SAP, IBM, and Coupa
SAP shows selected task autonomy across Joule, Ariba, and spend management. Its Buying Assistant can orchestrate specialist agents, while SAP’s procurement AI overview covers sourcing and spend workflows. A named SAP customer example describes a custom procure-to-pay agent automating goods-receipt work and accelerating invoice clearance. SAP’s 2026 availability guide says sourcing, procurement-contract, and invoicing assistants were planned for June, with more procurement assistants planned for September. We found no official release note confirming that the June roadmap shipped.
IBM watsonx Orchestrate publishes one of the broadest action catalogs. Its procurement agent documentation spans requisitions, approvals, purchase orders, goods receipts, suppliers, contracts, invoices, and three-way matching across Coupa, SAP, Ariba, and Oracle. The catalog supports bounded action technically. IBM’s named Dun & Bradstreet Ask Procurement case covers an assistive supplier-risk pilot rather than these transactional agents in production, so the evidence remains grade C under this rubric.
Coupa combines smart intake, orchestration, and an agent studio. Its 2026 Compose announcement marked third-party integration, MCP, and A2A support as expected in September 2026. Coupa’s June adoption update reports more than 350 customers with Navi agents in production across the portfolio. That vendor-reported count supports adoption, while it does not represent 350 closed-loop buying deployments. The public record lacks a named customer case where Coupa selects a supplier, commits, pays, and validates delivery without routine human intervention.
Spendflo and Vertice
Spendflo’s Flo AI launch describes agents that classify requests, check budget and policy, onboard vendors, raise POs, examine contracts, match invoices, and route standard invoices toward payment. The stated operating model keeps strategy, negotiation, risk, and exceptions with people. The launch is recent and the public material does not name a customer result for its touchless flow.
Vertice’s Ana uses contract context and vendor-pricing data to plan supported renewal negotiations inside buyer-defined thresholds. Its product page alternates between a user sending drafted emails and automatic countering, while the Vendr acquisition announcement claims direct negotiation. Those public descriptions may reflect different configurations or staged availability. Vertice also uses human experts for high-stakes work, and no named Ana customer result is public.
Broad autonomy claims that need action-level customer proof
Beam AI’s procurement agent claims supplier discovery, RFQ distribution, quote comparison, supplier recommendation or selection, PO generation, and order tracking. Approval routing and human monitoring remain part of the design, and its public case-study library does not identify a procurement customer using that full flow.
Blackbee AI describes agents spanning intake, low-risk approvals, vendor communication, POs, invoices, ERP synchronization, and payment. Its Procurement & PO Agent offers detailed workflow claims, while named production evidence for those actions is absent from the public record we reviewed.
Aerchain presents agents for negotiation, supplier evaluation, onboarding, contracts, and invoices. Named customer material supports adoption of the broader procurement platform; it does not isolate a live purchase or negotiation completed by an agent under delegated authority.
PRMAI is a 2026 rebrand of the established PRM360 source-to-pay workflow product. Current positioning promises AI-native orchestration from supplier discovery through payment. Public documentation does not yet define the agents’ approval logic, transactional authority, or a named customer deployment at that broader level.
These products may perform more inside private customer environments than their public material proves. Buyers should ask for action logs, eligibility rates, exception rates, and a reference customer rather than treating a workflow diagram as production evidence.
Valuable procurement automation that still leaves the commercial decision to people
Zip says it has launched agents for intake validation, risk review, contract work, document comparison, and RFP content. Its own agentic procurement explanation describes a human-run model in which machines initiate and prepare work. That is sophisticated orchestration with people retaining decisions, approvals, and negotiations.
Procol automates RFQs, auctions, approvals, supplier onboarding, reminders, and PR-to-PO conversion. Its current Clara page claims autonomous PR-to-PO, while concrete mechanics still include RFPs prepared for sending, one-click counteroffer suggestions, and award allocation with human review. ProcurEngine has customer evidence for digital RFx, auctions, counteroffers, and approval routing; its autonomous-negotiation page says people review and send every communication. Both remove substantial manual work while keeping buyers or approvers in charge of material decisions.
Qvalia’s Clarity agent analyzes line-item invoice, order, accounting, price-list, and contract data to surface discrepancies and off-contract spend. Tendios finds tenders, generates and validates documents, and supports public-procurement workflows. Their current public products help people understand and execute procurement work without independently committing spend.
Why no product qualifies as closed-loop autonomous purchasing
The 17 vendors collectively cover nearly every step: intake, supplier discovery, sourcing, negotiation, contract review, POs, supplier follow-up, invoice matching, and payment routing. The missing public proof sits at the seams between those steps.
A closed loop would need to handle an unstructured business need, choose among suppliers, negotiate within legal and financial authority, create the commitment, pay, verify that the product or service was delivered and useful, and use that outcome in the next decision. Public case studies rarely report touchless eligibility, exception rates, rollback rates, authority thresholds, delivery quality, or outcome learning.
The likely near-term model is human-governed, exception-based purchasing: people define policies, budgets, approved suppliers, and escalation rules; agents execute routine transactions; people take strategic, novel, or high-risk cases. That can create substantial value without pretending that an agent owns the whole commercial relationship.
Ten diligence questions for an AI procurement vendor
- What external or system-of-record actions did the agent execute last month?
- Which actions can occur without a human click?
- What dollar, category, supplier, and contract thresholds bound its authority?
- What percentage of eligible transactions completes without intervention?
- How are supplier-selection and negotiation outcomes evaluated?
- Does the agent create a legal or financial commitment, or prepare one for approval?
- Can it verify receipt, service delivery, and quality beyond matching documents?
- What are the exception, rollback, and incident rates?
- Can a named customer confirm these actions in production?
- Does every action have an inspectable audit trail showing data, policy, and authority?
Those questions also reveal a gap inside most procurement stacks: evidence about how a supplier performed across companies. A buyer can inspect Talkshi’s public profiles for Pactum, Keelvar, Coupa, Zip, SAP Joule, and IBM watsonx Orchestrate, then submit a first-hand outcome through the plain HTTP review contract. The Agent Signup Index separately tracks whether 50 software products let an email-only agent obtain its own API key.
Scope and caveats
- This comparison reviews public evidence available from July 29 through August 13, 2026. Product behavior can differ by configuration, geography, ERP, contract, and early-access status.
- Vendor-published named case studies are stronger than anonymous claims, while remaining different from independent audits.
- “Autonomous” has no consistent vendor definition. Some vendors apply it to analysis, some to triggered workflows, and some to external actions under delegated authority.
- Evidence grades assess the proof behind an autonomy claim. They do not score security, procurement breadth, user experience, price, or overall product quality.
- Tell us about a missing action-level case or a changed product boundary through the review API. Keep first-hand experience separate from vendor marketing claims.