- Home
- AI agents
- Enterprise AI agent examples
- Procurement
Plan · Enterprise AI agent examples
AI agents in procurement: what they do and what companies report
A procurement agent checks purchase requests against policy before a buyer sees them, catches duplicate vendors, answers requester questions and, within a mandate the procurement team sets, negotiates routine terms with suppliers. It works inside the procure-to-pay system the company already runs and routes anything outside policy to a buyer.
In this guide: 83 pages
Key takeaways
- Three procurement deployments are published with figures, all as vendor stories: SUEZ UK’s requisition checks on Coupa, Honeywell’s supplier negotiations and the intake agents OpenAI’s procurement team runs on Zip.
- The published agents work the long tail that buyers had no time for. Honeywell’s agents negotiate spend buyers could not cover before, and SUEZ’s agent checks every requisition as it is raised.
- A negotiating agent needs a written mandate: the terms it may offer, the limits it may not cross and the suppliers in scope. The mandate is its main control, so version it like a policy.
- Keep requester, approver and payer as separate identities. An agent that can raise, approve and pay a request removes the segregation of duties auditors test.
What do AI agents do in procurement?
They check requisitions against policy and correct them, stop duplicate vendors at intake, answer requester questions about orders and invoices, look up supplier risk, and negotiate payment terms, rebates and discounts with suppliers inside a mandate. Each task runs at a different level, with a different control.
Procurement tasks by what the agent is allowed to do. "Assists" means it drafts or flags and a person acts; "acts" means it changes something itself.
| Task | Level | What it touches, and the control we add | Seen in |
|---|---|---|---|
| Requisition policy check: category, quotes, rate cards | Acts | Corrects the requisition in the procure-to-pay system; exceptions go to a buyer | SUEZ UK (Pactum on Coupa) |
| Duplicate vendor check at intake | Acts | The vendor master; flags a duplicate before a new vendor is set up | OpenAI (Zip) |
| Purchase orders, vendor setup and invoice questions | Acts | The procure-to-pay system, as the requester | OpenAI ("Penny" on Zip) |
| Third-party risk lookup | Assists | Supplier data; flags risks for a person to judge | OpenAI ("Risk Buddy" on Zip) |
| Negotiation of payment terms, rebates and item discounts | Acts | Supplier conversations, inside a written mandate | Honeywell (Pactum) |
Most of the published value sits in volume a team would never cover by hand. Honeywell reports more than 2,500 autonomous negotiations, and OpenAI says its first intake agent took over a duplicate-vendor check that used to cost 15 to 20 minutes per request.
What have companies published about procurement agents?
Three companies have published figures, each through the vendor that sells the agent, so read them as claims rather than audited results. The table gives each figure as the source states it, who published it and what it counts.
Publicly reported procurement agents, as published by each source and opened on 2 November 2026. None of these companies is a QueryNow client.
| Company | What the agent does | Published figure | Source and date | How to read it |
|---|---|---|---|---|
| SUEZ UK | Pactum’s Alignment Agent sits on SUEZ’s Coupa system, intercepts every purchase requisition, checks it against policy and corrects it where needed | Requisition cycle time down 90% ("days to minutes"); 80% of manual reviews removed; more than 10% of buyer time freed for strategic sourcing | Pactum customer story, undated, read 2 November 2026 | Vendor claim; SUEZ UK’s head of procurement is quoted |
| Honeywell | Pactum agents negotiate payment terms, rebates and item discounts with suppliers, including spend buyers could not cover before | More than 2,500 autonomous negotiations; over 80% supplier engagement; more than half a billion dollars of spend managed through Pactum | Pactum customer story, undated, read 2 November 2026 | Vendor claim; Honeywell’s chief procurement officer is quoted. Spend managed measures volume, and the page gives no saving for it |
| OpenAI (procurement team) | Zip agents: intake validation for duplicate vendors, "Penny" for purchase orders, vendor setup, invoices and questions, and a custom "Risk Buddy" for third-party risk; more than 10 further agents | 1,400 hours saved with the first agent; the page calls this annual in one place and gives no period in the headline | Zip customer story, undated, read 2 November 2026 | Vendor claim; OpenAI’s vice president of procurement is quoted |
One procurement story stays out of the table: Siemens’ supplier onboarding with Salesforce is described in the future tense ("will handle") in the September 2026 release, so it is a plan. We count an example only when the source describes it running.
What controls does a procurement agent need?
A procurement agent needs a mandate for anything it negotiates, limits that sit outside the model, separate identities for raising, approving and paying, disclosure to suppliers, and an audit trail. Most of these already exist for buyers; the work is applying them to a non-human actor.
- Mandate. For negotiation, write down the terms the agent may offer, the floor it may not cross, the categories and suppliers in scope, and when it must hand over to a buyer. The procurement lead owns the mandate and approves each version.
- Limits. Keep spend and category thresholds in the procure-to-pay system or a policy engine, outside the prompt, so a change of limit is a configuration change with a named approver.
- Segregation of duties. The agent may raise or correct a requisition; approval and payment stay with other identities. Give each agent its own identity: a Microsoft Entra Agent ID on Microsoft, an AgentCore Identity on AWS, an Agent Identity on Google.
- Disclosure. Tell suppliers they are dealing with an AI agent and give them a route to a person. Under the EU AI Act, the Article 50 transparency duty for AI that interacts with people has applied since 2 August 2026.
- Audit trail. Log every change the agent makes to a request and every offer it sends, with the policy rule behind it. Auditors will sample these the way they sample buyer actions.
Two failure modes are easy to predict. Rules copied from a policy document nobody follows make the agent correct requisitions buyers would have let through, and requesters route around it. A mandate written into a prompt rather than kept as versioned configuration leaves nobody able to say which limits applied to a closed deal. Both are fixed by keeping rules and mandate as data with an owner and a version number.
Where should you start with a procurement agent?
Start with the requisition check on the procure-to-pay system you already run, because it works on your own documents and no supplier sees a mistake. Supplier-facing negotiation comes later, once the rules have a track record.
- Pull a quarter of requisitions and count how many a buyer returned or corrected, and why. Those reasons become the agent’s first rules.
- Run the check in shadow mode for a month: the agent proposes corrections and a buyer accepts or rejects each one. Track acceptance by rule.
- Switch on automatic correction for the rules buyers nearly always accept, and keep the rest as suggestions.
- Add the duplicate-vendor check at intake, where OpenAI’s team started.
- Pilot negotiation on one tail category with a written mandate, low value per supplier and a buyer reviewing every closed deal for the first month.
How do Microsoft, AWS and Google implement it?
The three published examples run on procurement vendors’ products: Pactum on Coupa, Pactum and Zip. This is what each cloud offers if you build your own, with statuses as the vendor pages showed on 2 November 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Build the agent | Copilot Studio with Power Platform connectors, or Microsoft Foundry Agent Service | Amazon Bedrock AgentCore Runtime and Gateway (GA 13 October 2025) | Gemini Enterprise Agent Platform with ADK and Agent Runtime |
| Limits on each tool call | Foundry tool call guardrails (preview); per-tool approval in Microsoft Agent Framework | AgentCore Policy, Cedar rules on every tool call (GA 3 March 2026) | Agent Gateway as the policy point (GA 18 June 2026) |
| Separate identity per agent | Microsoft Entra Agent ID (GA 1 May 2026), created for every new Copilot Studio agent since July 2026 | AgentCore Identity workload identities | Agent Identity (SPIFFE-based); API in preview since 18 June 2026 |
| A person approves | Approvals in Copilot Studio agent flows; multistage approvals in preview | AgentCore Policy blocks the calls a rule can decide; the rest go to an approval step in your workflow | ADK tool confirmation (experimental) |
The procurement suites hold the requisition and supplier data, so buying the agent with the suite is the shorter route. Govern it with the identity, policy and logging you already run.
Frequently asked questions
Can an AI agent negotiate with suppliers on its own?
Honeywell reports more than 2,500 autonomous negotiations through Pactum, so it is in production at one large buyer. Every such agent works inside a mandate: the terms it may offer and the limits it may not cross. Keep a buyer reviewing closed deals until the outcomes have a track record.
Which procurement task should an agent take first?
The requisition policy check. It works on your own documents, a buyer can review its suggestions in shadow mode, and the rules come from corrections your team already makes. SUEZ UK’s published figures come from exactly this task.
Do suppliers need to know they are talking to an AI agent?
In the EU, yes: the AI Act’s Article 50 transparency duty for AI that interacts with people has applied since 2 August 2026. Outside the EU we recommend it anyway, with a clear route to a buyer, because a supplier who finds out later stops trusting the terms.
How do you measure a procurement agent?
Count what the published examples count, against your own baseline: requisition cycle time, the share of requisitions a buyer touches, corrections accepted by rule, and for negotiation the terms achieved against the mandate’s target. SUEZ UK reports cycle time and manual reviews removed, two numbers a finance team can check.
Who owns a procurement agent?
The head of procurement or a category lead owns its behaviour, its mandate and its limits; IT owns its identity and connectors. Write both names into the registry record so an incident reaches someone who can change the rules.
More in Enterprise AI agent examples
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- How Pactum’s AI agents support SUEZ in eliminating manual requisition checks
- How Honeywell is running autonomous procurement at enterprise scale
- Building AI-first procurement at OpenAI with Zip
- Siemens and Salesforce deepen AI partnership to redefine industrial sales and service
- AI Act: regulatory framework for AI
- Policy for Amazon Bedrock AgentCore is now generally available
- Amazon Bedrock AgentCore is now generally available
- Gemini Enterprise Agent Platform release notes
- What’s new in Copilot Studio
- What’s new in Microsoft Entra Agent ID
- Amazon Bedrock AgentCore key features and benefits
- Agent Identity overview
- What is Microsoft Foundry Agent Service?
- Guardrails and controls in Microsoft Foundry
- Tool approval in Microsoft Agent Framework
- Tool confirmation (Google ADK)
- Multistage and AI approvals in Copilot Studio agent flows
- Use connectors in Copilot Studio
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner