If you lead procurement, you have probably been asked some version of this question in the last year: how much of this can the AI just do? It is a fair question, and the honest answer is that it depends on the workflow, not on the AI. The teams getting value from procurement AI agents decide autonomy one workflow at a time, and they move up only when the results earn it.
We call that approach the trust ladder. It has three rungs: manual, rule-based and hands-off. This post explains each rung, how to decide where a workflow belongs and which controls should stay in place at every level.
Autonomy decides who carries the risk when something goes wrong. A procurement AI agent that drafts a summary carries almost none. One that approves a contract carries a lot. The technology is the same. The consequence is not.
Gartner said in July 2025 that generative AI for procurement had entered the Trough of Disillusionment, with many organizations seeing uneven ROI. The obstacles it listed were fragmented, low-quality data, complex integration and skepticism about AI-driven insights. Those are trust problems as much as technology problems.
According to the 2025 Gartner Procurement Organization Design for the Future Survey, 67% of procurement organizations are discussing or planning to implement agentic AI. However, 33% are dissatisfied with the ease of implementation, and 21% worry about the rework it would introduce.
Interest is high and confidence is not. The gap closes when teams can see what an agent did, why it did it and how to stop it.
Each rung gives the workflow more room to act on its own. Each one also keeps a clear place for people.
| Rung | What runs | Good fit |
|---|---|---|
| Manual | People review, decide and act. The platform records each step and suggests answers marked for review. | New workflows, rare cases, anything with poor data |
| Rule-based | Conditions, approval brackets, assign-only-if rules and reminders route the work the same way every time. | Policy you can write down, such as approval limits and renewal reminders 90 days before renewal |
| Hands-off | Agent steps run when the step is reached, decide what they found and route on the result. People handle exceptions. | High-volume checks with clear outcomes, such as contract compare and playbook checks |
People do the work, and the platform makes it visible. Requests, answers, approvals and documents live in one place. AI helps without deciding: Opstream can suggest answers from an uploaded quote, and those values stay marked as suggestions until a person accepts them. Ask Opstream answers questions about requests and vendors, and it shows which tools it is using while it answers.
You write down the decisions people repeat, and the workflow follows them. Approval brackets set who signs at each amount. Conditions add Legal or Security only when the request needs them. Reminders start a renewal review a set number of days before the contract ends. Rules are predictable, which is why most procurement policy belongs here.
Agent steps run when the workflow reaches them and route on what they find. In Opstream, a Compare step picks the documents to compare, such as a renewal against the previous agreement, and reports changes by severity. A Playbook step checks a contract against your organization’s positions for that document type. The step outcome drives the next step: high severity goes to Legal, none continues.
Score each workflow on three questions. The answers tell you where to start.
A workflow with clean data, reversible outcomes and a clear rule can move up fast. A workflow that fails any one of the three should stay lower until it passes. This is why the data model comes before autonomy, not after it.
Some controls never go away, however much a workflow runs on its own.
Agents are digital workers with decisioning + execution capability. They need identity, permissions, supervision and accountability.
Source: The Hackett Group, “Agentic Governance and Controls Primer,” Kyle McNabb, Ammar Mohammed, June 2026.
Move one rung at a time and measure before each move. These six steps work for most procurement teams:
The trust ladder applies to every procurement use case. Contract review often reaches hands-off first, because a comparison has a clear result.Third-party risk at intake usually stays rule-based, with a person approving every high-risk vendor.
Yes, and it fits the same ladder. The Opstream MCP server connects an AI assistant your company already uses to your Opstream data. It answers within each user’s permissions. When a request is created through it, the same checks run as in the app, including the check for software you already own. Your assistant gets context, and your controls stay where they are.
The trust ladder is a way to decide how much a procurement workflow runs on its own. It has three rungs. Manual means people do the work and the platform records it. Rule-based means conditions you set route the work. Hands-off means agent steps run and route on their own result, with people handling exceptions.
Only for low-risk, well-defined cases with clean data and written rules, such as a renewal with no material changes under a set amount. High-risk vendors, large contracts and policy exceptions should always reach a person. Opstream lets admins choose which steps can auto-approve and which always need a reviewer.
Give every agent step a clear scope, show the evidence behind each finding and log every action. In Opstream, agent steps show their findings on the task card with a link to the full report, users can change the inputs and re-run, and audit logs record who or what did each action.
Rule-based automation follows conditions a person wrote, such as routing requests over a set amount to Finance. Agentic AI reads the request and its documents, decides what it found and acts on that, such as comparing a renewal with the previous contract and routing on severity. Most teams need both.
Yes. The Opstream MCP server lets an AI assistant your company already uses query your Opstream data, such as requests, vendors, spend and documents, within each user’s permissions. Requests created that way run the same checks, including the duplicate software check, as requests created in the app.
Start with an advisory step on a workflow you already run in the platform. The Auditor step is a good first choice because it only suggests corrections and never edits request data. Compare its findings with what reviewers decide for a few weeks, then let it route the clear cases.
We will map one of your workflows to the three rungs and show where agent steps fit today.
Book a Meeting