How Omnea builds agents

There's a version of autonomous procurement that feels a lot like babysitting: a preschool of enthusiastic agents who tackle every task with total confidence and not enough context, producing work you then have to check line by line. Plenty of vendors ship that version today, and it's no one's dream workplace.

Here's what the other version looks like. At Adecco, a typical procurement cycle used to take 23 days. Today it takes 4, facilitated by agents that clean, cross-check and route the request before it reaches approvers.

That's the version we're building at Omnea: agents that start as capable colleagues and grow into real specialists. This article explains how that works, and what we've learned while building a system where multiple agents work together on real procurement workflows.

01 — Definitions

Automation, workflow, agent

Automation is a rule: if the invoice is under $5,000, auto-approve. It does what it's told, forever, whether or not the rule still makes sense.

A chained workflow is automation with more steps: a fixed sequence of prompts where each output feeds the next. A great many things marketed as agents today are this. A workflow can't see that Supplier A needs a sanctions check, or that Request B duplicates an agreement you already hold.

An agent decides what information it needs, which systems to consult, and what to do next, then produces something that moves the process on: a request drafted for submission, a risk escalation ready to decide on, a contract routed for signing. A chatbot that answers questions about your contracts isn't one, because a person is still driving every step.

The test: if a person has to drive each step, it's a chatbot. If it does the same thing every time regardless of what it finds, it's a workflow.

02 — What separates them

Three things separate agents that deliver from agents that demo

1. What it can see

Too much context makes agents worse. As you load more into a model's context window, its ability to use any of it accurately degrades. Agents need a Goldilocks Zone: not too much, not too little.

Three things determine whether an agent has it. A single purchase can touch six to eight systems and five internal functions, so an agent that can't reach your ERP, CLM, GRC tools and finance queues is working from a partial picture. It needs the data at every level, not just the top: when you buy a security product from Microsoft, there's the supplier, the contracting entity, the reseller, the licence, and the teams actually using it, each with its own approval history. An agent working from a vague summary of "Microsoft" is working from a fraction of the decision. And it needs what the other agents already worked out, because procurement work is cross-functional and so is good agentic work.

At Omnea, all three run through the Context Engine, the part of the aOS that decides what each agent knows at the moment it acts. It draws on your policies and decision history, external signals such as pricing benchmarks from Tropic and sanctions screening from Dow Jones, and the record of every approval, rejection and correction that flows through the system. An agent can't cite a source it invented, because every reference points to something it actually retrieved.

"AI maturity was an explicit criterion in our RFP. A small team managing billions in indirect spend across many countries can't scale by hiring, and Omnea's agents stand out for catching bad data at intake, routing correctly, and reducing the volume of requests that need human intervention."

Kristina Adent, Head of Global Sourcing & Procurement, PayPal.

2. What it's allowed to do

Each Omnea agent operates in one of two modes, Advise or Act. An agent can draft and route a low-risk renewal unattended, while high-value contracts always get flagged for human review.

When an approver overrides a recommendation, or supplies a policy nuance that only applies in an edge case, that correction is captured. An admin then decides whether to act on it, so nothing gets applied in a black box.

3. Whether it gets better

The software wrapped around the model does two things: it sources and filters the data an agent needs, and it gives the agent the tools to act, drafting the request, routing it, updating the record. That combination is what turns a capable generalist into a specialist in your organisation. Without it, every agent starts every task as a new intern.

Because Omnea runs the end-to-end workflow, every decision loops back in, so each agent has more of your history to work from tomorrow than today. After a few cycles your agents have diverged from every other customer's, shaped by your decisions and your controls.

The Approval Agent is the clean illustration, using our own product data. It recommends a decision on each review task, and the approver makes the final call. Across its first few thousand live decisions, it went from matching the human reviewer 86% of the time to +95%.

The results

• €6.24M avoided spend tracked at Adecco in the first 90 days

• 38% cut in software request turnaround at Proofpoint

• 4-6 days taken out of TeamViewer's average request cycle time

"Seeing the impact of our AI procurement transformation project made me regret not starting sooner."

Remi Thomas, CFO, Proofpoint.

03 — In practice

How they work together

By design, Omnea's agents are specialists. Each does a narrow job very well, and the system connects them. Three examples from live customer workflows:

Intake and approvals. A requester uploads a quote as a PDF to work with a new AI consultancy. The Intake Agent drafts a properly routed request. The Duplicate Supplier Agent checks whether a similar supplier already does the same job. The Intake Review Agent catches missing data before an approver sees it. The Approval Agent then inherits a cleaner request plus your decision history. Each agent's work raises the quality of the next agent's input, which is where the compounding shows up most visibly. Intake completion time has dropped from around 27 minutes to around 11, and one fintech's first-pass approval rate climbed from roughly 44% to 92% over rollout.

Risk. The Continuous Monitoring Agent runs checks on every active supplier. When an alert lands, the Triage Agent reads it alongside how you actually work with that supplier, so an alert about a supplier's US operations matters less if your agreement only covers the EU. Findings that need attention are escalated with the rationale attached. At Reach, 70% of risk assessment work is now captured automatically.

Sourcing. During an RFx event, the Scoring Agent evaluates supplier responses, and every case where a human scores differently becomes precedent it references next time. The Supplier Suggestion Agent picks up those exclusion patterns, so suppliers who were never going to pass get flagged at the start rather than a week into assessment.

04 — The limits

Will procurement ever be fully autonomous?

Not completely. Human oversight will always be needed for risk decisions, governance and supplier relationships. Relationships in particular don't automate well: trust, category expertise, the judgment call when a strategic supplier has a bad quarter. Agents can prepare that work brilliantly. They shouldn't make the decision.

At one large enterprise we work with, 61% of new requests come in under $10,000, many with minimal information. Their team called it death by a thousand cuts. That work agents can and should own outright, and Omnea's Form Review Agent now checks and improves those requests automatically.

05 — Ask us this

Four questions to ask any procurement AI vendor

Including us.

1. Does the system get better with use, or just repeat itself? Six months in, what does the agent draw on that it didn't have on day one, and where exactly does my approver's override live?

2. Can your agents see each other's work? Or is each one starting from scratch on context another agent already gathered?

3. How many levels of my data does the system hold? If it's one, ask how it tells my three Microsoft agreements apart.

4. What does the agent deliberately not see? A vendor who answers "we give it everything" hasn't dealt with the selection problem.

A vendor building capable specialist agents will enjoy these questions. One running an agentic preschool will not.

Want to see agentic procurement in action? Meet the agents, and the system that makes them better every cycle. Book a demo →