Same models, different context: why Omnea's agents are ahead of the pack

Executive summary

• Every procurement AI vendor rents the same frontier models. The difference is the system that decides what each agent knows.

• Context is key for agents to perform well: the system providing it should control what agents see, what they can safely ignore, and what they remember.

• Legacy P2Ps remember without understanding. Orchestration tools route data without remembering. Omnea runs the whole end-to-end process and structures the full supplier record, so your agents get better every time they run, based on your decisions. That's the Context Engine.

• Over time, your agents in Omnea are different from everyone else's: continuously tailored to drive better outcomes for how your organisation works.

01 — Same models

The same AI models are used everywhere

Ask a generic agent whether to renew a supplier and it will give you a confident, plausible recommendation. The issue is, it will do that whether or not it knows which teams depend on the tool, what you would pay for something similar elsewhere, or the terms you negotiated last time.

For the isolated agent to get it right, someone would have to assemble all of that context by hand from multiple internal systems and email threads. Nobody does. Which means the decision has always been made on partial information, whether a person or an agent makes it. Getting the full picture in front of whoever is deciding is what was never possible before.

Now, no procurement AI vendor is training its own model - it would make no sense. If they were, they would be spending many millions on it, and it would still be sub-par compared to the leading models from the AI labs. Instead, procurement vendors are all renting the world's most powerful models from those labs, which means the models are not where the difference lives.

What separates one procurement AI from another is the software wrapping the model: what it can reach, what it remembers, and what it can judge to ignore or weigh lightly. An agent that treats a risk flag from five years ago the same as one from last week will perform worse than one that knows the difference.

The models are rented, the context is yours, and the outcomes depend on how that context is structured and accessed by agents.

02 — What went wrong

Where procurement tech went wrong

Legacy P2Ps remember without understanding. Orchestration tools route data without remembering.

Many procurement agents from legacy P2Ps and modern orchestration software look good in demos, but fail to perform in practice. Why is that?

The P2P and S2P suites were built to record and process what was purchased, back when that was the main job. Data hierarchy exists on paper, but that often depends on how well the system was implemented. Commercial context is often mired in attachments, free-text fields, and email threads. They are systems of record that keep everything and understand nothing.

The more modern orchestration tools were built to move requests between internal systems. They own the process without owning the record. Their agents are often little more than chained workflows which gather basic information and guide requests along pre-determined routes. While automations and bots may speed up the process, they don't make your process smarter or your data richer. The same request still gets a different answer depending on who picks it up.

03 — Context Engine

Inside the Context Engine

The Context Engine is the part of Omnea's aOS that decides what each agent knows at the moment it acts. Reaching your data was never the hard part, because deep bidirectional integrations are already our bread and butter. The hard part is knowing what to leave out.

Anthropic's research names the problem: context rot. The more you load into a model's context window, the less accurately it uses any of it. So the Context Engine includes an agentic harness that draws on four sources and filters hard so your agents can focus:

1. Your own context. Policies, thresholds, risk appetite, approved supplier lists, category rules, contract history. Captured at intake through 200+ bidirectional integrations, structured on the way in.

2. External signals. Price benchmarking from Tropic, sanctions and adverse media screening from Dow Jones, and SOC2 reports and pen tests from supplier Trust Centers, and a whole host of other data points from public and licensed sources.

3. Encoded expertise. What we've learned across hundreds of procurement transformations, built into how each agent is configured. No customer's data crosses into another's context (which is important, as we work with the most regulated businesses in the world).

4. The decision record. Every approval, rejection, escalation and correction. The tacit knowledge that lives in the heads of approvers and never made it into a policy document.

A useful parallel: the Context Engine is less like a search engine and more like the person on your team who's been there ten years. A search engine returns everything matching your query. That colleague tells you the three things that matter and reminds you what Legal made you change last time.

04 — Why we pull ahead

Why Omnea's agents pull ahead

Omnea's Agentic Operating System (aOS) adapts to your organisation, structures your data for agents to work with, and improves your operations. The result is a procurement system that drives compounding outcomes. Here's what makes the difference:

1. Your agents learn from your corrections. When an approver overrides a recommendation, or supplies a policy nuance for an edge case, that correction is captured in the Context Engine. An admin decides whether to act on it, so nothing gets applied in a black box. Agents also build on each other's outputs and share context rather than starting from scratch, which is where the system compounds. After a few cycles, your agents will have diverged from every other customer's, shaped by your decisions and your governance.

2. Your data is stored on multiple levels. When you buy a security product from Microsoft, you're handling the supplier, the E5 licence, the contracting entity, the reseller, the teams using it, and the agreements behind each with their own approval history. Risk, spend and renewal decisions each happen at a different level. Many legacy P2P tools structure the relationship around the supplier record; Omnea holds the record at a more granular level, so your agents can find what you actually bought. Same supplier, different risk profile depending on the job.

3. Omnea runs one process, not six. A single purchase can move through six to eight systems and five functions. Omnea runs the process across all of them: ERP and finance, CLM, GRC and TPRM, with security, legal and finance queues routed in parallel rather than in sequence. Every approval and escalation therefore lands in the same record, which is the rich understanding that other procurement AI vendors struggle to recreate.

We run the whole process, so we hold the record. That's why our agents know things other vendors' agents can't.

05 — Results

Compounding results, in practice

86→95+%: approval Agent match rate with human reviewers, first few thousand live decisions

70% risk assessment work now captured automatically at Reach plc

27 → 11 min average intake completion time across our customer base

23 → 4 days request cycle time at The Adecco Group, down from 10-20 clarification emails

Risk scoring reflects your risk appetite. An agent scores a vendor's inherent risk, a user overrides it, and that override becomes precedent the next score draws on. Over time the scores line up with your organisation's own patterns, such as EU data-transfer flags going with higher risk scores.

Agents capture and reflect your decisions. Repeated corrections on TPRM questionnaires are codified into the system, and continuous monitoring suggests actions that are most likely to be accepted, based on past performance.

The forms improve alongside the agents. Form Insights flags questions requesters repeatedly get wrong, so the intake process gets better with use.

And the cycle time falls. Proofpoint cut software request turnaround by 38%. TeamViewer's employees came to procurement earlier in the buying process, which gave the team room to negotiate. They've put four percentage points of EBIT, year on year, down to that one change.

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

Remi Thomas, CFO, Proofpoint.

06 — Getting started

You don't need clean data to start

If context is what makes agents useful, do you need to clean up your enterprise data before rolling out Omnea? No, you don't.

One customer, a recently merged diagnostics business, arrived with four separate ERPs, intake scattered across SharePoint forms, risk in one system and contracts in another. There was no realistic path to cleaning any of it first, and that turned out not to be a blocker.

Messy data is the starting condition for enterprise procurement, and making sense of it is what the Context Engine is built to do. You don't get organised and then adopt. You adopt, and Omnea gets you organised.

"Omnea shone an uncomfortable light on how bad, incomplete and patchy my data was, and then very quickly gave me the tools to fix it."

Graham Bettes, Procurement Director, Ocado.

07 — Ask us this

Four questions to ask any procurement AI vendor

And that includes us.

1. How is cycle ten different from cycle one? 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 you trace this recommendation to source? Which data, which policy, which prior decisions. If the answer is a model citation rather than my organisation's own records, the agent is just guessing well.

3. What does the agent deliberately not see? Anyone who answers "we give it everything" hasn't dealt with the problem that more context makes a model less accurate, not more.

4. How many levels of my data does the system hold? If the answer is one, ask how it tells my three Microsoft agreements apart.

08 — Start compounding

Start compounding

The Context Engine isn't a premium AI feature you switch on. It's the accumulated record in Omnea of how your organisation actually makes procurement decisions, growing as a by-product of work already flowing through the system.

Others remember without understanding, or route without remembering. Omnea holds the record and runs the process, so the outcomes compound.

That's why the timing question answers itself. There's no readiness milestone to wait for. Omnea's agents work from the Context Engine to perform better every cycle, starting from your first request.

Want to see the Context Engine in action? Bring your messiest data. It's exactly the starting condition the Context Engine is built for. Book a demo →