There's a version of procurement that most leaders can already picture: a purchase request lands, the system already knows your approved suppliers, the contract terms, the right budget owner, and the historical pattern for this request type. A document review agent assesses the MSA, flags issues against your legal playbook, and surfaces those findings directly to the CFO approval stage. Two previously separate review steps become a connected, documented chain. Approvers no longer reconstruct context from earlier in the process. Cycle times shrink. The audit trail builds itself.

This is not a distant prospect. It is already operating in production for some teams. But the gap between that picture and where most procurement organisations sit today is not a technology gap. It is a trust gap. And trust, it turns out, is something you build one step at a time.

The trust problem is not technical

The pressure to do more with less in procurement predates the current AI wave. What AI has changed is the scale of the expectation gap. Leadership sees autonomous agents in demos. Procurement teams see colleagues who are sceptical, processes that need redesigning before they can be automated, and a risk appetite that does not extend to AI making consequential decisions without a human in the loop.

Three things come up consistently when procurement leaders talk about what trust in AI actually requires.

  1. Control: teams need to know what data the AI is accessing, what actions it is authorised to take, and where the human remains in the decision.
  2. Context: AI needs to understand your policies, your workflows, your supplier relationships. Not just respond to a prompt typed into a general-purpose tool.
  3. Reliability: procurement teams care far more about measurable process improvement than AI feature counts.

As Rob Fagg, VP of Product at Omnea, put it: "Adoption fails when AI starts making decisions before people are ready to trust it. And that trust is earned through transparency and reliability."

This is the reason so many procurement AI pilots produce impressive demos and underwhelming rollouts. The model knows nothing about your supplier relationships, your approval thresholds, your risk policies, or how your team has handled similar requests before. Without a persistent organisational context layer — your policies, supplier data, contracts, spend history, and the accumulated record of past decisions — procurement automation software risks repeating the failure mode of enterprise sourcing suites a decade ago: technically live, functionally unused.

The three phases of procurement automation that actually deliver ROI

The procurement leaders getting measurable results from automation are not the ones who started with the most ambitious deployment. They are the ones who built the conditions for trust first.

Phase one is process hygiene — the unglamorous work that makes everything else possible. Before procurement automation software can be trustworthy, intake needs to be consolidated, approval workflows need to be clean, and data needs to be structured. Organisations arrive at automation with forty-plus intake forms that nobody can navigate. After consolidation, intelligent routing replaces the sprawl. This phase builds the data foundation AI needs to make recommendations worth trusting.

Phase two is embedded AI — not AI as a feature, but AI demonstrating reliability inside existing workflows before autonomy is extended. A review agent catches incomplete submissions before they progress. An approval agent surfaces recommendations with documented reasoning. A document review agent flags contract risk against your legal playbook. Decisions remain human. But the humans making them are faster and better-informed, and the track record of accurate recommendations starts to accumulate.

Phase three is agentic procurement, where agents handle the mechanics end to end for lower-risk requests and humans review exception cases. Most organisations are not here yet. The teams that reach it are the ones that treated phases one and two as prerequisites rather than shortcuts, because the autonomy in phase three has to be earned by the reliability demonstrated in phase two.

What one enterprise team learned

Vlad Craciunescu leads procurement excellence at a publicly listed enterprise software business and has been rolling out Omnea since early 2024. His team's experience tracks this model closely.

The decision they got right from the start: redesigning their procurement process from scratch rather than automating the existing one. That meant involving every downstream team — IT security, legal, GDPR, finance — in the workflow design upfront, with each team assigned ownership of their tasks. By the time a request moved through the system, it was genuinely end-to-end compliant rather than theoretically so.

What he would change: they took a big-bang approach, bringing every team in simultaneously. His advice to anyone starting today is to begin narrow — procurement plus one or two adjacent teams — get that to 80 or 90%, roll it out, fix it, then expand.

With that foundation in place, the business now has hundreds of monthly active users, with adoption growing by more than a fifth in the past quarter. The AI features delivering the most day-to-day impact are form insights and the review agent working in combination: one surfacing where form design is generating friction, the other catching incomplete submissions before they progress. Together they have improved the quality of data entering the approval process, which makes the downstream approval agents more reliable.

The evolution of their approval agents is instructive. Rather than pushing fixed approval rules across every team, each approval team now configures their own criteria. Decisions remain human. But the cognitive load on each approver is lower, and cycle times have improved.

To learn more about Omnea's AI solutions, read our AI page.

Autonomy is a destination, not a starting point

The teams furthest along this journey think about AI autonomy in levels rather than as a binary choice. At one end, AI generates a recommendation and a human applies their own judgment. A step further, AI pre-fills the context and the human confirms with a single click. Further still, routine low-risk requests — renewals under threshold, pre-approved supplier categories, standard reorders — route and approve automatically, with humans monitoring outcomes.

The line most organisations draw in practice is at risk and spend value. And the prerequisite most cited for moving further: a long enough track record of agent recommendations that the team has calibrated its trust — knowing where the agent is reliable, and where it needs more oversight. That track record cannot be shortcut. It is built by running AI in recommendation mode first, evaluating accuracy, and expanding autonomy as confidence grows.

The consistent design principle that emerges from practitioners: procurement automation software should make the user's next step obvious and reduce the effort required to complete it. When it does, adoption follows. One customer, Factorial, now has 95% of purchases following the procurement process — not through policing, but because the compliant path became less effort than the workaround.

If you want to see what this looks like across your specific workflows, book a demo and we'll show you.

The intake and review layer delivers the fastest visible ROI. The bottleneck in most procurement workflows is not the decision — it is the information gathering before the decision. Requests bounce back because they are incomplete. Approvers sit on requests because they lack context. Procurement automation software applied at intake — catching incomplete submissions, auto-filling fields from uploaded documents, routing to the right approver with context pre-surfaced — eliminates that friction without removing human judgment from the decision itself. Proofpoint cut software request turnaround by 38%. Typeform reduced overall cycle time by 62.5%. Form Insights has decreased form completion time by 63% across Omnea's customer base.

For teams further along, the approval layer delivers the most significant cycle time improvement once data quality is there. For compliance-focused teams, automated risk-tiering — where every vendor goes through diligence scaled to their risk profile — has driven measurable results: Reach plc now has 70% of risks autocaptured by Omnea AI, and Luno cut vendor screening time by over 50%.

The honest answer depends on how much trust the AI has earned. Most organisations start with AI generating recommendations that humans review. A step further, routine low-risk requests — renewals under threshold, pre-approved categories, standard reorders — can route and approve automatically, with humans monitoring outcomes rather than pre-approving every step. Fully autonomous operation across high-stakes decisions is not the right starting point for most enterprises, and probably not the right destination for the highest-risk decisions either.

The prerequisite for moving further toward autonomy: a long enough track record of accurate agent recommendations that the team knows where the system is reliable and where it needs oversight. That track record is built by running procurement automation software in recommendation mode first, evaluating accuracy, and expanding autonomy incrementally — something teams configure at the workflow level rather than as a single organisation-wide setting.

Yes. Review agents and approval agents can both be trained with explicit rules set by the procurement team, independent of historical context. Past submission history often provides the most accurate signal for whether a new submission is complete — but fixed rules can be layered on top, and this is the intended approach for hard requirements where you do not want the agent applying judgment.
Yes. This can be built into the form conditionality as a question, and the review agent can treat the answer as a hard rule. The caveat: enforcement relies on the requester answering accurately. If your IT team wants a harder gate — one that prevents the request from progressing entirely until the condition is met — that can also be configured within the workflow conditionality.