"AI agent" has become one of those terms that means everything and nothing. Omnea, a procurement automation platform out of London, is a useful concrete example of what an agentic workflow looks like when it's solving an actual operational bottleneck rather than being bolted on as a chat interface.
The problem wasn't "we need AI," it was a broken process
Before reaching for AI as a solution, it's worth looking at the actual bottleneck: enterprise procurement reportedly takes an average of six months and involves upward of a dozen stakeholders to move a purchase from request to approval.
That delay is what pushes employees to bypass procurement entirely, creating shadow IT and untracked vendor risk. The AI here isn't solving "how do we add a chatbot," it's solving "how do we compress a multi stakeholder approval chain without losing the compliance checks that chain exists to enforce."
That distinction matters for anyone building agentic systems. An agent that just answers questions faster isn't actually removing the bottleneck. An agent that can route a request, check it against policy, flag risk, and pull in the right approver automatically is doing something structurally different, it's replacing a workflow, not accelerating a query.
What "agentic" actually requires here
For an agent to safely automate a procurement request, it needs to reliably do several things in sequence: interpret an ambiguous natural language request, check it against evolving compliance rules (which change per regulation, per vendor category, per region), route to the correct human approver when a decision genuinely needs a person, and log every decision it made and why.
That last part, the audit trail, is arguably the most underrated piece of building agents for regulated or high stakes workflows. An agent that can't explain its own decision path is a liability in any process someone might need to audit later.
The regulatory pressure is a feature of the design constraint, not an obstacle
Procurement automation exists inside a tightening regulatory environment, obligations like DORA and the EU AI Act put real constraints on how AI assisted vendor decisions have to be documented and justified.
Rather than being a compliance headache bolted on afterward, this kind of constraint is arguably what forces agentic systems to be built well: reasoning has to be traceable, decisions have to be explainable, and "the AI decided" can't be the end of the audit trail.
The takeaway for anyone building agents
The genuinely useful agentic products aren't the ones that added a conversational layer to an existing dashboard. They're the ones built by starting from an actual broken workflow, mapping every stakeholder and constraint in that workflow, and only then figuring out where an agent can safely replace a manual step without losing the guardrails the process needed in the first place.
That's a much less flashy design process than "add an AI feature," and it's exactly the discipline that's easy to skip if you're chasing the current wave of UK startup news around agentic AI rather than the underlying operational problem.
Top comments (0)