Agent of agents: working towards total agentic orchestration in procurement

As AI agents take on a growing share of procurement work, the next challenge is not simply to automate more tasks. It is to connect specialised agents into systems capable of pursuing and delivering broader business goals. This is the promise of Total Agentic Orchestration (TAO), in which multiple AI agents, each with specialised skills, collaborate, negotiate and execute complex workflows under a central orchestrating AI super-agent. 

This does not imply a fully autonomous AI black box, however. Pardeep Gill, senior vice president of procurement, supply chain and fleet at AEP, one of the largest power utilities in the US, explains that human owners remain accountable, while decision authority can be progressively delegated to agents within defined boundaries, with human review for high-risk and exceptional decisions: “Someone has to own the output of the agent. Someone has to own the data that was used to make that decision.”

The shift is not simply quantitative, but qualitative. Procurement is moving from a collection of digital tools that perform individual tasks towards something closer to a digital department, capable of identifying changes, allocating work, resolving exceptions and co-ordinating activity across the procurement lifecycle. But achieving that capability will depend on more than the sophistication of the individual AI agents. It requires actionable workflows, enterprise-wide co-ordination, clear accountability and, increasingly, disciplined management of an agent’s entire lifecycle.

Stop staring at the dashboard

The first step towards orchestration is deceptively simple: moving AI from observation to action. Many organisations have already invested heavily in dashboards. Yet, although a dashboard can identify a problem, ensuring that the problem is owned and addressed is another matter. An AI-driven workflow that results in “a room full of people looking at the same red or green number and nobody actually owning the fix” is clearly not doing what the business intended it to, says Colleen Zupan, senior director, business development at GEP. 

What needs to happen, believes Mr Gill, is to turn information into precise work cues, with specific people responsible for the overall process and output. In sourcing, for example, one agent might identify a change in demand, another assess supplier capacity, while a third examines contractual implications. The orchestration layer can bring those perspectives together and trigger the appropriate workflow.

Mr Gill’s test is whether AI activity can be connected to an underlying business metric and converted into an ongoing responsibility, rather than just performing an analytical exercise. Information has little value if responsibility for responding to it remains ambiguous, he says.

AEP procurement has therefore invested in data engineering and data science capability, with dedicated teams focused on data structure, lineage and governance. This would be a business requirement in any case, Mr Gill adds, but AI has been a powerful catalyst in the move towards it.

Creating a common framework

As firms deploy more agents, co-ordination becomes a pressing issue. An agent optimised for one team’s objectives may produce recommendations that conflict with those of another. Without a common framework, procurement could find itself replacing fragmented human decision-making with fragmented machine decision-making.

Mr Gill points to the risks of independent development across the enterprise. “Building independent models without a common enterprise framework leads to duplicate recommendations, conflicting supplier interactions, regulatory exposure and data security issues,” he says.

What is needed is a mechanism for reconciling objectives that sits above individual agents and platforms. As Ms Zupan explains: “We build one co-ordination layer above the individual platforms, rather than letting each function build in isolation. Any conflict therefore gets addressed and resolved before it ever reaches the supplier.”

If executed properly, such governance is not a brake on innovation, but a condition for scaling it. “Companies that skip this get stuck in the pilot stage indefinitely because no one’s willing to sign off on expanding something they can’t explain,” warns Ms Zupan.

Conducting the AI orchestra

The case for orchestration becomes particularly compelling in industries where procurement decisions are inseparable from operational resilience. The utilities sector – AEP’s home turf – provides one such example. Demand patterns are changing rapidly, as the growth of data centres creates new pressures on power suppliers, while the consequences of disruption remain severe. 

“Data centres and AI have completely changed the expectations on utilities, and have given us an opportunity to be able to do so much more for the business,” says Mr Gill. “Inventory has to be looked at through two lenses at once: risk mitigation and cash discipline. Those are both legitimate, and the job is to balance them. AI helps you take a business-wide view instead of optimising one at the expense of the other.”

The value of orchestration, Ms Zupan argues, lies precisely in its ability to translate technical activity into something the wider business can support. “It produces a business outcome instead of a task that was automated,” she says.

Planning for the full lifecycle 

Once agents are up and running, they require management throughout their operational lives. Zupan cites a five-factor framework for successful agentic implementation: named ownership for every agent; decision authority documentation approved before these agents go live; standards that capture the reasoning chain and not just the output; independent oversight outside the agent’s own chain; and a retirement process.

Ms Zupan identifies companies’ failure to plan the lifecycle of their AI agents as a significant blind spot. “An agent feels like a one-time build, when in reality it’s an ongoing cost and risk centre,” she says. “The testing, monitoring, updating and eventually retiring of the agent – that all costs money and attention.”

Companies must therefore define the objective, test for adverse outcomes, deploy in phases, monitor performance and eventually retire the agent when it no longer serves its intended purpose. 

The final, retirement stage is the one that is most neglected, says Ms Zupan. “Almost nobody remembers to formally retire one. So, an agent that’s no longer fit for purpose doesn’t announce itself. It just keeps quietly making decisions, consuming budget long after the business case has expired.” 

This loops back to human accountability and ownership of process and output, says Mr Gill. “Years from now, no company should be in a position to point at an AI agent and say ‘the agent made the decision’. Accountability has to sit with a person, and that principle has to be designed in from the start.”