Undercover agent: Can process mining police AI automation?

Counter measures

While AI can make process mining more accessible and potentially more predictive, process mining can also provide AI with an understanding of how work actually gets done inside a particular organisation. Think of it as an extra pair of eyes, keeping watch on the comings and goings of work processes, reporting back in simpler terms, thanks to generative AI.

Rather than requiring specialist skills to interrogate event logs and build dashboards, users can increasingly ask questions in natural language. Michael Curry, president of data modernisation at Rocket Software, says AI can also take this further, enabling “predictions and simulation-based recommendations”, potentially moving process mining beyond identifying what went wrong towards anticipating problems before they occur.

Peter van der Putten, director of the AI Lab at Pegasystems and assistant professor of AI at Leiden University, sees a change in expectations. Businesses, he says, increasingly want AI that “drives action, not just insight”, bringing process mining much closer to the execution of the processes it was traditionally used to analyse.

Knowledge graphs built from process data can provide “the semantic grounding that limits agent hallucinations or agents acting outside organisational intent”

Things become more interesting when the relationship is reversed. Instead of asking what AI can do for process mining, businesses are beginning to ask what process mining can do for AI. An agent given access to company data may be able to retrieve information or execute a task, but that does not necessarily mean it understands where that task sits within a wider business process, what should happen next, or the rules and dependencies that govern it.

Process intelligence potentially provides some of that missing operational context. The key word here is “potentially”.

Schaffrik sees three emerging roles for process intelligence here. The first is helping to train agents using operational process data. Then it can help design them by providing a structured picture of processes and their rules. And finally, it can observe what agents do once they are deployed. As organisations give agents greater freedom, this last point becomes particularly important.

Van der Putten points to another possibility. He says process mining can itself be applied to the execution trails left by AI agents, identifying recurring patterns that could then be converted into “simpler, more predictable and lower-cost workflows”.

In theory, this gives an agent more than access to the underlying data. Process information can provide context around the sequence in which work happens, the relationships between different activities and the rules that have historically governed them. Schaffrik says knowledge graphs built from process data can provide “the semantic grounding that limits agent hallucinations or agents acting outside organisational intent”.

That can only be a good thing, surely. Curry certainly sees the potential.

“I think this is the promise of process mining in the world of AI,” says Curry. “Autonomous monitoring and agentic remediation represent an exciting future state for process mining. This really changes the nature of process mining, and transforms it into process governance and remediation. However, this is still aspirational for most organisations.”

He also believes process mining can provide agents with something approaching a set of rails on which to operate.

“Process mining can now be used as ground truth to instruct agents on the right steps to take under specific circumstances, making those decisions deterministic rather than probabilistic,” adds Curry.

Black box

But there is an obvious complication. AI agents remain probabilistic systems, no matter how much context is wrapped around them.

“Agentic AI will never be fully reliable as per its design. It will always remain somewhat of a black box since it is probabilistic in nature,” says Schaffrik. “Many decision-makers have this in mind when deciding about giving AI full autonomy over company operations.”

Giving an agent process context is of limited value if that context itself is incomplete or unreliable. Schaffrik identifies data readiness as the first bottleneck, with process data often fragmented across legacy systems, inconsistently logged or simply incomplete.

As agents begin to operate across multiple systems, businesses need to know not only what an agent is doing, but what it is permitted to do and who is accountable when something goes wrong

As agents begin to operate across multiple systems, businesses need to know not only what an agent is doing, but what it is permitted to do and who is accountable when something goes wrong. Schaffrik says most enterprises still lack the infrastructure needed to “register agents, enforce least-privilege access, maintain audit trails, and manage agent identity across heterogeneous environments.”

There is also a danger that businesses focus on controlling the agent without questioning the process it has been asked to follow.

“Enterprises are inserting agents into legacy, human-paced workflows and then wondering why step-change value doesn’t materialise because agents layered onto old processes inherit old constraints,” says Schaffrik.

It is a problem recognised by BearingPoint’s Strambi, too. Automating a process does not necessarily improve it. The problem becomes greater as more technology is layered on top. Strambi warns that businesses may be creating systems whose individual components can be understood, but whose overall actions become more difficult to explain.

“A process map feeding a prediction model feeding an agent that takes action is three layers of inference stacked on top of each other,” says Strambi. “You can explain each layer on its own, but it’s harder to explain the chain, and the chain is often what your customer experiences.”

Strambi says organisations have to decide where a human sits in the whole structure at the point at which you build that structure.

“If you leave it to chance, you get an architecture that works until the day it does not, and nobody can tell you why,” says Strambi.

Agent energy

There are examples of companies trying to put some of these ideas into practice. German energy company Uniper has put an AI agent built using Microsoft Copilot Studio into production in its recruitment and hiring operation, combining its HR data with operational context from Celonis.

During testing, Uniper ran the same prompts with and without that additional process context. Asked to identify bottlenecks in recruitment, the agent working from raw HR data produced generic responses. With process context, it was able to identify specific problems, including 27 candidates who had remained on shortlists for more than 14 days.

“With Celonis as our context layer, we’re building AI solutions that aren’t just tools we experiment with, but assistants we trust,” says Robert Daamen, head of Process Intelligence CoE at Uniper.

Design the human role before you deploy the agent. Every organisation has a threshold beyond which removing human judgement stops saving money and starts costing trust, and you have to find yours on purpose

Trust, however, does not mean autonomy. Strambi’s question about where the human sits in the structure is still ringing in the ears, but Daamen at least says that as far as Uniper is concerned, “humans will always be in the loop and decision-makers – agents will only support”.

Despite using AI agents within the hiring process, Uniper does not allow them to assess candidates or make decisions that affect them.

“Our HR AI agents do not analyse candidate profiles or decide actions that affect candidates,” says Daamen. “Instead, the agents focus on process guidance and process management by automating administrative tasks, such as providing hiring manager briefings and reminders via Microsoft Teams and managing job requisition overviews.”

The agents can also analyse position data, he says, including finding similar roles that are currently open or have been advertised previously.

Uniper therefore illustrates both the potential and the current limits of the model. Process intelligence can give an agent a richer understanding of what is happening within a business process, but that does not necessarily mean handing the process over to it. For now, at least, the agent is helping people make and execute decisions rather than replacing them.

Not every organisation will draw the line in the same place, of course. There are so many variables still, depending on legacy systems, data quality and the raison d’etre for an organisation to pursue automation in the first place.

For Strambi, however, responsibility does not become any less human as autonomy increases. “Accountability stays with the human who decided what the agent could do, and no vendor contract or model card changes that,” he says.

That makes deciding what an agent should not be allowed to do as important as deciding what it can do. For Strambi, organisations need to make those choices before deployment rather than discovering the limits of autonomy when something goes wrong.

“Design the human role before you deploy the agent,” he says. “Which decisions can it take alone, which need a person to confirm, which are off limits? Every organisation has a threshold beyond which removing human judgement stops saving money and starts costing trust, and you have to find yours on purpose.”

This perhaps brings process mining back to its original purpose. Its value has always come from making the workings of an organisation more visible. AI may allow it to predict, recommend and increasingly provide the context within which agents operate, but that same visibility could become even more important as those agents are given greater freedom.

In that sense, process mining may be as important for understanding what organisations should not automate as it is for enabling what they can.