Why deterministic automation, human oversight and agentic AI governance come before autonomy in enterprise process automation.
Agentic AI dominates the technology headlines, yet the leaders who run enterprise process automation are approaching it with notable restraint. Andreas Welsch, an AI leadership expert, attended a major process automation conference and shared his observations in an interview.
His central finding: the most experienced automation leaders are not racing to replace reliable processes with autonomous agents. They are combining rule-based automation, agentic AI and human judgment deliberately, and they are building governance before they build agents.
This article distills Welsch’s perspective on how process automation leaders approach agentic AI today and what business leaders should take from it.
Executive Summary
- Leaders automate deterministically where possible, use AI where needed and keep human judgment in the loop.
- Error rates compound across process steps, which makes unchecked agentic AI a business risk.
- A shared, end-to-end view of each business process is the foundation for scaling automation.
- Governance must be agreed and in place before the first agent ships, not after the twentieth.
- Every agent needs a named owner, a tester and a maintainer, with a plan for when people leave.
Key Takeaways
- Deterministic first. Welsch observed that reliable, repeatable results still depend on deterministic automation; agentic AI belongs where rule-based automation hits roadblocks.
- Accuracy compounds. A process step that is 90% accurate, repeated three times, delivers roughly 73% accuracy end to end, far below what most organizations require.
- Agentic AI as a last resort. Both the vendor and its customers described using agentic AI only when it is absolutely necessary.
- A mature conversation. Welsch noted that the discussion focused on the best approaches for the processes being automated rather than on replacing jobs.
- Governance is about people. It depends on bringing a diverse stakeholder group to agreement on policies that protect the organization without stifling innovation.
- Ownership must be explicit. Someone must be accountable for each agent’s testing, performance and maintenance across its life-cycle.
- Proof points decide. Measurable results such as time saved, money saved and revenue generated will show whether agentic AI delivers.
What is Agentic AI?
In the way automation leaders discuss it, agentic AI is the probabilistic counterpart to deterministic, rule-based process automation. Agents can take on process steps where traditional automation hits roadblocks, and coding agents can even build the bots, agents and lightweight applications that run a process.
Because agentic AI is probabilistic, it can behave in unexpected ways or not follow the script or workflow the process owner intended. That is the core tension Welsch describes: the technology is very capable and powerful, but it introduces risk and uncertainty into processes that organizations expect to run reliably every time. Managing that trade-off carefully is now a leadership task.
Deterministic Where Possible, Probabilistic Where Needed
The framing Welsch heard repeatedly was simple: deterministic automation where possible, probabilistic automation where necessary, and human judgment always in the loop.
He explained that deterministic automation remains key to reliable and repeatable results within a process. AI and agentic automation should primarily be used where traditional automation hits roadblocks, not as a wholesale replacement for it.
Executives and customers on stage reinforced this point. They shared how they combine traditional automation with agentic AI rather than choosing one over the other.
Key Insight: Automation leaders treat agentic AI as a targeted tool for the steps that rule-based automation cannot handle. The deterministic backbone of a process stays in place, AI fills specific gaps, and people remain accountable for judgment calls. This layered design keeps the process reliable while still benefiting from AI.
Why Accuracy Compounds Across a Process
One example Welsch highlighted makes the risk concrete. If a single process step is 90% accurate, chaining three such steps does not keep accuracy at 90%. It multiplies: 90% times 90% times 90% equals roughly 73%.
That is far below what most organizations need in terms of accuracy and confidence. Each probabilistic step added to a workflow increases the chance that the end result is wrong.
This is why, in Welsch’s view, the case for deterministic automation as the foundation is so strong. Agentic AI can add capability at individual steps, but the overall process still has to meet the reliability standard the business depends on.
A Shared View of the Business Process
The strongest consensus Welsch observed was around building a common view of the same business process, something he noted is not always the case today.
He described an approach that captures how a process actually runs by combining system architecture, process design, process documentation and stakeholder interviews. An AI agent analyzes those inputs, finds gaps and overlaps, and presents them so people can decide what to act on.
That information becomes a specification a coding agent can implement as a bot, an agent or a lightweight application. The result is then orchestrated as part of a larger end-to-end business process. A running record of how the delivered process drifts from what was intended closes the loop and informs the next improvement cycle.
Key Insight: Scaling automation starts with understanding the process, not with choosing a model. When documentation, system architecture, stakeholder knowledge and performance data are brought into one view, organizations can see where reality deviates from the intended standard and decide which steps deserve deterministic automation, AI or a person.
Beyond the Agentic Hype
Welsch contrasted what he saw with the broader technology narrative. Following the news from vendors, startups and AI labs, he said, “we should all be living in an agentic reality today. However, that’s not the case, and it’s in many cases, not even practical.”
What resonated with him instead was the combination of a rigid, repeatable, predictable process and outcome, augmented strategically with AI and people.
He was also struck by what was absent. He did not witness the common narratives about replacing jobs and people that surface at many vendor conferences. “It seemed that the conversation was very mature,” Welsch said. The focus was on the best means, technologies and approaches for the processes being automated.
He attributed this maturity to experience. The decision makers present have worked on process automation for a long time, setting up centers of excellence, establishing governance and understanding what can and cannot be automated. For them, agentic AI is an evolution of how they have been running, not something fundamentally new.
Moving Agentic AI From Pilot to Production
Welsch heard several examples of agentic AI working alongside traditional automation in production. One organization in the airline industry uses this mix to schedule overnight stays and travel requests for crew members, supporting more than 1,000 bookings each month.
A financial services organization described gathering information from several teams about how a process runs and how systems connect. That accelerated the path from information capture to process design to delivery. He also cited examples from loan processing in financial services, underwriting in insurance, and telecommunications.
Still, he cautioned that going from pilot to production remains critical and difficult. Agentic AI introduces risk and uncertainty because of its probabilistic nature, so testing the agent, the automation and the edge cases matters far more.
Both the vendor and its customers described using agentic AI only as a last resort, when absolutely necessary. To run a reliable, predictable process every time, Welsch said, “you need to have the guardrails, you need to have the reliability baked in.”
Agentic AI Governance Starts With People
In one session Welsch attended, a customer described how they set up their AI governance program. What stood out to him was that “at the end of the day, everything is about people.”
Governance means bringing a diverse stakeholder group together and defining frameworks and policies that protect the organization without stifling innovation.
The most striking remark he heard: unless there is absolute agreement on how the organization governs AI and agentic AI, it should not start implementing. Governance has to be in place before the first agent ships. Adding it once 20 or 100 agents are in production amounts to incident management and triage.
Key Insight: Getting agentic AI governance right up front costs more time and alignment, but it is far cheaper than cleaning up issues after agents are already running. Leaders who agree on rules, responsibilities and controls before deployment avoid turning their AI program into a permanent incident response function.
Who Owns the Agent?
Welsch emphasized that governance for agentic AI must answer practical ownership questions. In his words, it “means who is responsible for a particular agent who has tested the agent, who is responsible if the agent is not performing as intended, who is going to maintain it.”
That information has to be recorded and kept current. He also pointed to a life-cycle question that is easy to overlook: what happens when the named employee leaves the company, and who assumes their responsibilities for the agent.
On human-in-the-loop controls, the emphasis he heard was on oversight and clarity about how decisions should be made, as a critical part of governance overall.
He added that most organizations are starting from roughly the same position on these questions. Getting it right before building and scaling an agentic AI program was a message he personally shares.
Where the Friction and the Exposure Lie
The biggest friction point Welsch heard repeatedly was the probabilistic nature of generative and agentic AI. The technology is capable and powerful, but it needs careful management because of the risk it creates when it behaves unexpectedly or does not follow the workflow the process owner intended.
On organizational and regulatory risk, he said it was rather the opposite: regulated industries such as financial services, insurance and telecommunications are already using agentic AI together with traditional process automation today.
The organizations most exposed, in his view, are professional services firms and technology organizations that rely heavily on technical expertise and software development, where agents can take over large chunks of work.
At the same time, he noted the consistent need to involve humans in discovery, design and continuous operation. How specifications become code, deployments and feedback will need further detail over time.
Leadership Implications
- Design the process before the agent. Build a shared, end-to-end view of each process, including documentation, system architecture, stakeholder input and performance indicators.
- Default to deterministic automation. Reserve agentic AI for steps where rule-based automation hits roadblocks, and calculate how accuracy compounds across the full workflow.
- Agree on governance before the first deployment. Align a diverse stakeholder group on policies that protect the organization without stifling innovation.
- Assign named ownership. Record who built, tested and maintains each agent, and plan for handover when people change roles or leave.
- Measure business impact. Track time saved, money saved and revenue generated so agentic AI investments are judged on outcomes.
Why This Conversation Matters
The decision makers Welsch observed have spent years in their roles, learning what it takes to run reliable processes, set up centers of excellence and establish governance. Their structured approach to agentic AI is a useful counterweight to a technology narrative that suggests everyone should already be living in an agentic reality.
Welsch sees the clearest signals of progress in customer stories with measurable proof points: time saved, money saved and revenue generated. He also sees an opportunity in the end-to-end view of automating entire processes, whether a step is deterministic, probabilistic or needs a human in the loop.
For leaders responsible for AI strategy, governance and workforce transformation, the lesson is consistent with Welsch’s broader work on moving agentic AI from pilots to value: technology capability alone does not create business value. Process understanding, governance and accountable people do.
Conclusion
The automation leaders Welsch observed are not skeptical of agentic AI. They are disciplined about it. They automate deterministically where they can, apply AI where they must and keep humans accountable for judgment.
In Welsch’s view, getting it right before building and scaling an agentic AI program matters most: a shared view of the process, governance controls agreed up front, and named ownership for every agent. Those still laying the groundwork can start with becoming AI-ready.
Frequently Asked Questions
How are process automation leaders using agentic AI today?
Process automation leaders use agentic AI selectively, applying it where rule-based automation hits roadblocks. They keep deterministic automation as the backbone of their processes and keep human judgment in the loop, treating agentic AI as a targeted capability rather than a replacement. Andreas Welsch observed this pattern consistently among experienced automation leaders.
What does “deterministic where possible, probabilistic where needed” mean?
It means automating process steps with predictable, rule-based automation whenever that is feasible, and using probabilistic AI only where those rules cannot handle the task. Human judgment stays in the loop throughout, so the process remains reliable, repeatable and accountable. This framing guides how leaders decide where agentic AI belongs.
Why does accuracy matter so much in multi-step agentic workflows?
Accuracy matters because errors compound across steps. Three process steps that are each 90% accurate produce roughly 73% accuracy end to end, which is far below what most organizations require. Every probabilistic step added to a workflow increases the risk that the final outcome is wrong.
When should an organization put agentic AI governance in place?
Organizations should put agentic AI governance in place before the first agent ships. Waiting until 20 or 100 agents are in production turns governance into incident management and triage. Early alignment takes more time, but it is far less costly than cleaning up problems after deployment at scale.
Who should be accountable for an AI agent?
Every AI agent needs named people accountable for it: who is responsible for the agent, who tested it, who answers when it does not perform as intended, and who maintains it. Organizations also need a plan for transferring those responsibilities when the named employee changes roles or leaves the company.
What is the biggest barrier to moving agentic AI from pilot to production?
The biggest barrier is the probabilistic nature of agentic AI. Agents can behave unexpectedly or deviate from the intended workflow, which introduces risk and uncertainty. That makes testing the agent, the automation and the edge cases far more important before production, alongside guardrails and reliability built into the process.
Which industries are already using agentic AI in production?
Andreas Welsch heard production examples from airlines, financial services, insurance and telecommunications. These include scheduling crew travel and overnight stays, loan processing and insurance underwriting. In each case, organizations combine agentic AI with traditional process automation rather than relying on autonomous agents alone.
Which organizations are most exposed to agentic AI disruption?
Professional services firms and technology organizations that rely heavily on technical expertise and software development appear most exposed, according to Andreas Welsch. Agents can take over large chunks of that work. Humans still remain essential in discovery, design and the continuous operation of automated processes.
Is agentic AI fundamentally new for process automation leaders?
Not fundamentally. For experienced automation leaders, agentic AI is an evolution of how they have long worked: setting up centers of excellence, establishing governance, deciding what can be automated and measuring business impact. Their experience explains why their approach to agentic AI is measured rather than hype-driven.