Is there any good news for enterprises disappointed in their AI automation efforts? There is — but only if organizations confront a harder truth first. For companies struggling to realize value from AI agents, the problem is rarely the technology itself. It is the absence of process intelligence — the contextual foundation agents need to operate safely, efficiently and at scale.

At the start of 2025, hype over AI agents was everywhere. Today the hype remains — but reality is edging into the picture. Enterprises are heading into 2026 carrying the same quiet complaint: They bought AI agents, wired them into workflows and waited for magic. What they got instead were automated actions colliding with processes they didn’t understand. This is no doubt partly why Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, citing high costs, unclear business value and weak risk controls in its June 2025 forecast. 

AI agents did exactly what they were told, but they acted on fragmented system data rather than end-to-end workflows, and that mismatch exposed a problem many organizations prefer not to admit: They do not understand how work actually moves through their own systems.

That tension — between ambitious AI adoption and limited process understanding — emerged as one of Celosphere 2025’s clearest and most consistent messages. It cut through the AI-agent chatter with a blunt takeaway: Automation can’t fix what organizations don’t understand. Data processing company Celonis SE used the event to reinforce that point with evidence, customer examples and a steady drumbeat of process-first thinking that reframed how agentic AI should be deployed at scale.

“AI is everywhere, but it’s not working everywhere at the moment … and I think that’s really the exciting part of Celonis and process intelligence — [to] really make that work … at companies at scale,” said Manuel Haug (pictured), field chief technology officer of Celonis, in an interview with theCUBE’s Rob Strechay and Savannah Peterson, during Celosphere 2025.

Discussions at the conference covered the difficulty of wiring AI agents into enterprise systems and why process visibility matters at scale.

This feature is part of SiliconANGLE Media’s ongoing coverage of the rise of process intelligence, AI-driven execution and real-time enterprise operations. (* Disclosure below.)

AI agents don’t fail quietly

Most enterprises believe they know their processes. They usually don’t. Approval chains fork in email, exceptions pile up in spreadsheets, ownership blurs across teams.

Paul Nashawaty, principal analyst at theCUBE Research, points to numbers that back that up. In theCUBE’s latest application-development and process-optimization research, 72% of organizations said their biggest bottleneck was not tooling, but invisible inefficiencies buried in operational processes, a finding he has discussed in recent Celosphere-related analysis.

 AI agents walk straight into that fog. They surface late invoices, stalled pull requests or missed handoffs, but they cannot tell whether the issue is a real failure or a normal exception unless the process is visible end to end. The result is friction, not flow, because humans get flooded with alerts that feel smart but land wrong, according to Nashawaty.

That failure pattern has become one of the defining risks in agentic AI deployments — even when model quality and automation tooling are otherwise solid. It also explains why attention has shifted away from automation mechanics and toward something more fundamental: understanding how work actually flows.

Process mining isn’t enough anymore

Process intelligence did not emerge from theory — it was born out of  observation. For years, large enterprises used process mining to reconstruct workflows from system logs, which gave teams a kind of X-ray vision into how work actually moved across ERP systems, ticketing platforms and CI/CD pipelines. The insight was often uncomfortable. It showed rework, loops, dead ends and informal workarounds that never appeared in official process diagrams. But it stopped short of action, according to theCUBE Research’s Rob Strechay.

A new phase is taking shape across the market. Organizations are pushing beyond passive observation toward systems that not only map how orders, invoices, service tickets and code artifacts move in real time, but also recommend what to fix and when to intervene. That shift — from mining to intelligence — matters because AI agents do not invent better processes — they faithfully execute what already exists. When workflows are fragmented or misaligned, agents automate the mess with remarkable efficiency.

Strechay put it plainly in recent coverage: “If you bake broken processes into agentic workflows, you won’t get the ROI you expect. Mapping and improving them first is what separates useful AI from expensive mistakes.”

That distinction matters because AI agents do not invent better processes — they faithfully execute what already exists. The consequences of that reality become especially clear in everyday operational scenarios.

Celonis repeatedly pointed to order-to-cash as a proving ground. It’s messy, it spans teams, it breaks easily. For example, an AI agent may flag a late invoice and escalate it automatically because the ERP system said payment was overdue. This agent would be correct according to the data it saw. But it would be wrong in reality. Here, the customer pays only after full delivery, the shipment is split and the invoice is triggered early — creating a false escalation.

Without process intelligence, the agent spams finance, irritates sales and confuses the customer. With process intelligence layered in, the same agent recognizes partial delivery patterns, suppresses false escalations and routes the issue to logistics instead. That difference — which has an outsized impact on trust — has nothing to do with the model. It has everything to do with context.

In other words, AI agents don’t need better guesses — they need accurate maps of how the business actually runs.

In theCUBE’s 2025 process optimization research, a majority of enterprise leaders said AI must understand how the business actually runs before it can operate safely at scale. The same research found 23% of developers already use AI agents regularly, a number that is rising as pilots move into production.

Nashawaty described process intelligence as “the contextual fuel AI agents need.” 

“Without a real-time map of work, agents automate the wrong things or amplify small errors into big failures. The winners will pair agent orchestration with process intelligence,” he said.

That framing connects directly to value creation — across the top line, bottom line and green line. When organizations lack process clarity, automation often increases rework, resource waste and operational drag. Process intelligence makes efficiency and sustainability measurable by exposing friction before it is automated at scale.

Celonis is leaning hard into that thesis, positioning its platform as the layer that turns AI agents from enthusiastic interns into reliable operators. That shift marks a broader move away from retrospective insight and toward real-time decisioning inside operational workflows.

From dashboards to decisioning

Celonis’ Chief Product Officer Daniel Brown told theCUBE the company’s goal is to move beyond dashboards and embed decision intelligence directly into operational workflows so AI agents can recommend and trigger next-best actions with guardrails in place. 

“AI agents rely on understanding processes because context turns data into actionable, meaningful and safe insights,” he said. “Agents don’t just act on isolated steps — they need to see how actions connect.”

That framing resonates with practitioners who are tired of analytics that explain yesterday but cannot change today. It also aligns with the broader market shift toward what Nashawaty calls intelligent control planes, systems that observe, guide and intervene rather than merely report.

Celonis’ Process Intelligence Graph, a digital twin of enterprise operations, sits at the center of that strategy by giving AI agents a shared source of truth across systems. The broader market is converging on the same conclusion, even if not all vendors are equipped to act on it.

The global market for process mining and process intelligence tools was estimated at $1.4 billion in 2024 and is projected to grow to roughly $21.9 billion by 2030 — and possibly higher depending on adoption rates and market definitions — according to recent industry research.

Celonis is not alone in recognizing the problem. It is, however, ahead of most competitors in addressing it. Major cloud providers, including Microsoft, IBM and AWS, increasingly partner with Celonis rather than compete head-to-head. At the same time, RPA vendors racing to rebrand as AI companies continue to encounter the same limitation: Automation without process clarity scales poorly.

“Celonis isn’t a dashboard company anymore. It’s a decisioning layer,” Strechay said. 

That positioning explains why Celonis continues to show up in large-scale transformations across manufacturing, logistics, financial services and the public sector. It also explains why Celosphere landed as a moment of clarity rather than hype for many enterprise leaders navigating AI adoption.

A sober takeaway from a noisy market

Celonis’ position is clear: AI agents are inevitable. The more consequential message from Celosphere was quieter but harder to ignore — enterprise self-awareness must come first. AI agents amplify what already exists. They reward clarity and punish confusion.

Celosphere served as a moment of strategic clarity for enterprise AI leaders. The event’s most important message was not about advancing agentic capability, but about understanding the systems those agents are expected to operate within. Process intelligence turns that understanding into something operational rather than theoretical, according to Celonis.

The takeaway was unambiguous: Fix the workflow first, then automate. Process intelligence is not a supporting feature of agentic AI — it is the prerequisite for trust, resilience, sustainability and measurable impact at scale.

(* Disclosure: TheCUBE is a paid media partner for Celosphere ’25. Neither Celonis, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE