No AI without PI: What enterprises get wrong about AI return on investment

Enterprise AI has no shortage of ambition. What it lacks, in most cases, is measurable return.

That gap has been stubborn for years. In 2019, research from MIT Sloan Management Review and BCG found that seven out of ten companies reported minimal or no gains from AI. By late 2024, BCG said only 26% of companies had moved beyond proof of concept to generate tangible value at scale. And in 2025, The GenAI Divide: State of AI in Business reported that 95% of organisations were getting no measurable return from generative AI.

That is the reality despite the hype, the board-level urgency and the flood of investment. Enterprises are not short of models, copilots or agents. They are short of the operational foundations needed to make those tools useful.

That is what Celonis was getting at with its “No AI without Process Intelligence” message at Celosphere 2025. Stripped of the slogan, the argument is simple: AI cannot reliably improve a business process that the business does not properly understand in the first place.

The real reason AI ROI remains elusive

A lot of enterprise AI strategy still assumes that value comes from adding intelligence on top of existing workflows. In practice, that often means deploying an LLM assistant into a process that is already fragmented, exception-heavy and poorly mapped across departments and systems.

The result is predictable. The AI may generate plausible answers. It may even automate individual tasks. But if the wider process remains opaque, those interventions do not necessarily improve the outcome that matters most: faster order fulfilment, lower working capital, better customer retention or reduced operational waste.

This is why so many pilots look impressive in demos but disappoint in production. The issue is not always the model. More often, it is the absence of context.

Why process intelligence matters

Celonis’ core thesis is that process intelligence, or PI, should come before enterprise AI at scale. That means creating process transparency across the business: understanding how work actually flows, where bottlenecks sit, where variants emerge and where value leaks away.

In Celonis’ framing, this comes through digital twins, connected operational data and object-centric process mining. The point is not simply to visualise a workflow, but to build a living model of how the business runs in reality, not how it looks in a PowerPoint diagram.

Once that layer exists, AI becomes more grounded. A copilot can respond with operational context. An agent can act with an awareness of dependencies and constraints. Automation can target the point in the process where intervention will produce measurable business value rather than isolated activity.

Composable AI: upside and chaos

This matters even more as vendors and enterprises talk up “composable AI”.

In the best-case scenario, composability is a major advantage. If a company already has a process-intelligent digital twin, it can assemble AI solutions much faster. It can combine an LLM-based interface, a task-specific agent and workflow logic to solve a business problem such as late orders or excess inventory without rebuilding the data plumbing every time.

That is the upside Celonis is leaning into. In its recent announcement on composable AI, the company positioned process intelligence as the operational context layer that allows organisations to build modular AI solutions more safely and effectively. It also highlighted zero-copy integrations with Databricks and Microsoft Fabric, along with orchestration and MCP-based agent connectivity.

But composability also amplifies the downside when that foundation is missing.

Without process intelligence, composable AI can simply accelerate agent sprawl. Enterprises end up wiring more modules into broken workflows, disconnected systems and inconsistent governance environments. Instead of agility, they get complexity at speed.

That is why the most interesting part of the Celonis proposition is not just the AI tooling. It is the claim that PI should act as the map, memory and control plane for enterprise AI.

Where process-aware AI is already showing value

The strongest examples Celonis points to are not flashy chatbot demos. They are targeted operational use cases and include:

• Fujitsu cut excess inventory by 20% using AI-driven recommendations built on process intelligence. 
• Uniper reported double-digit millions in savings across 27 processes with an AI maintenance agent. 
• Vinmar improved productivity by 20% in a $3 billion business unit.
• Deutsche Telekom used process-aware intelligence to identify risky customer journeys earlier.

Vendor case studies should always be read carefully. Even so, the broader lesson holds: AI creates the most value when it is attached to a visible process, a measurable bottleneck and a defined commercial outcome.

That may be the real lesson for enterprises in 2026. Return on AI is not primarily about having the most advanced model. It is about knowing how the business actually works, where performance breaks down and where intelligence can intervene with precision.