Enterprises are pouring billions into AI. Copilots, agents, foundation models, prompt engineering teams, innovation labs. The tools are more capable than ever. And yet, according to a recent survey, 56% of CEOs1 say they've gotten nothing back from their AI investments.
The uncomfortable truth underlying most enterprise AI programs is this: The processes where AI could deliver the most value are the same processes that are hardest to change.
The reason they're hard to update is usually because they’re still running on outdated legacy BPM technologies: systems that were never designed for AI, that have accumulated years of technical debt, and that quietly block transformation at every turn.
Until enterprises reckon with that reality, the AI spend will keep climbing and the returns will keep disappointing.
The ROI is real. The risk is also real.
Ask any enterprise transformation leader where AI could move the needle and the answers are consistent: loan origination, customer onboarding, fraud detection, dispute resolution. In other words, high-volume operations where small efficiency gains can translate into massive outcomes.
But those are also the processes wrapped in years of process logic, built by developers who have since left, running on infrastructure that predates modern APIs, and integrated with data systems that weren't designed to be read by AI agents.
Because of this, the areas of your business with the highest ROI for AI are also the areas with the highest risk to change. Compliance isn't optional. For these processes, governance is critical, auditability is non-negotiable, and customer outcomes matter. So when an enterprise tries to layer an AI agent on top of a legacy process, one of two things happens: either the AI gets constrained into uselessness, or it operates outside the guardrails and introduces risk the organization can't accept. Neither outcome is delivering on ROI.
Why POCs don't scale
The most common response to this tension is the proof of concept (POC). Build something adjacent. Pick a low-risk process. Show the business what's possible. Get executive buy-in. Then scale. It's a reasonable strategy that almost never works – and the reason why reveals the real problem.
McKinsey research on AI high performers draws a clear line between organizations that use AI to augment individual tasks and those that use AI to redesign workflows end-to-end. The former generates incremental gains. The latter generates transformation. The difference isn't model quality or prompt sophistication. It's whether intelligence is woven into the process itself or floating above it.
Consider fraud detection. Deploying a model that flags suspicious transactions for analyst review is useful. Genuinely useful. But it doesn't change investigation cycle times, false positive rates, or regulatory reporting posture – because it sits outside the workflow. Flagged transactions still get routed manually. Case documentation still gets assembled by hand. Escalation paths still depend on whoever happens to be watching the queue. The process hasn't changed; it just acquired a smarter inbox.
Most POCs prove the first thing. Leadership greenlights them expecting the second. When the gap becomes clear at scale, the program stalls.
Legacy process platforms are why that gap exists. They weren't built for dynamic orchestration. They weren't built for real-time handoffs between automated steps and human workers. They weren't built to expose governed workflows as composable services. Extending them into a modern agentic architecture means building against the grain of the platform, and that friction compounds with every new capability you try to add. Point solutions that sit above the process layer can't fully deliver because they're not part of the process. They're commentary on it.
Now consider the same operation rebuilt so that detection, triage, investigation routing, evidence gathering, escalation logic, and SAR filing all operate as governed steps in a single orchestrated workflow – with human review embedded at the right points rather than bolted on at the end. That's a different category of outcome entirely. And it requires the process infrastructure underneath to actually support it.
What legacy process infrastructure is actually costing you
Legacy process debt tends to be treated as a maintenance problem – an annoying line item in the IT budget, something to deal with eventually. The actual cost is much larger.
The average enterprise wastes over $370 million annually on technical debt. A significant share of that is concentrated in legacy process infrastructure. And that's before accounting for the opportunity cost: 68% of enterprises report that legacy systems are preventing them from operating as effectively as possible. Another 68% say time spent on legacy maintenance is time that could be spent making the business more effective2. When you're running AI pilots next to a process layer that consumes that much organizational energy, you're not building toward transformation. You're painting over it, and the walls are still rotting underneath.
There's also a customer cost. 57% of enterprises acknowledge their reliance on legacy systems causes customers to leave due to slow or fragmented experiences3. AI tools applied to broken workflows don't fix broken workflows, they generate fast, confident outputs from a system that still can't close the loop for the customer.
The architecture has to change
The enterprises getting real returns from AI aren't the ones with the most sophisticated models. They're the ones that made the decision to redesign their process infrastructure, not just add AI on top of it.
McKinsey data backs this up: Over 50% of AI high performers are using AI to transform their businesses, redesign workflows, and build new applications – not enhance existing ones at the margins4. The word "redesign" is doing a lot of work in that sentence. It implies that the process layer itself is being rebuilt to accurately reflect and scale with the way their business actually runs, not how it runs when patched.
This is where the nature of the problem becomes clearer. Legacy process platforms were designed for sequential, human-routed, rule-based work. Route this document to this queue. Assign this task to this role. Escalate after X days. That's a useful model for a certain era of operations. It is not a useful model for an operation that needs AI agents making real-time decisions, handing off to human workers with full context, completing multi-system tasks autonomously, and doing all of it in a way that can be audited and governed.
Those requirements demand a different architectural foundation. One where workflows can act as a transparent guardrail with both business and IT alignment. One where AI agents take their cues from the workflow, so they operate predictably. One where adding a new capability doesn't require reverse-engineering a system that's been accumulating complexity for a decade.
The migration path is shorter than it was
One reason enterprises stay on legacy process platforms longer than they should is the belief that migration is inherently expensive, slow, and risky. That belief was largely accurate five years ago, but it’s no longer a given.
The emergence of AI-assisted transformation tooling – specifically tools that ingest existing process documentation and generate structured, deployable workflow designs from them – has materially changed the economics of migration. What used to require months of requirements gathering can now be compressed into days. Companies like Vodafone are taking workflow ideas to delivered in less than 40 hours of work. The institutional knowledge embedded in a decade of process flows doesn't have to be rediscovered. It can be translated and used to accelerate your delivery.
The key mechanism in each case is the same: Start with what already exists. Process documentation and legacy system outputs aren't obstacles to migration. They're the raw material for it.
The question worth asking
If your organization is measuring AI ROI in productivity gains per seat, you're measuring the wrong thing. Copilots are useful. AI-generated summaries are useful. Faster document review is useful. But none of those move the needle on the metrics that actually matter: cost-to-serve, customer resolution rates, compliance posture, time-to-decision in high-value processes.
Those metrics are owned by the process layer. And for most large enterprises, the process layer is still legacy infrastructure.
The $40 billion isn't being wasted because AI doesn't work. It's being spent on the wrong layer of the problem. Until the process infrastructure underneath is capable of supporting trusted, governed, end-to-end agentic automation, the most sophisticated AI tools in the world are working with one hand tied behind their back.
The enterprises that address this in the next 18 months won't just get better AI ROI, they'll be operating in a fundamentally different way than the ones that didn't. In industries where operational efficiency and customer experience are the competitive edge, that gap compounds fast.
One starting point worth knowing about: Pega Blueprint lets you upload existing BPMN files, process documentation, and requirements directly into the tool. From there, you can immediately begin mapping agentic steps that replace manual handoffs in your current workflows – without starting from scratch or commissioning a months-long discovery project. If you're sitting on legacy process documentation, you're closer to a business case than you might think.
See how enterprises are using Blueprint-led discovery to build the business case for process migration – download the Legacy BPM Migration Guide or start redesigning a workflow today.