The Integration Paradox: How Low-Code BPM Bridges Legacy ERPs and AI

Seventy per cent of ERP modernisation projects will fail to meet their objectives over the next three years. Not “underperform.” Fail. And yet the pressure to push AI into core business processes has never been higher. So what happens when you can’t afford to rip out the system — but standing still isn’t an option either?

You don’t replace it. You build above it. The organisations that have worked this out are using low-code BPM platforms as an orchestration layer between their legacy ERPs and AI, skipping the risk, the cost, and the 17-month timelines that rip-and-replace drags along with it.

The ERP trap nobody talks about

Nearly half of SAP ECC customers — roughly 17,000 organisations — will still be running their legacy ERP beyond 2027. More than 63% of European SMEs are on legacy systems too. These aren’t laggards. They’re companies that ran the numbers and decided the risk wasn’t worth it.

And the maths only really work at the top of the market. Enterprise ERP programmes still run into the millions over 17-month timelines, with 55% going over budget, 68% running long, and cost overruns averaging 189%. Overall failure rates sit between 60% and 70%. But that isn’t the Southern European reality. The typical mid-market company here didn’t spend EUR 7 million on its ERP; it spent somewhere between EUR 500,000 and EUR 1 million, six or more years ago. The system still runs — it’s just ageing, and paying full price to replace it, with those odds attached, makes little sense.

What these companies want now is cheaper and more inventive. SAP customers feel it most sharply: instead of signing up for another multi-year, multi-million programme, they want a more agile alternative at 10-20% of the SAP-equivalent budget. And they can have it. Uniksystem has already migrated SAP customers onto its platform successfully, keeping what worked and dropping the cost and rigidity that didn’t.

Phased migration does better, succeeding 73% of the time against 42% for big-bang swaps. But even phased migration assumes the ERP eventually gets replaced. The sharper question — the one most consultants would rather not raise — is whether it should be replaced at all.

897 applications. Twenty-eight per cent connected.

Here’s the figure that should be keeping CIOs up at night. The average enterprise now runs 897 applications. Only 28% are integrated. Companies deploying AI agents juggle even more — 1,103 on average, a 45% jump — because AI spins up new point solutions faster than IT can wire them together.

What you get is spaghetti integration: brittle connectors, data re-keyed by hand between systems, and shadow processes quietly living in spreadsheets because the ERP can’t handle them. IT spends 55% of its budget just keeping the lights on, and 19% building anything new. The integration backlog is eating the innovation budget alive.

Ask IT leaders why AI projects stall and 95% point to integration hurdles as the primary blocker. Not model quality. Not a shortage of data scientists. Just the plain inability to connect AI outputs to legacy process inputs. Data silos hit 90% of enterprises, and 62% admit their systems aren’t configured for AI at all. The problem was never that AI doesn’t work. It’s that AI can’t reach the processes where it would actually earn its keep.

The bridge that already exists

The low-code BPM market hit EUR 9.7 billion in 2024 and is on track for EUR 26 billion by 2033, with enterprise adoption already at 86%. That’s not a wave of companies replacing ERPs. It’s companies building an orchestration layer above them.

Gartner has a name for it: the composable enterprise — assembling business capabilities from several systems through a coordination layer, instead of forcing everything through one monolith. By 2027, the firm expects 80% of AI-generated business applications to be 80% composable.

In practice, the ERP stays put. It keeps doing what it’s good at, which is financial accounting, inventory and master data. The processes that need flexibility, AI-driven decisions and cross-system coordination move up to a BPM platform sitting above it, connected through APIs, database links or service calls.

Portuguese enterprise software firm Uniksystem built its whole platform around exactly this. Rather than going head-to-head with Primavera, SAP or Oracle, it runs as the orchestration layer above them, connecting through Java-based integration and embedding AI agents straight into production workflows. It’s a very different bet from “replace everything,” and it sidesteps the 60-70% failure rate that comes bundled with that approach.

What the low-code BPM layer delivers is everything the ERP was never built to. Process changes ship in days rather than months — no ABAP rewrites, no PL/SQL development cycles, just workflow changes deployed through visual configuration. AI plugs directly into process nodes, so document intelligence, classification agents and decision models slot in without anyone touching ERP code. A single workflow can span the ERP, the CRM, document management and external services at once. And governance comes built in, with approval chains, escalation rules and audit trails designed for NIS2, GDPR and the EU AI Act from the start rather than retrofitted after the fact.

Where this is already working — and what the numbers look like

None of this is theory waiting for a proof point. The bridge pattern is already running in production across sectors, and the results are measurable.

Look at European banking. BNP Paribas runs more than 750 AI use cases in production and is aiming for 1,000. Nordea has 12 AI-powered virtual agents live across Nordic markets, handling upwards of 220,000 conversations a month at resolution rates above 90%. HSBC has 600-plus AI use cases in flight. Not one of them replaced its core banking system — every one built an orchestration layer to connect AI to the legacy transaction engine underneath.

The Portuguese banking sector tells the same story through Uniksystem. Seven banks use its platform to process over 90,000 non-compliance cases a year. Document intelligence hits near-100% accuracy pulling data out of unstructured documents, then feeds it straight into BPM workflows that route, classify, escalate and archive, with no human touch for the routine cases. Because process changes hot-deploy via XML configuration, they go live with no downtime — a world away from the months-long cycles ERP modifications demand. The payoff: 280% ROI on automation, and new processes live in two to four weeks.

Insurance shows the pattern from another angle. One Nordic insurer automated claims document processing with AI-powered OCR and NLP and reached a 70% automation rate. The decisive factor wasn’t the extraction itself. It was the BPM workflow routing extracted data to the right claims handler, triggering approvals and archiving records in line with regulation.

Even Toyota fits the mould. Toyota Motor North America tore out more than 70 interconnected spreadsheets and replaced them with an agentic AI planning layer — built by Deloitte on AWS — sitting above its existing ERP and supply chain systems. The legacy stayed. The intelligence got layered on top.

Then there’s the wider Portuguese landscape. Portugal leads Southern Europe on ERP adoption at 52%, comfortably above the EU average of 40.5%. That installed base — Primavera, PHC, SAP, Dynamics, Sage — is legacy investment companies will hold onto for years. The opportunity was never replacement; it’s augmentation. Uniksystem’s platform already integrates with all of them, and it runs over 7,000 employees on the UnikPeople platform while processing EUR 8 million in monthly payroll — proof the orchestration-above-ERP pattern scales well past process automation into full enterprise operations.

The architecture that actually ships

Strip the integration paradox down and it resolves into three layers.

At the bottom sits the legacy ERP — SAP, Oracle, Primavera, PHC. This is the system of record for financial data, master data and regulatory reporting. Stable, proven, and best left alone.

Above it runs the low-code BPM layer: the process fabric that connects the systems, routes the work, enforces governance and gives people a modern interface to work through. This is where agility actually lives, with a new approval workflow designed, tested and deployed in days.

On top come the AI capabilities — document classification, data extraction, anomaly detection, predictive routing. These attach to BPM process nodes as services rather than standing alone as apps, and the BPM layer hands them the context, governance and audit trail that make AI safe to run in an enterprise.

Uniksystem’s compliance-first design, built for Banco de Portugal, NIS2 and EU AI Act requirements, means governance isn’t bolted on afterwards. Every AI decision, every routing action, every approval is logged, auditable and explainable. For anyone operating under European regulation, that isn’t a feature. It’s the minimum viable architecture.

Get the layering right and AI opens up without the ERP migration risk. Get it wrong — bolt AI straight onto a legacy ERP, or run it as a standalone experiment off to the side — and you join the 70% that fail. The companies pulling ahead aren’t the ones with the newest ERP. They’re the ones with the smartest orchestration layer sitting above it.

Five questions to ask before you build your integration layer

Which processes genuinely need the ERP? Financial posting, tax reporting and master data governance belong there. Almost everything else is a candidate for orchestration.

Where does data get re-entered by hand? Every point where someone copies data from one system into another is an integration gap BPM can close today.

Which compliance requirements span multiple systems? Regulatory workflows that touch the ERP, document management and external reporting are precisely what BPM orchestration was built for.

How fast do your processes actually need to change? If the honest answer is “faster than the ERP development cycle allows,” you already need an orchestration layer — you just haven’t built it yet.

Is your AI producing outputs nobody acts on? When models generate insights that sit in a dashboard instead of triggering an action, the missing piece is the process fabric connecting prediction to execution.

This was written in collaboration with Uniksystem, a Portuguese company building low-code BPM and AI-powered automation for European enterprise. More at www.uniksystem.com.