Why Project Data Quality Decides If Your AI Plans Succeed or Stall

Executives pour millions into artificial intelligence pilots for enterprise resource planning systems. They chase agents that promise to handle procurement, forecast demand and even post journal entries without human touch. Yet many of those efforts quietly fade. The reason sits in plain sight. Your AI plans prove only as strong as the project data feeding them.

Consultants at Midagon studied more than 20 live ERP transformations. Data quality, availability and the mess of unstructured content topped every list of obstacles. Less than half the programs even placed AI inside project scope. When they did, the focus stayed on vague future readiness instead of concrete production use cases. The pattern repeats across industries.

But data problems run deeper than dirty records. Decades of heavy customization in core ERP platforms create brittle foundations. Angela Maragkopoulou, chief information and digital officer at Sunlight Group Energy Storage Systems, speaks plainly. “Don’t sell standardization to your board as technical cleanup,” she told CIO. “Frame it for what it is: the price of admission to AI.” Her team pursued a clean core. The payoff arrived later in consistent data models and embedded analytics that made AI feasible.

ERP Software Blog laid out the case directly in September. Project data — the records, workflows, master files and transaction histories accumulated during implementations — forms the bedrock. Without trusted information from those projects, AI models amplify errors at scale. Inaccurate inventory figures lead to flawed forecasts. Inconsistent customer definitions produce unreliable segmentation. The blog notes that organizations seeing real returns started by getting their ERP data in order first.

Recent reporting sharpens the picture. A Bernstein analysis published this week describes ERP at a structural inflection point. AI pushes systems from mere records toward execution platforms. Autonomous agents need trusted data, clear process rules, permissions, governance and audit trails. Those elements live inside mature ERP environments. Vendors that translate intelligence into controlled workflow execution stand to gain. Manufacturing Dive reported on September 22 that data quality and sensitive decision-making still limit AI adoption in factory ERP systems despite vendor pushes like Microsoft’s Dynamics 365 agentic offerings.

SAP, Microsoft and others race to embed agentic AI. SAP’s Autonomous Enterprise vision, detailed after its Sapphire event, positions the ERP as the brain for AI-native operations. Yet experts caution that without clean foundations the agents stumble. McKinsey researchers examined the divide between AI agents and ERP. They advise leaders to map every priority workflow backward to specific ERP data elements, transactions, events and business rules. Skip that step and modernization efforts drain resources without delivering differentiation.

Customization once delivered competitive edge. Now it blocks progress. IDC research shows the share of organizations preferring customized ERP barely budged from 52 percent in late 2024 to 51 percent by August 2025. Tech debt from those choices can consume 18 to 29 percent of AI implementation costs and stretch a 30-month project to 36 months, according to an IBM study cited in CIO coverage. Bain reports that 80 percent of ERP transformations already miss goals. Add AI on top of that mess and disappointment follows.

Project data carries unique weight. It captures decisions made during rollout. It records configurations, testing outcomes, data migration rules and exception handling. When that information stays fragmented or poorly documented, AI lacks context. Retrieval-augmented generation systems falter. Agents make decisions on stale or conflicting inputs. One manufacturing leader described in Saigon Technology’s September 23 analysis how master data quality determines whether embedded AI delivers or corrupts ledgers.

So what separates winners from stalled initiatives? Several patterns emerge. Treat data remediation as a funded workstream inside the ERP program, not an afterthought. Midagon’s Jarkko Peltonen urges exactly that. Name a business owner for AI now rather than leave it to IT. Establish clear governance for tooling and scaling. Avoid fragmented adoption of Copilot, ChatGPT and vendor-specific features without alignment.

PwC’s June guidance on SAP S/4HANA stresses a clean core, lean architecture and strong master data governance as prerequisites for scalable AI value. AI agents can boost real-time transparency, automate processes and cut implementation timelines by up to 40 percent. Those gains remain theoretical until data readiness catches up.

TechTarget explored agentic AI in ERP this summer. Experts there repeat a consistent warning. Readiness comes down to data. Most enterprises lack the universal foundation required. A data lakehouse approach helps, yet few have built it. Business context matters as much as raw accuracy. Agents need to understand not just numbers but the rules, permissions and exceptions that govern real operations.

Recent web searches reveal the conversation accelerating. A September 24 TechTarget piece on preparing enterprise data for advanced AI notes that only 7 percent of companies surveyed by Accenture achieved necessary readiness for scaling generative, agentic and physical AI. Over 80 percent paused or adjusted initiatives due to data risks. Cloudera found 73 percent struggle with processing and preparing data. Governance, traceability and protection become non-negotiable.

Global Banking & Finance reported September 23 that the next constraint on enterprise AI centers on reliable, current and governed information. Customer records conflict across systems. Product data stays incomplete. Permissions vary. When AI surfaces these weaknesses, back-office issues turn into visible operational risks. NIST’s AI risk framework underscores data quality, validity and reliability across the full lifecycle.

Organizations that act early follow a practical sequence. They audit existing project data for accuracy, completeness and consistency. They standardize definitions across finance, operations and supply chain. They build canonical schemas and change-data-capture pipelines instead of nightly batches. They define what clean means for each field through data contracts. Only then do they layer AI features.

Digital Fractal outlined a 90-day readiness audit for CIOs in mid-September. The first month produces a data quality report, architecture recommendations and ranked use cases. Later phases focus on pilots and governance. Gartner predicts 40 percent of enterprise applications will include task-specific AI agents by year-end. Many will fail without outside help on agentic architectures.

Implementation projects themselves stand to benefit. BCG estimates generative AI can cut ERP transformation effort by 20 to 40 percent in areas like requirements gathering, data mapping and testing. One European energy company used the technology to accelerate data migration cycles and reach production-quality information faster. Yet those productivity lifts assume the underlying project data was captured and governed properly in the first place.

Executives who treat AI strategy as separate from data strategy set themselves up for frustration. The two remain inseparable. Project data — the output of past implementations, upgrades and migrations — provides the historical record, the business logic and the exceptions that agents must respect. Ignore it and models produce confident but wrong answers. Address it head-on and AI moves from experiment to execution layer.

Manufacturing giants offer glimpses of what works. Nestlé shifted its ERP to cloud in 2022 as part of a broader digitization push. Schreiber Foods partnered with Ascendion to deploy AI agents across dozens of sites. Success in both cases rested on prior attention to data foundations. Microsoft’s specialized Dynamics 365 program connects demand, supply, production and cost data so agents can act rather than merely recommend. Still, experts quoted in Manufacturing Dive stress careful execution. Sensitive decisions cannot rest on systems built only to record transactions.

The message lands with force. Companies that standardize first gain optionality. They reduce integration debt. They create semantic models that Copilot and similar tools can trust. They position ERP as the control layer for autonomous processes instead of a legacy constraint. Those that delay data work watch their AI investments compound existing problems.

Ownership questions matter too. Most programs lack a single accountable leader for AI. Responsibility drifts to IT and stays disconnected from business outcomes. Midagon recommends naming that owner early and establishing light governance. Fragmented tooling choices only multiply risk.

Look across the research. From Midagon’s survey findings to Bernstein’s market analysis, from CIO interviews to ERP Software Blog’s direct warnings, one truth repeats. AI amplifies whatever sits underneath. Clean, governed, contextual project data unlocks speed and insight. Messy, siloed or undocumented information creates costly failures at scale.

Leaders preparing 2027 budgets would do well to allocate for data remediation inside core ERP initiatives. The expense feels painful in the moment. Yet it buys admission to the next wave of productivity. Skip it and the agents will simply automate yesterday’s mistakes faster. The choice belongs to the executive team. Data readiness is no longer optional. It defines what comes next.