ERP is still the foundation

While AI dominates the headlines, enterprise resource planning (ERP) remains the digital backbone of the modern organization. It is the system of record that AI models draw data from and the home of the processes AI agents automate.

An ERP that consolidates processes and establishes reliable, well-governed data gives every AI initiative a tailwind. However, a convoluted, poorly designed one creates headwinds for a decade or more. 

The dependence of AI ROI on ERP quality has become evident. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data, and reports that 63% of organizations either lack the data management practices AI requires or are unsure whether they have them. For Chief Data Officers (CDOs), this raises an important question: how can organizations ensure the data foundation established through ERP will support AI initiatives for years to come? 

Compounding the issue is that what it means to be AI-ready differs from what traditional ERP programs were built to deliver. Financial and operational systems were designed to be accurate and auditable on the assumption that a person sat between the data and the decision, supplying from memory whatever context couldn’t be ingrained in the system. That assumption is now expiring, and the data foundation has to carry a load it was never designed to bear.

Higher stakes, new approaches

In the era of AI, the stakes for a successful ERP implementation have never been higher. Yet these projects still fail regularly, and even the successful ones are often painful. 

Research shows that roughly 70% of newly implemented ERP systems are projected to miss their business case goals, with up to 25% failing catastrophically.

Plenty of ink has been spilled on remedies – stronger executive sponsorship, better change management, stronger data stewardship, less customization – all of which are valid. Yet one strategy that can significantly improve any ERP program hides in plain sight: spending more time on assessment and design – the phases where data architecture, ownership, and governance are determined – and less time on software selection.

Where the time actually goes

An analysis of nearly 1,400 ERP projects found that companies spend an average of 17 weeks choosing a system. At the large enterprises I’ve worked with, the timeline frequently stretches across multiple quarters. Meanwhile, assessment and design are often compressed in an effort to accelerate the path to go-live. 

So why do sophisticated companies spend so much time on selection? Partly because it is the most comfortable part. 

Vendor demonstrations, scorecards, and status slides produce a steady stream of visible outputs that signal progress while delaying the more difficult steps of unpacking years of technical debt, governance gaps, and undocumented data cleansing processes. Further, a prolonged and broadly staffed selection also diffuses accountability. Its apparent rigor serves less to sharpen the choice than to distribute responsibility should the system later disappoint.

AI has made software selection more seductive. Demonstrations now showcase copilots, agents, and natural-language analytics performing flawlessly against pristine sample data. These demos are valuable for illustrating what’s possible, but they say little about what is achievable with your organization’s own data. 

Experienced data executives know AI is only as effective as the structure, quality, and governance of the data it inherits – the work established during assessment and design. As embedded AI capabilities converge across ERP platforms, the underlying data foundation increasingly becomes the differentiator. Choosing a platform based primarily on today’s AI demonstrations is deciding a long-lived question on short-lived evidence.

Despite all the effort invested in selection, the platform itself is often less important than most organizations assume. The major systems converge on standard processes, and the failures that follow tend to trace back to the condition of the data and the governance around it rather than to the badge on the box. 

Lidl offers a cautionary tale. The grocer spent roughly €500 million over seven years implementing a market-leading ERP before abandoning the effort in 2018 and returning to its legacy system. The platform was sound. The program faltered because Lidl’s design centered on a nonstandard inventory-costing convention that should have been sunset with the legacy ERP rather than replicated in the new one.

An inventory-costing convention is a data model decision as much as an accounting one, since it dictates how cost is structured, valued, and reported through every downstream system that consumes it.

The reallocation

The time reclaimed from software selection is ultimately an investment in AI readiness. Assessment and design determine whether future AI initiatives inherit trusted enterprise data or decades of technical debt. Selection still demands diligence, but time is best spent where it improves the decision rather than where it merely feigns rigor. 

Four practices support a faster and sharper selection.

1. Validate the front-runner instead of staging a pageant.

Most organizations already have a natural leading candidate given their industry, size, and existing technology. Identify the assumptions that must hold for that candidate to be the right choice, test them directly, negotiate firmly, and proceed.

2. Go around the vendors to actual users.

Advice from vendors and integrators tends to be overly optimistic. Find comparable organizations that run each system under consideration and ask them the same candid questions:

• What does the platform handle worse than expected?
• What workarounds have they built?
• Which of the advertised AI capabilities are working well in production?
• Would they choose it again?

3. Make your differentiating requirements the primary lens.

Every major platform manages standard processes similarly well, so the evaluation should concentrate on the areas of meaningful divergence. 

Examples include complex domains like advanced manufacturing, supply chain and warehousing, multi-currency reporting, and asset-intensive or field-service operations. Weigh just as heavily how naturally each platform accommodates your company- and industry-specific data model. A purpose-built vertical solution often carries those nuances natively, whereas a general-purpose platform may need to be adapted to fit them.

4. Show, don’t tell—with your data.

Vendors invariably conduct software demonstrations with pristine sample data and idealized processes, implicitly assuming your organization starts from the same place. Instead, provide real data extracts for your most demanding processes and ask each finalist to build those workflows with your data in a sandbox your team can pressure-test. Watching a platform digest your uniquely structured data and handle your most complex exceptions reveals more than any vendor-led demo ever will. Applying the embedded AI capabilities to that same data will also give a sober indication of your own readiness.

Assessment: An accurate starting point

The time recovered from selection belongs first in assessment. It is the disciplined surfacing of every bolt-on legacy system, process workaround, and data-quality problem in the current environment. Additionally, it includes the harder questions of who owns and stewards each critical data domain and whether those responsibilities are actually understood and exercised. 

In the AI era, assessment also requires cataloging the undocumented, judgment-based “shadow processes” performed outside the system. This work defines the pre-work needed to enable AI automation.   

Target learned the cost of skipping data assessment when it expanded into Canada. Racing to open 133 stores on data it had not validated, the company found that only about 30% of its product data was accurate, compared with 98% in its US operations. The errors compounded until shelves stood empty while warehouses overflowed, a collapse that cost billions and forced the retailer out of the market within two years.

Design establishes where the organization intends to go, but assessment provides a realistic picture of where to begin. The distance between the two is what produces credible plans and defensible budgets.

Design: Where AI inherits its foundation

For many organizations, how the ERP is designed ultimately matters more than which ERP is chosen. Rigorous process and UX design are critical for standardization, scalability, and adoption. Sound design also includes:

• blueprinting the system landscape, 
• mapping data flows, 
• defining how data will be structured and governed, and 
• establishing which datasets serve as the source of truth. 

These decisions carry the longest half-life in the program because every future report, integration, and AI initiative inherits them.

The design agenda has also expanded for the AI era. In organizations staffed only by people, gaps in the data model were survivable because institutional knowledge supplied the missing context. Employees knew which customer record was authoritative, how to interpret misused fields, or which legacy data to ignore. 

AI agents don’t inherit this tacit institutional knowledge. They can only act on what’s consistently available in the system, making metadata (definitions, ownership, lineage, business rules, and entity relationships) a critical design deliverable for AI enablement. 

Future-state ERP design must increasingly encode both data and its meaning because AI systems reason most effectively over what has been made explicit.

The decade-long dividend

The pattern behind most ERP disappointments is remarkably consistent. Organizations over-invest in the choice of software and under-invest in the harder work of understanding their current data environment and designing an inspired future one. The remedy is to reverse those priorities. This entails making the selection decision quickly and on the evidence, then reinvesting the recovered time and attention into an unflinching assessment and a deliberate, well-governed design. 

The ERP being designed today is the environment in which a workforce of people and agents will operate for the next ten years. Its data model will determine how effectively the meaning required for breakthrough AI automation is encoded into the system. When done well, the result is more than a system that clears the failure statistics. It becomes the data foundation that everything else, including your AI ambitions, is built on.

Ultimately, an ERP program is not simply an implementation of financial and operational software; it is the design of the enterprise’s future data architecture. That architecture will determine how effectively the organization governs information, enables AI, and scales automation over the next decade.