Your Data isn't wrong, but your model might be!
A Vaimo perspective
A lot has already been said about AI project failure. RAND estimates that more than 80% of AI projects fail, twice the rate of IT projects that do not involve AI.¹ S&P Global found that 42% of companies abandoned most of their AI initiatives before they reached production, up from 17% a year earlier.² An MIT report put the share of generative AI investments producing no measurable return at 95%.³
The explanations are just as consistent, and very few of them are about the technology.
No clear objective. RAND's most common root cause is a misunderstanding of what problem the AI is meant to solve.¹ Harvard Business Review calls it the "experimentation trap": scattered pilots that never connect to real business value.⁴
An organisation not built to absorb it. Researchers at Harvard Business School argue that initiatives fail because organisations are not set up to sustain them, lacking aligned incentives and redesigned decision processes.⁵ HBR's example: General Motors used generative design to produce a seat bracket 40% lighter and 20% stronger than the original, which never reached production because the supply chain was built for stamped steel.⁶
Leadership. Companies that succeed have leaders who connect technical work to business strategy, build trust and drive adoption.⁷
Governance. Gartner expects 40% of agentic AI projects to be cancelled by the end of 2027, with governance for systems that act on their own among the main obstacles.⁸
Data. Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data.⁹ In an HBR Analytic Services study, only 7% of respondents described their data as completely ready for AI.¹⁰
Each of these deserves attention. In this paper we focus on the last, because it is the least visible and the most often misread. When a pilot disappoints, the data is usually blamed for being wrong. In our experience it rarely is.
The data gap
Most disappointing AI pilots do not fail because the model is weak. They fail because it answers using whatever information it has to, interpret, combine and figure out on its own what the answer should be.
This is rarely dirty data in the old sense. Each system is fine on its own terms. The ERP knows what it knows, the PIM is accurate about products, the CRM holds the customer record. The difficulty starts when something has to work across all of them at once.
Semantic gap can be divided in 2 concepts.
The gap between systems
Take a question someone might genuinely ask: "why is my order delayed?" Answering it might touch different aspects and business concepts: what is a delayed order? Where can I find the root cause? In which order do I need to look into this? Do I need to access systems real-time? The order system, the warehouse, the carrier's events, the stock position and the support history. Five systems, five owners, and five slightly different ideas of what a customer is.
A person handles that by opening several tabs and applying human judgement. An AI assistant does not work that way. It gathers what it can reach and provides an answer, with the same confidence whether it understands the context or not.
The semantic gap
The second gap is harder, because nothing in your architecture flags it.
Ask three teams what an active customer is and you will often get three answers. Sales may say anyone who has bought more than once. Marketing may say anyone who purchased in the last twelve months. A dashboard may apply whatever default was set the day it was installed.
None of these definitions is wrong, however nobody decided which one wins.
This worked as long as an analyst sat between the data and the decision, because they knew which number to trust. Give the same environment to an agent and it will infer what active means, then infer something slightly different next week, and different again for the next use case. The numbers disagree and nobody can say which is correct.
An agent has no concept of your business. It has no instinct for suspicion either: an experienced employee notices when a stock figure looks implausible, a model reads what it is given and acts.
Closing both gaps
The semantic layer
A semantic layer closes the semantic gap. It is the agreed definition of your business terms and measures. Concepts such as order delayed, active customers, available for sales are defined once, centrally, so every consumer of your infrastructure calls the same definition instead of inventing its own.
The architecture gives the definition somewhere permanent to live and stops it drifting apart again. It does not remove the need to agree on a definition between operations, finance, marketing and commercial teams. That is usually the harder part.
The Operational Data Store
An operational data store closes the first system gap. It is a curated copy of the business data, kept continuously up to date from the master data sources. An answer spanning five systems does not require five live calls stitched together in code. It also absorbs the load, because one user question can become dozens of lookups in seconds and source systems were sized for people clicking through screens. Another aspect: operational systems keep today's state and overwrite yesterday's, so any question beginning with "why" depends on history somebody must keep.
Layer | What it holds |
|---|---|
Raw | What each source sent, as it arrived |
Curated | Cleaned, joined entities such as orders and customers, with ownership attached |
Aggregates | Pre-computed totals and trends, so common questions return quickly |
Semantic models | The agreed meaning of terms and measures |
Two conditions to make it trustworthy:
Data flows in as it changes rather than overnight, because an agent does not qualify its answers and a nightly load means confidently reporting yesterday's position all day.
Copy keeps its permissions: ownership travels with each record, the check happens against the consumer asking (person, agent…) at the moment the request is made. Every data access is logged so you can always show who saw what, and when.
Your systems of record keep owning the data. The operational data store is where the organisation reads, not where it decides.
Where to start
We offer a short AI gap assessment. We review your current AI use cases and the data foundation beneath them, identify where meaning, connection or freshness is breaking down, and set out what can be defined once and reused by everything that follows.
References
Ryseff, J., De Bruhl, B. and Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation. rand.org/pubs/research_reports/RRA2680-1.html
S&P Global Market Intelligence (2025). Voice of the Enterprise: AI & Machine Learning. spglobal.com
MIT Media Lab / Project NANDA (2025). The GenAI Divide: State of AI in Business 2025, as reported in Furr, N. and Shipilov, A., "Beware the AI Experimentation Trap", HBR, August 2025.
Furr, N. and Shipilov, A. (2025). "Beware the AI Experimentation Trap." Harvard Business Review. hbr.org/2025/08/beware-the-ai-experimentation-trap
Israeli, A. and Ascarza, E. (2025). "Most AI Initiatives Fail. This 5-Part Framework Can Help." Harvard Business Review. hbr.org/2025/11
"Match Your AI Strategy to Your Organization's Reality." Harvard Business Review, January 2026. hbr.org/2026/01
van den Broek, R., Hellauer, S. and Wang, D. (2025). "What Companies with Successful AI Pilots Do Differently." Harvard Business Review. hbr.org/2025/09
Gartner, as cited in "Agentic AI", HBR sponsored content from Deloitte, May 2026. hbr.org/sponsored/2026/05/agentic-ai
Gartner (2025). "Lack of AI-Ready Data Puts AI Projects at Risk." gartner.com/en/newsroom
Harvard Business Review Analytic Services, sponsored by Cloudera (2026). Taming the Complexity of AI Data Readiness. hbr.org/sponsored/2026/03