New research finds 70% of companies navigate AI output errors, driving over 60% to prioritize shared knowledge foundations and cost savings over new AI revenue
AI hallucinations are prevalent, as one in four (25.6%) companies reported incorrect or fabricated AI outputs occurring very frequently while almost half (45.2%) stated AI confabulates occasionally. Only 27.9% said it happens rarely, according to a new study from MarketsandMarkets.
Surveying 380+ enterprise leaders across industries, the new Graphwise-sponsored Enterprise Semantic Backbone: A Foundation for Reliable and Scalable Agentic AI report found how AI is grounded decides how often it hallucinates. Among organizations relying on vector search alone, 53% report very frequent hallucinations and none report “rarely.” Among those using knowledge graph-based grounding, none report very frequent hallucinations and 60% report “rarely.” Yet only about 19% use a knowledge graph (hybrid 32%, vector search 31%). To download the report, click here.
Despite this, AI is increasingly seen as a core strategic capability that underpins business transformation as nearly half (49%) expect AI to fundamentally transform their business or significantly change how they deliver value within 2–3 years. Ironically, only 28% of organizations have deployed an enterprise-wide AI strategy.
As AI makes a decisive shift from experimentation to production, three areas organizations expect to see the strongest return on their AI investment in the coming years, include faster decision-making (62%), operational cost savings (61.5%), and a shared knowledge foundation (60.2%). Only 17.4% expect the return to come from new AI-generated revenue.
This reflects AI’s evolution into a strategic capability rather than a functional tool, reshaping enterprise architecture, operating models, and competitive positioning. In terms of AI-adoption, a little more than a third (36%) have scaled multiple use cases, 30% have run several pilots, and 6% are still exploring.
AI Lacks Trust and Why AI Outputs Need Validation
Trust remains a key concern as not one of 384 respondents rated deterministic, traceable AI behavior as unimportant. The trust levers enterprises want most include data governance 62.8%, citations and references 61.5%, human review 59.9%, rule validation 58.6%. Ironically, only about 1 in 3 validate AI outputs against business rules. Validation before outputs reach users is fully implemented happens in fewer one of four (35.7%) of organizations, partially at 29.4% and is still in development at 34.9%.
The data shows that a fundamental architectural gap exists, as current approaches improve access to data but do not provide a unified, governed, reasoning-capable foundation. More than two out of five (42%) organizations continue to rely on spreadsheets to keep the meaning of data consistent, followed by taxonomies and ontologies (41.9%), governance frameworks (39.6%), metadata catalogs (38.8%), and enterprise knowledge graph (37.2%). Spreadsheet users most often reported frequent meaning conflicts 37% vs 26–30% for ontologies and knowledge graphs.
Vendor Independence Helps Reduce AI Costs
Likewise, two-thirds have no neutral home for business, which means stating the knowledge layer of AI systems is frequently distributed across systems (35.7%) and locked inside one vendor (30.7%). Only 32.3% keep it on a vendor-neutral, standards-based layer. Almost seven in ten (67%) rated vendor independence of their AI and data architecture as critical or very important. Most cited pains included migration risk and switching costs (14.1%) and platform lock-in (13.5%).
A Semantic Backbone creates the needed foundation because it includes an Enterprise Knowledge Graph coupled with machine-readable governance that acts as the “semantic nervous system” of the organization. It is built as infrastructure rather than an overlay, serves as a tangible enterprise source of truth, and simultaneously powers applications, such as a domain knowledge hub, a metadata hub, a digital-twin model, and serves as a contextual grounding layer for agentic AI – rather than a single function.
“Enterprises have to stop treating knowledge as an application-level enhancement and begin treating it as core infrastructure. Achieving production-scale AI success requires establishing a Semantic Foundation built on governed, standards-based knowledge assets,” said Andreas Blumauer, SVP of Growth at Graphwise. “This semantic backbone then activates enterprise intelligence through agent orchestration, reasoning, and automation, providing the foundation necessary to deliver reliable, explainable, and scalable agentic AI at enterprise scale.”
The post 49% of Enterprises Expect Rapid AI Growth, but Only 28% Have a Strategy first appeared on AI-Tech Park.
