What CEOs need to know about the potential of AI in Enterprises.

AI in the Enterprise: Overhyped in Execution. Undeniable in Strategic Impact

Most CEOs hear the same promise:

"AI will transform productivity, collapse costs, and redefine work."

That promise is misunderstood.

AI is not underperforming.

Enterprise readiness is.

What the Evidence Actually Says :

2025 adoption versus impact gap

McKinsey’s State of AI 2025 confirms that while adoption is now widespread, only a minority of organisations are achieving material financial impact.

The implication is clear. The limiting factor is no longer model capability, but enterprise execution, data readiness, and integration at scale.

Most enterprises see little financial impact.

That gap is not a about choosing "a" LLM.

It is an operating model problem and a IT Legacy problem.

Why Skepticism Is Rational

MIT-backed research reported in October 2024 showed that roughly 95% of enterprise GenAI initiatives have no measurable P&L impact.

McKinsey’s State of AI: Global Survey 2025 found that while AI and generative AI adoption has surged, only ~39% of respondents report enterprise-wide EBIT or profit impact, despite widespread use and localized benefits at the use-case level. This indicates a large gap between adoption and measurable enterprise value, reinforcing the narrative that most implementations are not yet delivering full financial returns.

The dominant failure mode is clear:

Enterprises deploy generic, consumer-grade AI tools and expect enterprise-grade outcomes.

That approach does not scale.

The Hard Limits of Generic LLMs in Enterprises

Models such as ChatGPT, Gemini, Claude, LLaMA are powerful.

But they are also structurally misaligned with enterprise needs at 3 levels:

1. Timeliness. These models lack real-time awareness of internal data, current regulations, and evolving business context.

   
   

2. Transparency. They do not consistently provide sources or explainable reasoning, making them unsuitable for regulated decisions.

   
   

3. Accuracy. Hallucinations remain a systemic risk. The IAPP warned in 2024 that hallucinations are a governance issue, not a temporary technical flaw. In 2025, regulators, privacy bodies, and enterprise governance frameworks continued to treat hallucinations as a structural risk tied to probabilistic generation, not as a transient model maturity issue. The core problem persists because LLMs still optimize for linguistic plausibility rather than factual certainty, which means governance, source grounding, and human oversight remain mandatory regardless of model generation.

For CEOs, this is a risk management problem.

Build a Strong Data Foundation

No model can outperform the quality of the data behind it.

Strong data governance, clear ownership, and consistent definitions across teams make AI outputs more reliable and far easier to scale.

Enterprises that invest early in data quality reduce downstream rework, compliance exposure, and AI model instability.

This is not glamorous "AI work".

But it is the most decisive work.

Remember: Garbage in. Garbage out.

Integration with Legacy Systems Is the Silent Killer

Research consistently confirms that system integrations is one of the largest barriers to AI at scale.

It is essential to modify or upgrade parts of the IT infrastructure before deploying AI broadly.

Most enterprises still depend on IT legacy systems.

Mainframes, on-prem databases, rigid CRMs, outdated APIs, and decades-old custom applications remain core to daily operations.

Why this matters.

AI does not operate in isolation.

It must read from databases, call APIs, trigger workflows, and update systems of record. When the underlying infrastructure is slow, closed, or brittle, AI workflows simply cannot function.

Organisational impact:

• Integration timelines become unpredictable

• Pilots succeed in isolation but fail in production

• Workarounds introduce new security risks

• AI initiatives remain disconnected from core business processes

Practical solution paths with Symantra:

• Use middleware and custom deep integrations layers to bridge AI tools and legacy IT systems

• Modernize high-impact legacy components selectively

• Move retrieval and reasoning workloads to cloud-based orchestration layers

• Favour modular integrations over monolithic redesigns

• Prioritize RAG pipelines to unlock legacy knowledge

Insufficient Proprietary Data and Low Data Quality

Every AI initiative depends on:

1/ high-quality,2/ compliant, and 3/ accessible data.

Many organisations lack all 3.

• Data is often inaccurate, fragmented, poorly governed, or legally ambiguous.

• Proprietary datasets suitable for training or retrieval are limited.

• Governance and privacy frameworks are incomplete or absent.

Why this matters:

AI models are only as reliable as the data feeding them.

Poor data increases hallucinations, reduces accuracy, and amplifies bias. Inconsistent or stale data leads directly to unpredictable outputs.

Impact on organisations:

• AI tools deliver contradictory answers

• Teams cannot trace data lineage or usage rights

• Slow data access creates delivery bottlenecks

• Sensitive information leaks into the wrong workflows

• Regulatory exposure increases due to misuse of personal or confidential data

Solution approaches with Symantra:

• Establish an enterprise-wide AI governance framework with clear ownership

• Standardize data definitions across business units

• Invest in automated data quality and validation tooling

• Use data augmentation, federated learning, synthetic data, and selective partnerships

• Centralize data access approvals and monitoring

• Implement retrieval-augmented generation (RAG) with Symantra (we have deep, hands-on experience designing and deploying RAG that reliably ground AI outputs in enterprise data) using high-quality, curated corpora so models answer from enterprise truth, not probabilistic guesses.

The CEO Imperative

McKinsey’s State of AI 2025 shows widespread adoption but limited scaled impact. Roughly 85% of organisations use AI, while only about 30% achieve material enterprise value.

The difference is discipline.

AI is comparable to the early Internet.

Value was not created by building websites.

It was created by redesigning operating models.

The same rule applies now with AI.

CEOs who treat AI as "a tool" will see only incremental gains.

In contrast, CEOs who treat AI as a business and data architecture transformation will reshape cost structures, decision velocity, and competitive positioning.

That is the real advantage.

Want to improve the data quality and IT Stack foundation?

Want a white label custom AI agent trained on your data?

Need the AI skills for custom AI project?

Talk to Symantra.

Did you know?

In 2025, 20% of European Union enterprises with 10 or more employees reported using at least one AI technology in their operations, up from about 13,5% in 2024, reflecting notable growth in AI adoption across the EU business landscape. The highest shares of these enterprises in 2025 were in Denmark (42.0%), Finland (37.8%) and Sweden (35.0%). The most common use of AI technology by EU enterprises was to analyse written language (11.8%).

The Eurostat December 2025 data shows that the most widely adopted AI use cases in EU enterprises focus on written language analysis and generation, directly validating the growing importance of AI in knowledge-intensive work.

This trend reinforces Symantra's thesis that AI is becoming a structural capability for knowledge management and decision-making rather than a peripheral productivity tool.

The largest productivity losses in EU enterprises stem from information fragmentation, manual document handling, and slow access to expertise.

AI investments delivering the fastest ROI are those targeting knowledge access, regulatory analysis, and internal decision support.