GenAI adoption is accelerating — but production deployments remain difficult.
A recent survey of 300 senior IT and data leaders shows why: the biggest obstacle isn’t the model.
It’s enterprise data architecture.
Let’s break down what the research revealed — and what it means for building GenAI systems on AWS.
The Shift: From Experiments to Production
Enterprise GenAI is moving out of experimentation.
Survey results show:
38% of companies plan production deployments in 2026
7% plan to scale existing GenAI use cases
Only 18% remain in exploration
Two years ago, only 2% had production deployments.
But as companies move toward production, the biggest issues appear.
The Top Obstacles Are Data-Related
Four of the top five deployment challenges are related to data, including:
Guardrails for responsible AI (76%)
Lack of skills (66%)
Enterprise data readiness (62%)
LLM reliability (52%)
Cost is no longer the primary concern.
The real issue is that GenAI stresses enterprise data architecture in ways BI systems never did.
The Current Enterprise GenAI Stack
The survey shows the most widely used technologies in GenAI deployments today are:
RAG — 85%
Prompt engineering — 72%
Vector databases — 64%
On AWS, a typical RAG pipeline looks like this:
This works well for unstructured data.
The survey found that 89% of organizations use unstructured data (documents, knowledge bases, support articles) for GenAI use cases.
But this only solves half the enterprise problem.
The Hard Problem: Operational Data
Enterprises still rely heavily on operational systems:
CRM
ERP
financial systems
billing platforms
The survey shows 66% of organizations use operational data in GenAI initiatives.
Unlike documents, operational data requires:
real-time access
cross-system joins
strict governance
privacy controls
This is where most GenAI architectures struggle.
What a Real Enterprise GenAI Architecture Looks Like
Production GenAI systems typically require multiple data layers, not just a vector DB.
Example AWS architecture:
This architecture combines:
vector search
real-time operational data
governed enterprise data
The Core Data Challenges
The survey identified the biggest concerns around enterprise data for GenAI:
Data quality and consistency — 59%
Fragmented data across systems — 50%
Security and privacy — 50%
Real-time access — 33%
None of these problems are solved by better prompts or bigger models.
They require better data architecture.
The Real Lesson
The biggest insight from the research is simple:
Enterprise GenAI success depends less on models and more on data architecture.
Models are improving rapidly.
But enterprise data is still:
fragmented
slow
hard to govern
Until that changes, many GenAI projects will remain stuck between:
PoC → Pilot → Production failure
📊 The full findings are available here
Top comments (0)