The GenAI strategy stack — a framework for enterprise AI adoption

Every week, I talk to companies working through their GenAI strategy. The conversations often start with a specific use case — "we need a chatbot" or "we want to add AI to our product." It's a natural starting point: begin with what users see. But over time, I've found that the most successful strategies also invest in the layers underneath.

After dozens of these conversations, I've noticed that a coherent GenAI strategy consistently breaks down into four layers. I'm calling it the GenAI Strategy Stack.

Layer 1: Surfaces — where users interact with AI

This is the most visible layer — chatbots, copilots, AI-powered search, embedded intelligence in existing products. It's where executives get excited and where demos get built. And for good reason: this is where value is delivered to actual humans.

But here's the trap. Surfaces are deceptively easy to prototype and brutally hard to productionize. A weekend hackathon can produce a compelling chatbot. Making that chatbot reliable, accurate, and trustworthy at scale? That depends entirely on the layers beneath it.

Examples: customer-facing chat assistants, developer copilots, AI-enhanced dashboards, document summarization interfaces, voice agents.

The key question for your organization: are you designing surfaces based on real user workflows, or are you building AI features because you feel like you should?

Layer 2: Tooling & integrations — how AI connects to your ecosystem

This is the layer that separates a toy from a tool. A language model on its own can answer questions. A language model connected to your CRM, your codebase, your ticketing system, and your internal knowledge base can actually do work.

This layer includes API orchestration, agent frameworks, retrieval-augmented generation (RAG) pipelines, function calling, and integration protocols like MCP (Model Context Protocol) that let AI interact with external systems in a standardized way.

Most organizations underestimate the investment here. It's not enough to have AI that can talk — you need AI that can act. And acting means deep, reliable integration with the systems your business already runs on.

Examples: AI agents that file tickets, update records, query databases, trigger workflows, pull context from multiple internal tools before responding.

The key question for your organization: what actions would AI need to take in your systems to deliver real value — and do you have the integration layer to support that?

Layer 3: Plumbing & infrastructure — the operational backbone

This is the layer no one wants to talk about in a board meeting, but it's the one that determines whether your AI initiatives survive contact with production.

Model routing and gateway management. Guardrails and safety filters. Observability and monitoring. Cost management and optimization. Latency budgets. Failover strategies. Prompt management and versioning.

This layer is where engineering discipline meets AI ambition. Without it, you get unpredictable costs, inconsistent responses, no visibility into what's working or failing, and no ability to improve systematically over time.

Companies that skip this layer end up with a collection of disconnected AI experiments that nobody trusts and nobody can maintain.

Examples: LLM gateways, prompt versioning systems, usage dashboards, content filtering pipelines, A/B testing frameworks for AI features, model evaluation and benchmarking infrastructure.

The key question for your organization: if your AI features broke at 2 AM, would you know? Would you know why? Could you fix it?

Layer 4: Data — the foundation everything else sits on

This is the layer that ultimately determines the ceiling for everything above it. You can have the best surfaces, the most sophisticated tooling, and bulletproof infrastructure — but if your data is fragmented, stale, ungoverned, or inaccessible, your AI will underperform.

Data for GenAI isn't just about volume. It's about quality, freshness, structure, and access patterns. It's about knowing what data you have, where it lives, who can access it, and how to serve it to an AI system in a way that produces accurate, grounded results.

This includes knowledge bases, vector stores, embedding pipelines, data governance frameworks, and the organizational processes that keep data clean and current.

The hard truth: most enterprises have spent decades accumulating data across dozens of systems with inconsistent formats, duplicated records, and unclear ownership. GenAI doesn't fix that problem. It exposes it.

Examples: curated knowledge bases, embedding and indexing pipelines, data quality monitoring, access control and governance policies, ground-truth datasets for evaluation.

The key question for your organization: if you asked your AI a question about your business, would you trust the answer? If not, it's a data problem.

The mistake most companies make

They build top-down.

It makes sense intuitively. Start with the thing users see. Ship a chatbot. Get feedback. Iterate.

But what happens is predictable: the chatbot works in a demo, struggles in production, hallucinates because it lacks grounded data, can't take actions because there's no integration layer, and can't be debugged because there's no observability. Six months later, the organization has AI fatigue and a shelf full of abandoned prototypes.

The companies I've seen succeed with GenAI build bottom-up — or at minimum, they invest across all four layers in parallel. They treat their data layer as a strategic priority, not an afterthought. They build infrastructure for observability and cost control early, not after the bill shocks hit. They design integrations around real workflows before they design the interface.

“Your data strategy is your AI strategy. Everything else is a user interface on top of it.”

Using the framework

The GenAI Strategy Stack isn't prescriptive about specific technologies. It's a diagnostic tool. When you're planning your GenAI roadmap, map your current investments and capabilities to each layer:

Surfaces: what AI-powered experiences are you building? Are they tied to real user needs?

Tooling: what systems does your AI need to connect to? Do those integrations exist?

Plumbing: can you operate, monitor, and improve your AI systems in production?

Data: is your data ready to power AI? Is it accessible, accurate, and governed?

If you find that all your investment is concentrated in one layer and the others are empty, that's your signal. Sustainable GenAI adoption requires strength across the full stack.