A short break gave me the distance to see what my teams had been telling me for three years.
For three years, I gave the same advice to anyone rolling out AI in an enterprise: pick the best model for each job, mix and match, don't lock into one vendor. Best of breed, every time.
I was wrong — and half-right. It took a week away from the noise to see why.
The Signal I'd Been Ignoring
Every time we announced another tool, the response from our power users wasn't excitement. It was fatigue. "I just got the last one working — why do I need this?"
I had reasons. They had a point.
What I'd missed was that vendor sprawl is a tax on adoption. Every new tool resets the learning curve, fragments the workflow, and erodes trust in the strategy. The org doesn't have the breadth or the appetite to keep switching. Treating every task as a separate procurement decision optimized for the wrong thing.
The fix isn't to give up on choice. It's to apply different rules to different layers of the stack.
The Two Layers, and the Rule for Each
Horizontals are the AI capabilities everyone uses the same way: drafting emails, summarizing meetings, extracting action items, cleaning spreadsheets, light analysis. The task is the same in finance, legal, or operations — only the content differs.
The rule: pick one vendor. Whichever already lives in your stack. If you're a Microsoft 365 shop, that's Copilot. Google Workspace, Gemini. Salesforce-heavy, Einstein. The frontier models converge on these tasks within months of each other; your organization will not switch fast enough to capture the gap. Standardize, train once, move on.
Verticals are the AI capabilities built into how your business actually runs: claims pricing, ticket triage, anomaly detection, underwriting, fraud scoring, demand forecasting. These are tied to your data, your processes, and your competitive edge. The task is not the same across industries — it's the opposite of generic.
The rule: don't pick an AI vendor. Pick your hyperscaler.
This is the part that just changed. OpenAI's models are now available on AWS Bedrock — Azure exclusivity is over. Claude has been on Google Vertex AI for two years. The major model families are converging onto the major clouds. Inside your hyperscaler, swapping from one model to another is genuinely closer to "another endpoint" — same auth, same governance, same billing, same compliance posture. Your engineers can pick the right model per use case without your procurement team renegotiating contracts.
What This Looks Like in Practice
Take a claims process at an insurer. The bolt-on version: an existing workflow with a chatbot pinned to the front for intake, and an AI-assisted summary at the end for the adjuster. The process itself is unchanged — slow, sequential, human-gated at every step. The AI is decoration.
The AI-first version: an agent ingests the claim, pulls policy data, cross-references prior claims, validates documentation, and either auto-resolves the routine cases or hands the adjuster a pre-investigated package. The adjuster works on judgment calls, not data gathering. Cycle times drop from days to hours.
Same data. Same regulations. Same staff. Different process design.
"The bolt-on captures maybe 10–15% efficiency. The redesign captures 50%+ — and changes what the team is capable of."
Most enterprise AI investment today is going into the bolt-on version. That's the real waste, not vendor choice.
A 90-Day Roadmap
For leaders looking at their own stack:
Days 1–30: Audit and consolidate horizontals. List every AI tool currently deployed for general productivity. Map them to the platform already in your IT estate. Pick one, sunset the rest on a timeline, and tell people clearly. The clarity itself is half the win.
Days 30–60: Pick your hyperscaler and set the model-access pattern. If you're already heavily on AWS, Google Cloud, or Azure, the answer is mostly already made. Stand up Bedrock, Vertex, or Foundry as the default access layer for vertical use cases. Define the governance, security, and procurement path once, so engineering teams don't re-litigate it per project.
Days 60–90: Pick one vertical to redesign, not bolt onto. One process. One cross-functional team. One AI-first redesign with a measurable outcome. Resist the temptation to do five at half-speed. The point of the first one is to prove the redesign muscle, not to transform the company.
The Harder Part
The framework is simple. The discipline is hard — because most of the resistance you'll hit is internal. Vendor sprawl feels like optionality. Bolt-ons feel like progress. Picking one tool feels like settling.
It isn't. It's the difference between an AI strategy your organization can actually execute and a portfolio of pilots that never compound.
The major models are converging. Your strategy should consolidate where they have, and stay flexible where they haven't.
Pick the layer first. The rest follows.