Dr. Strangelove AI: How I learned to stop worrying and love multi-model strategies

Last week, I watched a CTO proudly announce their "all-in" commitment to a single AI vendor.

My heart sank.

Not because the vendor was bad; they're excellent. But because I've seen this movie before, and it doesn't end well.

The uncomfortable truth about AI monogamy

Here's what the data tells us: 90% of AI initiatives fail to scale beyond pilots (McKinsey 2024). Three-quarters of companies struggle to achieve any tangible AI value whatsoever.

The common thread? Single-vendor dependency.

Think about it. Would you build your entire business on one supplier? One customer? One product line?

Yet that's exactly what we're doing with AI.

Why vendor lock-in is more expensive than you think

McKinsey dropped a bombshell recently: CIOs are making 500%-1,000% errors in their AI cost calculations.

That's not a typo.

The culprit? They're not accounting for how vendor dependencies compound over time. What starts as a simple subscription becomes an architectural straightjacket.

Only 4% of companies have developed truly cutting-edge AI capabilities. The rest are trapped in what a colleague recently called "pilot purgatory” - endless proofs of concept that never quite deliver the promised transformation.

The rise of AI orchestration

Something fascinating is happening in the market. The AI orchestration sector is exploding - growing at nearly 20% annually toward $42 billion by 2033.

Smart organizations are building what I call "AI switchboards."

Imagine this: Your customer service team uses GPT-4 for initial responses (fast, cheap and gets the job done). Claude Opus reviews them for empathy and tone (specialised, but more expensive). A specialized model handles technical queries (trained, context specific). Another checks compliance.

Each model doing what it does best. No vendor politics. Just outcomes.

Companies using this approach report 15% revenue increases and 22% productivity improvements. Not from picking the "best" model, but from orchestrating multiple models intelligently.

The practical reality of multi-model strategies

I recently spoke with a financial services firm that epitomizes this approach.

For risk assessment, they use one model. For customer communications, another. For fraud detection, a third.

Their secret? They built a platform that treats AI models like APIs - plug and play, easy to swap, no architectural rework required.

BCG calls this the 70-20-10 principle: 70% focus on people and processes, 20% on technology and data, only 10% on the actual algorithms. I spoke about this in my last article (Admittedly – in a little too much depth…)

The firms getting this right pursue half as many AI opportunities but achieve twice the ROI.

Building your AI mesh architecture

McKinsey coined a term I love: "agentic AI mesh."

It's a composable architecture where any agent, tool, or model can integrate without system rework. Think microservices for AI.

The governance implications are profound. Instead of trusting one vendor's black box, you're creating a system of checks and balances. Models compete for accuracy. Different perspectives emerge. Blind spots disappear.

Gartner predicts 40% of enterprise AI solutions will be multimodal by 2027. Not because it's trendy, but because it works.

The competitive advantage of being vendor-agnostic

High-performing organizations are three times more likely to use AI across diverse functions: risk, legal, compliance, strategy, supply chain.

They're not asking "Which vendor should we choose?"

They're asking "How do we orchestrate the best capabilities from multiple providers?"

It's the difference between marriage and strategic partnerships. One locks you in. The other keeps everyone honest.

Time for tough questions

Ask yourself:

• Can you switch AI models without rewriting your applications?
• Are you measuring success by business outcomes or vendor metrics?
• Do you have a plan for when your vendor quadruples their prices? (Remember VMWare…)
• What happens when a competitor's model leapfrogs your vendor's capabilities?

If you're uncomfortable with the answers, you're not alone.

But here's the good news: It's not too late to change course.

The path forward

Start small. Pick one use case. Build it vendor-agnostic from day one.

Create abstraction layers. Implement model switching capabilities. Develop internal orchestration expertise.

Most importantly, shift your mindset from "picking winners" to "enabling options."

The future doesn't belong to organizations that bet everything on one AI vendor.

It belongs to those brave enough to embrace strategic AI polyamory: multiple relationships, optimized for outcomes, not vendor revenues.

Your move.