"Model-Agnostic" Doesn't Mean What You Think It Means

Every company selling an AI-powered CX agent will tell you they're "model-agnostic." The pitch is reassuring: no matter what model comes next, you'll always have the best one working for you. They swap foundation models behind the scenes. They route queries to whichever LLM fits the task. You're future-proofed.

Except you're not. ‍

Being model-agnostic inside a closed box is still vendor lock-in.

The vendor can swap models all they want internally — but if your business logic, prompts, and policies are trapped inside their proprietary system, you can't swap vendors. That distinction that matters.

When you hand your business logic, brand policies, and system prompts to a managed AI vendor, that IP gets encoded into their proprietary orchestration layer — their bespoke routing architecture, their prompt chains, their constellation of internal models. You can't export it. You can't port it. And if you want to leave, you're starting from zero.

That's not agility. That's a cold start problem dressed up as a managed service.

Switching models shouldn't mean starting over

The enterprise AI landscape is repricing in real time. Inference costs are dropping. New models are shipping monthly. Open-weight alternatives are closing the gap on proprietary labs. The enterprises that win are the ones who can swap models the way you swap SaaS tools — quickly, cleanly, without rearchitecting.

But most can't. A 2026 Parallels State of Cloud Computing survey found that 94% of IT leaders are concerned about vendor lock-in, with nearly half reporting they're very concerned. And in CX specifically, CX Today reports that enterprises switching AI platforms lose months of accumulated context — prompt engineering, fine-tuning, edge cases — that can't be ported to a new provider. Every switch resets the clock.

And the risks aren't just about missing the next breakthrough. If your vendor's preferred model shifts to more expensive reasoning architectures, your token costs surge — and you have no leverage to route to a cheaper alternative. If their latency degrades, you wait for them to fix it. Your cost structure and your customer experience are both hostage to decisions you didn't make.

Prompt engineering is a depreciating asset

Here's what most managed CX vendors actually sell: prompt engineering wrapped in proprietary software. Complex orchestration layers designed to squeeze performance out of today's models through heavy-handed prompt chains and routing logic.

That value proposition erodes every time a foundation model gets smarter. As models natively improve at reasoning and instruction-following, the need for elaborate prompt scaffolding shrinks. You're paying for complexity that the models themselves are making obsolete — and you're locked into a vendor whose entire business depends on that complexity persisting.

Structural bias is a feature, not a bug

When a single vendor controls your model routing, you inherit their alliances. And in this market, those alliances run deep. One of the most prominent AI CX companies is led by a co-founder who simultaneously chairs the board of a major foundation model lab. Another has built its entire stack around a single model provider's API. These aren't quiet partnerships — they're structural dependencies baked into the product.

That creates an obvious incentive problem. If a rival lab releases a model that's faster, cheaper, or more accurate for your use case, is your vendor going to route you there overnight? Or are they going to wait for their preferred partner to catch up?

True model agnosticism requires impartial routing. That's hard to deliver when your vendor's cap table has opinions.

The alternative: own your intelligence layer

The architecture that solves this isn't complicated in principle. Your business logic, policies, and prompts should live in open frameworks you control — not inside a vendor's black box. The model should be a swappable utility, called via a standard RESTful API, not a walled garden you're renting access to.

Frameworks like ADK and open agent tooling already make this possible. Build on stateless interfaces. Retain every prompt, every policy, every piece of logic inside your enterprise. And give yourself the ability to route to the best model for the job — today, next quarter, and three years from now — without a migration project and a seven-figure bill.

The AI landscape is moving too fast to build on someone else's lock-in. Choose architecture over dependency. Choose portability over promises.

Your CX stack should work for you, not for your vendor's partnership strategy.

That's the architecture we built Scaled Cognition around.