RESOLVING THE LEDGER
This is one outlet's own report from TechRadar — the article as it was filed. Other outlets are covering the same event; open the full story to compare every source side by side.
See the full story · 1 sources
A reckoning is underway in enterprise AI . Chief executives who spent two years bolting frontier models onto their businesses are now asking harder questions.
What did we actually get for the money? Where does our data go when it flows through someone else's model? And when the model we standardized on last year isn’t the best one this year, how much of our business have we quietly handed over to a vendor we don't control?
That last question is the one that should keep leaders up at night, because for most companies, the honest answer is: far more than they think.
The conversation about AI has mostly been about which model is best.
That’s the wrong question.
In a field where leadership changes hands every few quarters, “best model” is a snapshot, not a strategy.
The question that actually matters is architectural: when the state of the art moves — and it always does — can you move with it, or do you have to rebuild your business every time?
Models are temporary. Plan accordingly
Here's what a decade of running AI at production scale teaches you that no benchmark chart will: models are disposable, and they're getting more disposable by the year. The transcription engine that led the market when we started is a footnote now.
The computer vision system that dominated three years ago has been lapped over and over. We've swapped best-in-class engines in and out of live customer workflows hundreds of times — across speech recognition, translation , object detection, face redaction, and now large language models — and each cycle turns over faster than the one before it.
An company that’s hard-wired to one provider inherits that provider's roadmap, pricing, and priorities as its own. When the provider raises prices, you pay. When it deprecates the version you built on, you rebuild.
When a smaller open-source model, fine-tuned on your own data, would actually do the job better and cheaper, you can't reach for it because your workflows only speak one dialect. That isn't a partnership. It's a dependency — and depending on a moving target is about the most expensive position you can be in.
That's the reasoning behind building systems where model lock-in isn't an option from day one: an orchestration layer that connects and manages hundreds of commercial, open-source, and proprietary models across different cognitive tasks, routes each job to whatever engine is best for it, and swaps models out as the state of the art shifts — without anyone having to rebuild a workflow.
The model becomes a component, not a foundation. The real foundation is the orchestration layer and the data underneath it, and those stay owned by the enterprise, not the vendor.
The discipline that makes model plurality real: evals
Model plurality sounds good in a keynote and collapses in practice without one thing: the ability to prove, on your own data, which model is actually better for your job. “Best” isn’t a leaderboard position.
A model that tops a public benchmark can badly underperform on your accents, your camera angles, your legal thresholds, your definition of “good enough.” Public benchmarks measure general capability.
They tell you almost nothing about how a model will perform inside your specific workflow. So the real currency of the next era isn't the model — it's the evaluation.
The companies pulling ahead are the ones who can put any engine up against any other on their own content, with their own quality bar, and make swap decisions based on evidence instead of vendor marketing.
That means scoring engines against each other continuously, on real customer data, so “which model” stays a measured decision you can remake any time the field shifts — not a one-time choice you're stuck with.
That evaluation muscle is itself a strategic asset. It's what turns a pile of interchangeable models into a compounding advantage — and it's precisely the capability an enterprise forfeits the moment it standardizes on a single black box.
Ownership over your own audio and video
There's a reason this matters most in audio and video. Text is basically commoditized at this point. The proprietary, defensible, hard-to-replicate data in the enterprise is the recorded record of what a company actually said, did, made, and witnessed — decades of broadcast, footage, calls, and captured events. It's multimodal, it's rights-encumbered, and it can't be replaced, which also happens to make it exactly what this generation of AI is hungriest for.
The appetite for training and tuning data has outrun what the open web can supply. What's actually needed now is what enterprises already have sitting in their archives: vast, rights-cleared, real-world, multimodal data, plus expert human judgment about what “good” looks like.
Which makes the default posture of the last two years perverse — companies have paid premium prices to push their proprietary audio and video through third-party models, with limited visibility into what's retained, learned, or one day competed against them. If your data is the scarce input everyone's after, the last thing you want to do is hand it over as a byproduct of your software bill.
The alternative is building the enterprise business the other way around: turning an organization's raw archives into AI-ready, enriched assets it actually owns, with rights and governance metadata baked in at the point of creation rather than tacked on afterward — because when the data in question is a witness's voice, an athlete's likeness, or a rights-encumbered broadcast, provenance and consent aren't optional extras.
From there, rightsholders can put that data to work themselves — licensing it to model developers and cloud providers on their own terms, with consent, provenance, and compensation built into the deal.
The world is moving toward fine-tuning and open source
Watch where the sophisticated buyers are going and the pattern is unmistakable The old reflex — route everything to whichever frontier model is ...
AIPROPX is an independent multi-source news index — we track, compare, and connect coverage from across the web into one place you won't find anywhere else.