MIT says 95% of AI deployments fail. After sitting down with Amira Labs’ Kyle Soos and Stefan Cardenas, I reckon I know why — and it’s not the tech that’s broken.

By Ben Anchor — Tuesday, 2 December 2025 · Listen to the podcast episode

I’ve been banging on about AI in broadcast for the better part of two years now, but I’ll be honest: most of what I see amounts to expensive theatre. Proof-of-concepts that never ship. Pilots that gather dust. Vendors flogging ‘AI-powered’ solutions that are really just GPT wrappers with a broadcast skin.

So when Kyle Soos and Stefan Cardenas from Amira Labs joined me on the podcast this week, I was curious whether they’d give me the usual spiel or something closer to the truth. Turns out, they gave me both — and the gap between the two is where all the interesting problems live.

The 95% Problem

Kyle kicked things off by citing an MIT study: 95% of AI deployments fail. That’s a staggering number, and it’s not because the models are rubbish or the executives are clueless. It’s because nobody’s thinking about the workflows.

As Kyle put it: “There’s sort of that experimentation mindset of like, let’s just understand the technology first. This is all new to everyone for the most part. So having someone to really guide organizations through that is really quite an effective means.”

That rings true. I’ve seen it firsthand at Ancast — broadcasters excited by the promise of AI, but utterly paralysed when it comes to operationalising it. They know what they want (faster turnarounds, fewer errors, lower costs), but they don’t know how to get from A to B without ripping out everything that currently works.

Stefan added another layer: “Part of the reason why 95% of all AI endeavours fail is that the data isn’t actually there or isn’t accessible.” He’s right. If your streams don’t have metadata an agent can plug into, you’re not doing AI — you’re doing wishful thinking.

Language Sense: A Case Study in Getting It Right

Amira Labs is working with a top-three US broadcaster to automate language identification across international distribution feeds. Prior to their solution, engineers would literally hold a phone up to a screen and listen to each audio track. Hundreds of tracks. Per day. Per facility.

Kyle described the old workflow: “It takes 30 seconds to a minute to check per track. And when you think about doing this at a thousand plus tracks, there’s like an entire day right there someone could spend just to check a few.”

Their Language Sense product now checks hundreds of tracks in two to three minutes, flags errors proactively, and delivers notifications so engineers only deal with exceptions. It’s local, on-prem, and purpose-built for broadcasters — not a ChatGPT API with a JSON endpoint.

This is what real AI deployment looks like: boring, specific, and operationally grounded. It’s not sexy. It won’t win awards at IBC. But it saves time, reduces errors, and actually ships.

Agents Are Coming (But Not Tomorrow)

Everyone’s calling 2025 ‘the year of agents’. Stefan was more measured: “I think it’s definitely the year where you’re going to start hearing a lot of talk about agents, but I think it’s going to be a while before you actually see agents out in the field and actually working in production and operating by themselves.”

He’s experimenting with thinking models — the kind that can reason through a problem, diagnose audio issues on Channel 45, and suggest fixes without a human having to triage 500 streams first. But he’s also realistic: you need the metadata pipeline in place before agents can do anything useful.

Kyle echoed this: “If you try to just do too much at once, you’re probably going to face a lot of complications. It’s sort of like you’re biting off more than you could chew.”

That’s the discipline broadcasters need right now. Not moonshots. Not vendor promises. Small, focused tasks that scale. Language identification. Macro-blocking detection. Colour banding flagging. Do those well, and the agentic stuff follows.

Human-in-the-Loop Isn’t Going Anywhere

Stefan made a point I’ve been trying to get across for months: “Nobody really feels safe sort of letting AI run on its own. That’s natural, because even if you were to ask chat GPT a question right now, sometimes it gets it wrong, even on basic things. So imagine trying to have an AI help you broadcast the Super Bowl or something. And the AI gets it wrong.”

Exactly. The goal isn’t to replace the operator. It’s to stop wasting their time on tasks that shouldn’t require human judgement in the first place. Sam Altman asked recently: was it meaningful work to begin with? If the answer’s no, automate it. If the answer’s yes, augment the person doing it.

Amira Labs is threading that needle. They’re not selling a future where broadcasters sack half their engineering team. They’re selling a future where that team has time to focus on the macro-level decisions, the storytelling, the creative stuff — because the AI’s already triaged the exceptions.

My Take

I came away from this conversation more convinced than ever that *broadcast AI only works when it’s built with broadcasters, not sold to them*. Kyle and Stefan get that. They’re in the weeds with engineers, understanding workflows, architecting for scale, and deploying on-prem models that respect the operational realities of live media.

The hype cycle will churn on. Vendors will keep flogging LLM wrappers. Executives will keep commissioning pilots that go nowhere. But the organisations that succeed will be the ones doing what Amira Labs is doing: starting small, solving real problems, and building the metadata foundations that make everything else possible.

95% of AI projects fail. The 5% that don’t? They’re the ones that understand the difference between a demo and a deployment.