That's not a Twitter take. That's MIT's research team, after looking at 300 deployments and interviewing 150 leaders. $30 to $40 billion has been spent globally on generative AI in enterprises, and only about 5% of pilots have turned that spend into measurable revenue acceleration.

So the question isn't whether AI works. It's why almost nobody is making it work.

The answer is boring, which is probably why it doesn't get repeated enough. According to the lead author of that MIT study, the failure has very little to do with model quality and almost everything to do with how companies pick the problem, partner up, and execute. Translated for operators: it's a strategy and execution problem, not a tech problem.

Which brings us to the actual decision you're trying to make. If you want AI working in your business this year, you have three doors. SaaS tools, an in-house team, or an AI agency. Most companies pick the wrong one for their stage, then wonder why the project stalled.

Let me walk through each door honestly, because I've been on both sides of this. I've scaled a company. I've hired engineers. I now run an AI agency. I'll tell you when an agency is the right call, and when it isn't.

Door One: SaaS Tools (You're Renting Intelligence)

The SaaS path is the easiest to walk through. Sign up, plug in a credit card, and you have an AI feature. Sales agents, customer support bots, content generators, you name it.

Here's the catch. You're not buying intelligence. You're renting it. Every workflow that runs through someone else's platform is a workflow that doesn't compound for your business. The data, the prompts, the logic, the institutional knowledge of how your company actually operates: all of that lives on their server, not yours.

That's fine for commodity use cases. It's a disaster when the AI is supposed to be the thing that differentiates you from your competition. And it gets worse over time, not better. The vendor raises prices because they can. They tweak the model and your edge cases break. They get acquired and the roadmap shifts. Meanwhile, you've trained your team to operate around the limits of someone else's product instead of building a capability you own.

The MIT data confirms this in a slightly different way. "We've seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects." That quote, from one of the executives interviewed, is a polite way of saying most AI SaaS is a thin layer over a foundation model someone else built. You're paying a markup to use a public API.

Door Two: Building In-House

The opposite extreme is to hire your own AI team. Real engineers, real ownership, real proprietary capability. On paper this looks like the strategic move.

In practice, the numbers are brutal in 2026. The average time to fill an AI engineering role is three to six months. Senior roles like RAG architects, multi-agent system engineers, and MLOps leads can take six to nine months. The top 10% of AI talent rarely shows up on a job board, and recruiting them takes signing bonuses in the $50,000 to $200,000 range.

Then there's total cost. A functional in-house AI team typically runs €350,000 to €700,000 in year one. Time to first production is six to twelve months. Compare that to an agency engagement which ships in six to twelve weeks at a fraction of the cost.

This is the part that gets overlooked. When you commit to an in-house team, you're not just hiring engineers. You're committing to running a recruiting process in the most competitive talent market on the planet, building MLOps infrastructure from scratch, surviving the inevitable departure of the senior person who set up your stack, and still needing to figure out which business problem to point them at first.

I'm not saying don't do it. If AI sits at the heart of your product and is going to be your moat for the next decade, you have to build the team. Eventually. But for most companies, the first phase of AI adoption is not the moment to also become a recruiting machine.

Door Three: An AI Agency (When It Works)

The AI agency model exists for one reason. It compresses the time between "we should do something with AI" and "we have something working that makes us money." When picked well, it does four things at once that neither SaaS nor in-house can do on day one.

A real strategy, not a list of features. A good AI consultancy starts with the business, not the tech. Where is the bottleneck that costs you the most? What's the workflow where you're hiring people you don't want to hire? Where does institutional knowledge keep walking out the door? Those are the conversations that lead to ROI. Vendors don't have them because they need you to fit their product. In-house teams often can't have them yet because they're still ramping.

A roadmap that survives contact with reality. Strategy without sequencing is a wish list. An agency that has shipped fifty AI projects knows which one to do first, which dependencies to clear, which integrations are going to take three times longer than you think, and where the demo magic falls apart in production. That pattern recognition is what you're actually paying for.

Execution power on day one. The pipelines, orchestration frameworks, evaluation harnesses, deployment patterns: those already exist. You're paying for delivery, not for someone to build the same wheels every other team has built. Agencies ship four to five times faster because the foundation is already there.

Real AI Agents, not flowcharts with a chat interface. This is where the market gets noisy. Most things sold as "AI Agents" are decision trees with a friendly LLM wrapper. Real agents reason, decide, take actions across systems, and recover from edge cases. The difference shows up the moment your volume goes from a hundred a day to a thousand. A serious AI agency builds for that volume from the start, because they've been burned by the alternative.

The Discovery Difference

There's one more thing that matters, and it's the hardest to put on a comparison chart. The right AI consultancy doesn't just take your order.

When a client comes to us and says "we want a chatbot," we want to know why. What's the actual bottleneck? Is the real problem inbound volume, or is it that responses are inconsistent? Is the chatbot the answer, or is it a routing agent that sends issues to the right human in thirty seconds instead of three hours? Half the projects we ship look nothing like what the client originally asked for. That's not a problem. That's the work.

Vendors can't do this because their product is fixed. In-house teams often can't yet, because they haven't seen enough variations of the same problem across enough industries to recognize the pattern. The reason an agency model works isn't that we're smarter. It's that we've seen the movie before. Forty times.

When an Agency Is the Wrong Call

I told you I'd be honest. There are cases where you shouldn't hire us, or anyone like us.

If AI is going to be your core product differentiation for the next decade, you eventually need the team in-house. Get the agency to bootstrap you, then bring it inside. The best model for most companies is hybrid: agency for early delivery and expertise, in-house for long-term ownership once the product is validated.

If your data cannot leave your environment for compliance or sovereignty reasons (defense, certain government work, some healthcare), a typical agency setup may not fit. There are still ways to work, but the calculus changes.

If you only need a tiny piece of automation that an off-the-shelf SaaS handles cleanly, just buy the SaaS. Don't overbuild. We've turned away projects where the honest answer was "this isn't worth a custom build yet."

If you have no idea what problem you're solving, an agency won't save you. It'll just help you fail faster and more expensively. Pick the problem first.

The Real Calculation

Strip away the noise and the build-vs-buy debate comes down to four numbers. Time to first production. Year-one total cost. Probability of actually shipping. Capability owned at the end.

SaaS wins on speed and cost up front. It loses on capability owned and usually on the long-term economics.

In-house wins on capability owned and long-term economics, eventually. It loses badly on time to production, on cost in the first year, and on the simple question of whether you'll even pull off the hiring process.

A good AI agency sits in the middle on cost and time, but wins on probability of shipping and on the speed at which you start learning what actually works in your business. That last point is what matters most for the next twelve months. The companies that pull ahead in AI right now aren't the ones with the biggest team or the fanciest model. They're the ones who shipped something that worked, learned from it, and shipped the next thing. Execution velocity is the moat.

What This Means For You

If you're a decision-maker staring at an AI budget and a pile of pitches, here's the test I'd apply. Ask the vendor where they think you should not use AI. Ask the in-house candidate how they'd prioritize their first three projects with no engineers yet hired. Ask the agency to walk you through a project they killed because it wasn't worth doing.

The answers tell you whether you're talking to a partner or a salesperson.

We started What's Next AI because the gap between what's possible with AI and what businesses actually have running is enormous, and most of the bridge is operator work, not model work. Strategy, sequencing, integration, change management, the unglamorous parts. That's where the 95% of failed pilots fall apart. It's also where the 5% that win pull away.

If you want to talk through where AI fits in your business this year, and where it doesn't, let's have that conversation. No demo theater, no buzzwords. Just an honest read on whether you should build, buy, or partner. Reach out through the site or find me on LinkedIn.