The Bottleneck Was Never the Model
The big AI labs spent May admitting, in their own way, where the actual value in AI lives. It isn't in the models. It's in the work of fitting them inside a specific business. That work is doable at small-business scale. It just looks different.
AI has never been more capable than it is right now. Businesses have never spent more trying to use it. And by one number that's been making the rounds in the last few weeks, the results are about as bad as they've ever been. A recent report from MIT (the Massachusetts Institute of Technology, whose research tends to set the tone for how the rest of the industry talks about AI) found that about 95% of the AI projects large companies have tried inside their own operations produced no measurable impact on the bottom line.
That's not a few projects that didn't pan out. That's almost all of them.
The interesting question isn't whether the technology works. By every benchmark, it works better every month. The interesting question is why so much of what's being spent on it disappears without a trace. The answer is finally getting named, and the way the big AI labs are responding to it tells you most of what you need to know about where the actual value lives.
What the labs just did
In May, OpenAI, the company behind ChatGPT, announced a new $4 billion subsidiary they're calling the Deployment Company. Its entire job is to send engineers into customer organizations and help them actually put AI to work. Not sell more software. Send people. They bought a London firm called Tomoro the same day to get a hundred and fifty of those engineers on staff from day one.
Anthropic (the company behind Claude) and Google are hiring for the same role. The industry name for it is Forward Deployed Engineer. The plain version: a software builder who works embedded inside a customer's business, sits with their team, learns how that business actually operates, writes the code that runs inside their real systems, and stays around after launch to keep it working.
A year ago, this job barely existed outside of a couple of firms. As of this spring, it's the most fought-over role in AI.
Why this category suddenly exists
The reason this category exists is something the demo videos don't show.
A demo runs on a clean slate. The model gets a clear question, the data it needs is already prepared, the answer comes back in seconds, the screen looks impressive. Nothing about that environment looks anything like a real business.
A real business runs on the specific way you take orders. It runs on the spreadsheet your bookkeeper trusts more than the official system, the customer who gets a different price for reasons that go back to 2023, the exception you make for the supplier you've worked with for eight years, and a half-dozen other workarounds nobody ever wrote down. It runs on the customer who always calls instead of using the form. It runs on the email that always gets forwarded to the wrong inbox first.
None of that is in the model. The model doesn't know any of it. So the moment you try to use AI on real work inside a real business, you find that everything interesting about the project sits in the gap between what the model can do in the abstract and how your business actually operates.
That gap can't be bought. It can't be subscribed to. The only way to close it is for somebody who understands both the technology and your business to sit down and build the piece that fits. That's what the new job title is naming.
The thing nobody quite says
Here's the part the announcements don't spell out.
The models aren't the bottleneck anymore. They haven't been for a while. AI capability has been doubling every few months for almost two years, and the gap between what's possible and what businesses actually get out of it has stayed roughly the same size. If anything it's gotten wider, because the ceiling keeps rising and the work of closing the gap doesn't get any easier when you raise the ceiling.
That work is labor. It's not a product. You can't fix it by buying a better subscription, switching to a smarter model, or adding another tool to the mix. The only thing that closes the gap is somebody sitting down with you, understanding what you do, and building the specific piece that fits your specific business.
That's what billions of dollars are now lining up behind.
What this looks like at small-business scale
Here's where it gets interesting if you run a small business.
The big labs need armies of Forward Deployed Engineers because their customers are enormous. A Fortune 500 company has six layers of approval, an IT department, a compliance committee, a legacy system from 1998 that has to keep running, and twenty different teams that all have to agree on what the new tool should do. Closing the gap for them takes a team because the gap is huge.
A small business doesn't have any of that. You have an owner who knows the business in their head. You have a handful of tools that mostly work. You have one or two real problems where the right piece of software would meaningfully change how the week goes.
At that scale, you don't need a team. You need one person who shows up, learns how you actually do things, builds the piece that fits, and stays around to keep it working. That's the same shape of work the big labs are spending billions on. It's just sized for the actual problem.
In some ways you're better positioned for this than a large enterprise. There's no board to convince. There's no legacy system that has to keep running. A decision that takes a 200-person company six months of meetings, you can make in a conversation. The piece that takes them a team of fifty to build, you might need built once, by one person, and then maintained occasionally.
How I work, briefly
I should be direct here. This is the work I do at Wallman Solutions. One person, scoping and building and maintaining the same project, sized for small businesses.
It's not a novel concept. The big labs putting a name on the category and writing billion-dollar checks to scale it doesn't change what the work itself is. It's helpful that they're describing it out loud, because it makes it easier to explain why buying an off-the-shelf AI product and hoping it fits keeps not working out the way the demo suggested it would.
If you're reading this
If you've been watching the AI news and wondering what any of it actually means for a business your size, or you've already tried a tool that didn't quite stick, that's a conversation I'd rather have early than late.
— Nathan