Walk into almost any company right now and someone will tell you they're “doing agents.” It's on every roadmap and every all-hands. And the numbers back it up. In a 2025 Gartner survey, 75% of IT application leaders said they were piloting, deploying, or had already deployed some form of AI agent inside their organization.
So that part is real. Everyone is, in fact, deploying agents.
But another question you might be wondering: are those agents actually getting any work done?
Mostly, no. MIT's 2025 report on the state of AI in business looked at 300 public AI deployments, surveyed hundreds of employees, and interviewed company leaders. The finding that got everyone's attention: about 95% of generative AI initiatives showed little to no measurable return. Only around 5% were delivering real, rapid value. Everyone is busy, almost nobody is getting paid back for it.
This is the gap. There's enormous enthusiasm but almost no results on the other. The obvious next question is: why? Why are so many teams shipping agents that never actually move the needle?
Why everyone's suddenly building their own
The short answer: it has never been easier to build your own agent.
You can sign up for a platform on a Monday, drag a few boxes around, connect it to your tools, write a prompt or two, and have something that looks like a working agent by lunch. No procurement cycle. No engineering ticket. No waiting. That's genuinely impressive, and it's a big part of why adoption shot up so fast. The barrier to starting basically disappeared.
But “you can do it yourself” and “you should do it yourself” are two very different things.
Think about plumbing. You can walk into Home Depot this afternoon and buy every single thing you'd need to redo the plumbing in your house. The pipes, the fittings, the shutoff valves, the torch, the tape, all of it. Nothing's stopping you. The materials are right there on the shelf, and they're not even that expensive.
That doesn't mean you should be the one rerouting the water lines behind your walls.
Because access to the materials were never the hard part. Knowing how the whole system fits together, what holds under pressure, and what happens when something goes wrong six months from now behind the drywall? That's the hard part. Having access to everything you need is not the same as knowing what to do with it. And when you get it wrong, you don't find out right away. You find out when the ceiling's wet.
That's roughly where a lot of teams are with agents right now. They've got the materials, the platforms made sure of that. What they're missing is everything that comes after.
The missing playbook
Here's the real problem, and it's a little counterintuitive.
Most platforms that let you build your own AI agents hand you the tools but not the playbook. They give you the ingredients, the workspace, the buttons. What they don't give you is the answer to the question that actually matters: what's the right way to build this for your business?
And to be fair, that's not because they're holding out on you. It's because writing that playbook is genuinely, brutally hard. Your business doesn't work like the business down the street, so a playbook that works for you would have to be built more or less from scratch, for you. At scale, across thousands of customers, that's not a feature you can ship. So they don't.
Bear with me for another analogy, but think of it like a cooking class. You show up and the place is incredible. There's every ingredient you could ever want, all of it fresh, all of it laid out and ready. You think, this is going to be amazing, I can make anything!
Then you find out there's one recipe. It's chicken piccata.
Don't like chicken piccata? That's too bad. There's only one recipe, because writing a custom recipe for every single person in the class (and then standing over each of their shoulders showing them how to actually cook it) would take more time and effort than the class could ever afford. So everyone gets chicken piccata, whether it fits what they came to make or not.
For a cooking class, that's a mild disappointment. You eat the chicken, you go home. The stakes are basically zero.
Implementing AI is not meal prep. When the “meal” is a system that's supposed to do real work inside your company, handed-down generic recipe and all, getting it wrong costs you a lot more than a so-so dinner. It costs you months, it costs you trust internally, and it costs you the thing you were actually trying to build.
Why agents fail once they hit the real world
Let's talk about what actually happens when these agents go live, because this is where the gap between “deployed” and “working” actually shows up.
AI is not one-size-fits-all. This might sound obvious, but almost everything about the self-serve approach quietly assumes the opposite. Every business runs differently. You've got your own data, in your own shape, sitting in your own tools. You've got your own workflows, your own edge cases, and that one weird exception everyone swears is rare but somehow turns out to be 30% of the volume. A generic agent has no idea any of that exists.
So here's what tends to happen. Out of the box, you can get the easy part working. You can glue together a few automations and offload that one repetitive task you do every morning (the copy-paste-between-two-tabs kind of thing). That works, and it demos beautifully. Everyone nods. The screenshot looks great in the deck.
But that's not the real work. The real work is the stuff where you'd otherwise be asking another person to do it. The “this is actually replacing what a teammate would handle” work. That's the work that matters, and that's exactly the work where self-serve agents fall down. They make it to a polished demo and then they just never go live. They can't handle the moments where the right move depends on context the agent was never given.
For example, picture an agent that's supposed to handle incoming customer requests. In the demo, it's flawless. A customer asks a clean, well-formed question, the agent answers, everyone claps. Then it goes live and a real customer shows up annoyed, half-explaining a problem that spans three previous tickets, asking for something that technically isn't allowed but that your team always makes an exception for when the account is big enough. A human on your team handles that in about four seconds, because they know the unwritten rules. The generic agent has no idea the unwritten rules exist. So it either does the wrong thing confidently, or it punts the whole thing back to a person. Which means you didn't actually replace the work, you just added a step in front of it.
Multiply that across every workflow you wanted to hand off, and you can see how a company ends up with a pile of impressive demos and not much that's actually running in production.
The data lines up with this, too. That same MIT research found that AI tools brought in from specialized partners succeeded roughly 67% of the time, while tools companies built internally on their own succeeded at about a third of that rate. The difference wasn't enthusiasm or budget. It was whether the thing was actually built to work in that specific environment.
What actually works: agents built for how you work
This is exactly why we design and build the agents for you at Amodal.
You don't get a box of materials and a “good luck.” An actual engineer builds the thing around how your team actually works, on top of your actual tech stack, and making it stable enough to handle the nuance that breaks everything else.
The difference is the same one as the plumbing. Anyone can hand you the pipes. What you actually want is someone who understands how your house is put together, builds the system to fit it, and makes sure it holds up under pressure long after they've left—including when something weird happens, because something weird always happens.
That's the whole reason the 5% pull it off and the 95% don't. It was never about who had access to the best tools. By now, basically everyone has access to the same tools. It's about whether what got built was actually shaped to the business it was supposed to serve: its data, its workflows, its exceptions, its reality.
So if your agent demos great but somehow never makes it into the work your team does every day, you're not doing anything wrong, and you're definitely not alone. You've just hit the exact wall that almost everyone hits. Having the materials was never going to be enough. The hard part was always going to be the build.
If you're tired of agents that look good in a demo and then quietly stall before they ever go live, that's the part we'd love to talk through with you. Tell us how your team actually works, and let's figure out what it would take to build something that holds up in the real thing, not just the screenshot. Ready to chat? Schedule a 30-minute discovery call with our team.
Sources: Gartner, 2025 survey on AI agent adoption (gartner.com); MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” via Fortune (fortune.com).