You Don't Need an AI Agent (Yet)

A client of mine was terrified of falling behind. Not by a competitor; a trend was chasing him, and trends never get tired. He had heard a lot of talk about OpenClaw lately, and he was ready to believe all of it.

He runs a small business, and every conversation we had circled the same fear: everyone is doing AI now, and we're not. He asked me about AI agents the way people ask about fire insurance, eagerly and without having read the policy. I asked him what he wanted the agents to do, and he told me he wasn't sure yet. He was ready to pay for a whole system before anyone asked what it would do, and that's how you end up with the answer locked in and the question still missing. I should have known right then that the request was the problem, not the solution, because I've been that client myself.

This post is about that question. I'll tell you what my client almost paid for, and I'll tell you what he needed in the end. It was cheaper than he feared, and it wasn't AI at all. I'll show you how we got there, because it's a route you can copy.

The client who was already behind

The panic was doing the talking, and I recognized it in the first five minutes. My client had a real problem, but it wasn't shaped like an agent: his projects live on different sites, he spends his days driving between them to keep the quality up with his own eyes, and he comes home too tired to look for new clients. He asked me about timelines and budgets, and I could see he was already picturing the demo in his head. I asked him what he wanted the agents to do, and I kept asking until the answers got small, but they never got smaller than "run the business", which was my first clue that we were dealing with a feeling, not a process.

Here's the part I remember best. He believed one tool could run large parts of his business, and the tool had a name: OpenClaw. He had heard a lot of conversations about how it magically streamlines business flow, and he had started repeating the magic himself: it checks and replies to emails, books calendars, searches the internet, and handles his daily routine while he sleeps. It was a beautiful picture of a desk job, and his week had almost no desk in it.

I remember thinking that this was a fear wearing a shopping list, and I kept that thought to myself for now. I don't blame him for the panic. The clock ticks for a reason: small business owners read the press, hear competitors claim AI in their pitches, and get asked by customers whether they "do AI" yet. Every one of those signals says the same thing: this is happening, and you're not in it. That feeling has made a lot of sensible people spend serious money.

But panic is a terrible way to gather requirements, and I say that as someone who's done his share of panicked buying. I told my client we'd start with the boring questions, and he agreed, mostly because I asked nicely. When you're afraid, you buy the shape of a solution instead of the solution itself, and you'll argue with anyone who tries to change the subject to the actual problem. My client had decided on the answer before the question existed. He was selling himself on agents the way people sell themselves on sports cars: he hadn't decided where he was driving, but he knew he wanted the leather seats. I'm not saying he was wrong; I'm saying he was shopping.

Not every problem is an AI problem

So I stopped talking about AI, and that was the whole trick.

I spent several rounds asking about his business instead: who does what, which tasks repeat, where the work gets stuck, which steps take two hours and show up every single week. We mapped his week on a whiteboard together and kept it boring on purpose, because we've all been burned by shiny, and the friction points showed up first. There were no agents, no models, no roadmap slides in those conversations, only the plain, true shape of a working week. I asked him to name the chore he hated most, and he answered before I finished the sentence: the drive between sites. What came out was unimpressive and completely normal: every "AI need" had an ordinary workflow problem hiding under it, and every workflow problem had a name you could write on a whiteboard.

I'll admit my first instinct was the same as his. When we finished the whiteboard, I wanted to propose an agent too, because it would have been a nicer story. But I've learned that the nice story costs more than the boring one, and the real answer was sitting right there in front of us.

His real bottleneck wasn't the routine he'd described. It was physical: his afternoons were eaten by driving from site to site to check quality in person, and no agent on earth can walk a job site. What might help there is something much more focused: a simple detection system using YOLO, the computer vision model that finds objects in images. Running on the photos he already takes, it flags what needs his eyes. And for the other half of his goal, finding new clients, he didn't need anything built at all. A ChatGPT or Claude subscription and a few decent prompts would have him drafting project docs and outreach emails by the end of the week.

Now the part I keep coming back to. He believed agents were a silver bullet for every problem at once, and I don't think that belief is his fault, because the industry manufactures it. I've seen that same confident certainty in plenty of demos, always wrapped around a happy-path story, and I've caught myself wanting to believe it too.

Watch an agent demo with me, and you'll see why I say that. The tool does the thing, smoothly, on the first try, in a tidy chat window, and it looks like magic, which is the point. What you don't see is the cases that break, the guardrails you have to build around it, and the person who has to watch it work. I'm not saying agents are useless; I'm saying the demo is a movie trailer, and I've sat through enough trailers to know the real film is mostly paperwork. I remember the first demo I ever fell for, and I don't want you to pay my tuition.

Reality is narrower, and I've learned this the expensive way: an agent is good at one well-defined task you can describe exactly, not at "run my business", and its failures are quiet. A human who messes up a step stops and says so, while a misconfigured agent mails the wrong thing to the wrong person, and you find out next week. I've had that phone call, and I never want to make it again. I'm not afraid of agents; I'm afraid of the pitch that sells them. I keep that phone call in my head as a reminder, and it keeps me humble.

Yes, the objection is coming, and it's a fair one, and I've heard it from myself more than once: an LLM could eventually do all of this better, maybe someday, but "someday, maybe" is not a requirement. A requirement is something you can act on this quarter with the money you have. If an off-the-shelf tool handles the chore today for the price of a coffee subscription, the better model that doesn't exist yet isn't a business case; it's a daydream with a GitHub repo.

Even the parts an LLM could genuinely help with, the docs and the emails, are subscription work, not build work. And the site checks, his real bottleneck? A multimodal LLM could look at those photos too, but a general model staring at every site photo is overkill; a focused detection system does the same job at a fraction of the cost. I know this dream well, because I've kept a few empty repos alive on it myself. I tell myself the same thing I told my client: the future can wait until it's my problem, and I'd rather ship the daydream with a real task attached to it.

Buy before you build

Here's the reframe that would have saved my client the most money: starting AI exploration does not mean building a pipeline. It means renting one. I told him that, and I watched his shoulders drop, because he'd been bracing for a much bigger bill, and it made me unreasonably happy. I've seen that relief before; it's the sound of a budget coming home.

The "we need a tailored solution" instinct is the expensive one, and I've felt its pull myself. A tailored build means scoping, hiring, architecture diagrams, and a maintenance burden that follows you forever. It feels like progress because it produces excellent meetings, while the chore it was meant to fix keeps happening by hand. I've sat in those meetings, and I've watched the to-do list refuse to shrink, and I know exactly why.

Take a generous budget: three or four of popular AI subscriptions like ChatGPT or Claude all at once, and you're still under a thousand dollars a year. I've lost more than that to a single afternoon of debugging my own code, and I didn't even get a logo out of it. For a small business, a little money on a service beats a lot of money on a build every single time, because someone else maintains the service, documents it, and keeps it running while you sleep. I keep that comparison in my head, because we both know how the numbers end.

So why do tailored builds keep seducing people? I think it's because they feel like progress, and progress is a nice feeling even when nothing improves. They produce diagrams, demos, and meetings where things are decided. A subscription produces none of that, and nobody holds a meeting about it, which is exactly the point. I've been in both meetings, and I know which one I'd rather skip.

You don't need Einstein to draft a daily email, and you don't need me to tell you that, but here we are. You need a decent typist who knows when to shut up. A small business doesn't need a research lab; it needs its invoices to stop eating Tuesday afternoons. I'd rather my clients hire the typist, and I tell them so at every chance I get.

Start small, spend on the actual problem

Here's the loop I use with every client, including the one I'm telling you about now: pick one recurring chore that annoys you. Not the whole business, a single chore: the weekly report that takes an hour, the site visits that eat your afternoons, the customer emails that all ask the same three questions. Rent a tool for that one thing and measure for a month, because a month of real numbers beats a year of opinions. If it saves more time than it costs, keep it and pick the next chore; if not, cancel it and try another one. That's the whole loop, and I'm slightly embarrassed to charge for it.

Let me be honest about the line, because I've crossed it the wrong way before. Sometimes off-the-shelf automation genuinely is not enough, and that is the moment an LLM or an agent earns its keep. When the chore is too weird for a template, too varied for a fixed workflow, and big enough that building around it pays for itself, then yes, an agent becomes the right answer. Notice the order, though: the cheap attempt comes first, and the agent is what you reach for after you've measured, not before you've understood. Trust me on this one; I've paid the tuition for skipping the cheap step, and I don't want you to repeat my course.

The discipline underneath is simple to state and hard to keep. Before any build, ask why, and then ask why again. I keep asking until the answer is a real task with a name, not a fear wearing a buzzword, and we both know which answers those are. If the answer contains the word "trend" or "behind", you're not done yet. I ask my clients this until they get annoyed, and the annoyance is usually where the real requirements live. My client got annoyed with me, and he forgave me when the invoice came in smaller.

Don't let the trend write your plan

The trend is not your requirements department, and I say that as someone who sits next to the trend all day. It's a salesman, and like all good salesmen, it wants you to decide before you ask questions. I've seen the invoice that follows that decision, and it's usually bigger than the problem was. I'm not saying the trend is lying; I'm saying it's busy. We both know this dance.

Most businesses that "need AI" need something else first, and my client's story was the proof: a recurring chore automated, a process understood, a workflow bought instead of built. Ask why. Then start small, and spend every penny on the actual problem, which will turn out to be cheaper than the panic. I've watched that arithmetic work, and I'd rather you didn't pay to learn it twice. It's the best deal I can offer.

And when a real AI need finally shows up, you'll be ready for it. You'll have a budget that isn't already buried in a pipeline nobody asked for, and you'll know exactly what it should do. I'd rather have that waiting list than a demo folder, and I'll know the need is real, because we'll have the same conversation we had this time.