The AI Cost Panic Doesn't Apply To Your Business
Every week there’s a new headline about AI costs spiraling out of control. Token bills in the millions. Agentic systems burning through compute. Enterprise AI budgets doubling.
And every week, I watch founders and executives read those headlines and quietly decide that AI is too complicated, too expensive, or too risky to touch right now.
That’s a mistake. And I think it comes from a fundamental confusion about what kind of AI we’re actually talking about.
The discourse is about a different problem than yours
The cost horror stories are real, but they’re almost exclusively about one thing: overly complex, multi-agent orchestration pipelines built at scale, often without clear ROI guardrails. We’re talking about autonomous systems making thousands of LLM calls per task, with memory layers, tool chains, and retry logic compounding at every step.
There are valid use cases for that kind of complexity, but it represents a small slice of how AI actually gets used in practice. Most businesses (including mine) need something much simpler: targeted automation for specific, repeatable workflows. The economics of that are completely different.
What my actual setup looks like
I run ODS Tech with a handful of tools that took days to set up, not months. And I try to bring custom agents to the platforms I already use, like Slack.
My recruiting agent “Zizou” is a Slack bot connected to a backend that uses the Claude Agent SDK. It has a detailed prompt describing the talent profile I’m looking for and tools to search and evaluate candidates. I describe a need in Slack. It shows me a set of profiles, then I reach out personally to assess.
My engineering agent “Ronaldo” is another Slack bot. I’ve created Slack channels that map to specific GitHub repositories so the agent has the right context. I drop a task in the right channel, the bot opens a Pull Request in the respective repo, and GitHub Copilot reviews it for anything I might have missed. If Ronaldo can’t do it or I just feel like doing it myself then I use Claude Code, which has genuinely changed how fast I can move on a greenfield problem. Yes, that one has a real cost but it fully justifies the productivity gain.
For prototyping and validating new ideas, I use Lovable. When a concept is worth exploring but not yet worth a full build, I can have a working prototype in front of a client or stakeholder in hours. It’s changed how I approach early-stage validation entirely: less “let me scope this out” and more “let me just show you.”
For the business side (SOWs, proposals, project briefs, recurring documents) I have Projects in Claude, GPTs in ChatGPT, and a set of reusable prompts and templates that handle the boilerplate. What used to take hours now takes minutes.
None of this is exotic. None of it requires an AI engineer on staff.
The hidden cost isn’t compute. It’s configuration.
Here’s what I actually spend time on: writing good prompts, being precise about what I want each tool to do, and connecting the right systems together with intention.
The businesses losing money on AI aren’t losing it on tokens. They’re losing it on scope: building general-purpose systems when specific-purpose tools would have done the job at a fraction of the cost.
The question worth asking isn’t “can we afford AI?” It’s “do we know what we actually want AI to do?”
If you have an answer to that second question, the economics almost always work out.
What this means for your business
You don’t need a six-figure platform contract or a dedicated ML team to get real leverage from AI right now.
You need clarity about your highest-friction workflows, a willingness to experiment with focused tools, and someone who can connect the pieces without overbuilding.
The gap between businesses getting real ROI from AI and those still watching from the sidelines isn’t budget. It’s specificity.
If you’re trying to figure out where to start, happy to share what’s worked. Reach out or leave a comment.