Stop Building "AI Agents." Most of What You Need Is a Workflow.
Lenny just published a framework that quietly says what a lot of small business owners need to hear: you don't have an agent problem, you have a workflow problem.
I just read Lenny's newsletter this morning on AI agents, and the framework he shared is going to save a lot of small business owners a lot of money. I had to break it down because I see this exact mistake every week with the people I work with.
The whole point of the post: most of the stuff people call "AI agents" isn't actually agents. It's automation with an LLM stapled on. And treating those two things as the same thing is the reason so many AI projects feel like they're burning cash.
The three buckets — pulled straight from the post
Lenny breaks AI work into three categories, and the percentages he puts on each are the part I keep thinking about:
Category 1 — Deterministic Automation (60–70% of opportunities). This is the boring one. You define the workflow yourself, and an LLM just handles content at specific steps. Email triage. Lead routing. Invoice classification. Pulling unstructured stuff into structured stuff. The tools here are the ones you already know — n8n, Zapier, Make.com. Timeline to ship: weeks.
Category 2 — Reasoning & Acting (ReAct) (25–30%). This is what people think they're building when they say "AI agent." The LLM picks what to do next from a set of tools, observes the result, and loops. Tools: LangGraph, CrewAI, AutoGen. Timeline: months.
Category 3 — Multi-Agent Networks (the rest). Multiple specialized agents talking to each other across domains. He basically says: don't go here yet. Reserve it for later-stage stuff once the simpler categories are paying their bills.
What I love about this framing — and what I'd add my own spin to — is that 60–70% number. That's not a small piece of the pie. That's most of the work most businesses actually need.
In my own business, this lines up exactly
I'll tell you the truth: the AI work I've shipped that has actually mattered for my business has almost all been Category 1. My assistant — the one I've trained on my tone and voice and all my context — that's a deterministic system with an LLM doing the talking. My content pipeline is a workflow. My lead-handling is a workflow. The places where I tried to be too clever and let an LLM "figure it out" are the places where things broke or felt unreliable.
The small business owners I talk to keep wanting to skip Category 1 and go straight to "a smart agent that handles everything." And I get it — it sounds cooler. But it's expensive, it's slow to build, and it's harder to debug when something goes sideways. You end up with a beautiful Category 2 demo and a Category 1 problem that's still unsolved.
The signs you've outgrown a category
Lenny gives clear warning signs for each category. The ones I keep hearing in my own conversations:
- Your workflow has more than 30 nodes. That's the moment a deterministic flow starts feeling like spaghetti, and it's a real signal you've graduated into ReAct territory.
- The requests you're handling are genuinely ambiguous. If users could ask for anything, a fixed flow won't keep up.
- You need cross-domain reasoning. A workflow can't decide between two different sub-systems on its own.
If none of those describe your situation — and for most online businesses, none of them do — you don't need to step up a category. You need to ship the workflow.
The success metrics also change by category
This is the piece I'd push hardest on with a small business. Lenny calls out that you should be measuring different things at each level:
- Category 1: completion rate, error rate, time saved.
- Category 2: reasoning accuracy, tool-use precision.
- Category 3: coordination efficiency, hand-off quality.
When I see people get into trouble, it's usually because they built Category 2 and they're trying to measure Category 1 outcomes — "why isn't this just doing the task?" Because you didn't actually build a thing that does the task. You built a thing that decides what task to do, and then sometimes does it.
Where I'd add my own spin
The thing Lenny doesn't say outright but is implied throughout: most small business owners are in the minority right now if they're not using AI at all. The next leg up isn't "build an agent." The next leg up is "automate the things you keep doing with your hands." Email. Content repurposing. Lead enrichment. Reporting. Internal docs.
If you can't list five Category 1 workflows you'd benefit from, you're not ready for an agent. You're ready for a Saturday afternoon with [ClickFunnels](https://www.[clickfunnels](https://www.clickfunnels.com/signup-flow?aff=39183cc3-2122-42a8-9eb8-b295ed7d8554 "Try ClickFunnels").com/signup-flow?aff=39183cc3-2122-42a8-9eb8-b295ed7d8554), n8n, or Zapier and a list of repetitive tasks.
The Bottom Line
Most "AI agent" projects in small business are workflows wearing a costume. Build the workflow first. Ship it. Measure it. If and only if it actually graduates — 30+ nodes, ambiguous requests, cross-domain needs — then step up to ReAct. Multi-agent networks come last, if ever. The mistake isn't being ambitious. The mistake is paying for Category 2 complexity to solve a Category 1 problem.