卡在 AI 工作流的中间地带：确定性流程、AI 组件与智能体工作流

Whenever I hear about a new startup, I pull out my research playbook. First, I understand the pitch, then find backgrounds of the team, & tally the total raised.¹

Over the weekend, I decided to migrate this workflow to use AI tools, & the process taught me something important about how we’re actually integrating AI into our work.

Tools are small programs that expand AI capabilities. ChatGPT might call a web search tool to read a blog post I’d like to summarized. Claude might call the terminal tool to change file permissions in my current directory. Gemini might call a tool to find the latest stock price of the most recent IPO I’ve been following.

I replaced each step in my workflow with an AI tool: a web search & summarization tool, LinkedIn research tool, & a capital fundraising history tool. I hadn’t changed the workflow itself—just swapped out the individual components within it.

This upgrade revealed something crucial: there are three distinct classes of programs emerging in enterprise software.

Deterministic workflows are my original startup research process—the same steps, in the same order, every time. These excel at mechanization, executing identical processes with small deviations or calculations at each step.
Deterministic workflows with AI components represent my current setup. I still follow the same research sequence, but now Gemini & ChatGPT handle the summarization. The AI makes individual steps smarter while I maintain control over the overall process.
Agentic workflows hand decision-making to the AI entirely. The system decides what to research, in what order, & which tools to call based on the input.

These excel at handling broad universes of potential inputs—like customer support where a user might ask “Why won’t my password reset?” or “Can I integrate your API with Salesforce?” or “My data export is corrupted”—questions that require completely different investigative paths.

Security incident response works similarly: when an alert fires, an agentic system might investigate network logs, check for similar patterns in historical data, or escalate to human analysts based on threat severity—decisions that can’t be predetermined because each incident presents unique characteristics.

I learned two things from this migration:

Programming with AI tools is remarkably simpler. AI categorizes companies far better than any rule-based system I could write.
I hadn’t built an agentic workflow—I was just upgrading my deterministic process with intelligent components. & that’s exactly what I wanted.

I don’t want an AI deciding how to diligence a company. I want it to diligence every AI software company the same way, every time. The consistency of my process combined with the intelligence of AI gives me the balance I need: repeatable methodology enhanced by superior pattern recognition.

Maybe I’ll evolve toward fully agentic startup diligence someday, especially as the models improve.

But for now, this hybrid approach delivers the reliability of deterministic processes with the power of AI—& that’s the sweet spot for most enterprise applications today.