Caught within the Center of AI Workflows by @ttunguz

Every time I hear a couple of new startup, I pull out my analysis playbook. First, I perceive the pitch, then discover backgrounds of the group, & tally the overall raised.¹

Over the weekend, I made a decision emigrate this workflow to make use of AI instruments, & the method taught me one thing necessary about how we’re really integrating AI into our work.

Instruments are small packages that develop AI capabilities. ChatGPT may name an online search instrument to learn a weblog publish I’d prefer to summarized. Claude may name the terminal instrument to vary file permissions in my present listing. Gemini may name a instrument to search out the most recent inventory worth of the newest IPO I’ve been following.

I changed every step in my workflow with an AI instrument: an online search & summarization instrument, LinkedIn analysis instrument, & a capital fundraising historical past instrument. I hadn’t modified the workflow itself—simply swapped out the person elements inside it.

This improve revealed one thing essential: there are three distinct lessons of packages rising in enterprise software program.

• Deterministic workflows are my unique startup analysis course of—the identical steps, in the identical order, each time. These excel at mechanization, executing an identical processes with small deviations or calculations at every step.

• Deterministic workflows with AI elements characterize my present setup. I nonetheless comply with the identical analysis sequence, however now Gemini & ChatGPT deal with the summarization. The AI makes particular person steps smarter whereas I preserve management over the general course of.

• Agentic workflows hand decision-making to the AI solely. The system decides what to analysis, in what order, & which instruments to name primarily based on the enter.

These excel at dealing with broad universes of potential inputs—like buyer assist the place a consumer may ask “Why gained’t my password reset?” or “Can I combine your API with Salesforce?” or “My knowledge export is corrupted”—questions that require utterly completely different investigative paths.

Safety incident response works equally: when an alert fires, an agentic system may examine community logs, verify for comparable patterns in historic knowledge, or escalate to human analysts primarily based on risk severity—selections that may’t be predetermined as a result of every incident presents distinctive traits.

I discovered two issues from this migration:

1. Programming with AI instruments is remarkably easier. AI categorizes corporations much better than any rule-based system I might write.

2. I hadn’t constructed an agentic workflow—I used to be simply upgrading my deterministic course of with clever elements. & that’s precisely what I wished.

I don’t need an AI deciding the best way to diligence an organization. I would like it to diligence each AI software program firm the identical means, each time. The consistency of my course of mixed with the intelligence of AI offers me the stability I want: repeatable methodology enhanced by superior sample recognition.

Possibly I’ll evolve towards totally agentic startup diligence sometime, particularly because the fashions enhance.

However for now, this hybrid strategy delivers the reliability of deterministic processes with the ability of AI—& that’s the candy spot for many enterprise purposes as we speak.