Startup Founder Shares Where to Leverage AI, and What It Can’t Replace

This as-told-to essay is based on a conversation with Tim Desoto, a 49-year-old founder and CEO, based in San Francisco. The following has been edited for length and clarity.

I’ve been working on my startup, an AI-powered shopping platform, since late 2024. With all the advancements in the tech industry, people might think of AI as a hammer, treating everything else as a nail, and in my experience, that just doesn’t work.

I don’t have a tech background, and since starting my business, I’ve learned a lot about where to leverage AI, and where not to. I’m always trying to be flexible about when to switch from AI to human intervention, and vice versa.

Over time, it’s become clearer to me where and where not to use AI for my business.

I rely on my network and social media to help me decide which AI tools to use

I’m fortunate to live in San Francisco. I go to meetups to hear what people are using, and I’ve attended some developer conferences. My ears always perk up when I hear that a new tool or version of something is working for somebody else in my network.

Right now, there’s a strong focus on agentic workflows. OpenClaw generated buzz as an open-source autonomous agent project, and Moltbook amplified that attention by making agent-to-agent interaction visible in a social environment. Claude Cowork is also gaining traction, particularly among teams looking for enterprise-ready agent workflows with clearer guardrails.

Beyond the agents themselves, the focus is moving from “what can agents do?” to “how do we run them reliably and securely at scale?”

Whether I’m looking at X, LinkedIn, or other platforms, there’s a lot of really great work being done to share these updates.

I use AI for all the typical time-saving tasks

My paid stack includes business plans for models such as Claude Max, Gemini Ultra, and ChatGPT Business, along with AI-powered development and productivity tools such as Cursor, Figma Make, Notion AI, Superhuman Ask AI, and Lovable.

Gemini’s image models have become incredible. The latest updates to the new model have really improved everything. I noticed faster performance, more stable reasoning, and stronger multimodal capabilities, especially in image generation. I was impressed by how consistent the images remained during modifications. I’ve even noticed some improvements in the responses it gives about real-time information.

I divvy up which tools I’m using based on where the latest developments are, and use them either as they’ve been designed or as I think I can use them in my current operational flow. For example, I use Lovable, an AI website builder, to make slide decks.

My ‘AI conveyor belt’ exercise helps me test different AI models

Usually, I start with a written prompt, then go multimodal, talking out loud to the model. I’ll talk back and forth with it about my idea and try to get the agent to push back because I know that some AI models tend to be more agreeable.

Once I get an output that I’m happy with, I use a different model to get a different view. For long-form analysis and structured insights, I lean toward Claude and Gemini. Gemini’s inline source linking is particularly useful for verification and deeper research. For structured reasoning and formal writing, I primarily use ChatGPT and Claude.

Sometimes, I’ll push a document out to multiple models at the same time and see what comes back simultaneously. For creative exploration and multimodal work, I use both Gemini and ChatGPT to generate early-stage concepts, mockups, and visual inputs. Some models are better than others at certain tasks, but I’m always getting a more well-rounded perspective by feeding content to multiple models.

The process can take as little as 15 minutes, or higher-impact decisions can span several hours to a couple of days, depending on complexity.

I eventually needed human developers

When I was vibe coding the alpha version of my product, I would hit spots where 30 or 40% would be wrong. I didn’t know what exactly the problem was. I would have multiple screens running the code to figure it out, and I’d continue to use AI against AI until I could get to about 95% confidence.

I contracted a few developers to help move my product forward. Now I have a product developing at a faster, much more robust, and scalable rate.

As much as I can do with AI, it’s amazing what technical people can do with AI tools that a non-technical person can’t.

The human perspective is still irreplaceable in my business

I reached out to a lot of informal mentors and friends in the space who could be helpful early on in my process, but recently I formalized having advisors to bounce things off of. That has been a huge feather in my cap.

These advisors have their own expertise to draw on, and they know many smart people in this space working on projects that have helped us identify potential blind spots. It’s helped me connect with potential partners in ways that I think would be more difficult as a solo founder.

I feel like I have a clearer view now of what I can trust AI for, compared to what I thought at the beginning of my journey. While issues like hallucinations and agreeability can be mitigated, long-term strategic judgment and taste still require human oversight.

AI can generate possibilities, but choosing the right direction remains a human responsibility.

Do you have a story to share about running an AI-powered business? Contact this reporter, Agnes Applegate, at aapplegate@businessinsider.com.