If you had told me I'd be a consultant 3 years ago, I would've have laughed at the idea. That's because I read an article in high school that shaped my belief that consultants generally smart, but lacking in their own vision so they are hired to solve other people's problems. And every person I graduated with who ended up in consulting only reinforced that idea.

So how did I end up here?

How I got here

In going to market with Woz, I realized that AI coworkers for tech companies was extremely crowded. The Ramps and Stripes of the world are going to build it themselves. And the path to winning the YC startups of the world was to do what Viktor did and raise a ton of money and spend $2M/mo on distribution.

I am a competitive person, but it wasn't clear how I'd compete. I didn't have any unique insights, so my only option was to find a niche. So then the idea was to focus on mid-market companies struggling to roll out AI with little to no internal AI capability.

That led to a lot of interesting conversations with folks in the physical industries. A lot of startups are bringing AI to construction, but I don't really know anyone focused on woodworking or millwork.

And the first thing I realized in talking to these companies is that they are way earlier in the AI adoption curve than I thought. Most of these companies weren't even using off-the-shelf tools like Claude yet let alone full blown AI coworkers. And it sure as hell wasn't going to happen by itself because most employees remained skeptical about the hype around AI.

"Sure it can code, but that's because it was designed by people from Silicon Valley." Every conversation I had seemed to share this undertone.

After talking to a bunch of companies, I decided to actually step into 2 organizations to see what's going on.

How it's shaping up

The initial motivation was to FDE myself into companies and do whatever it takes to roll out AI successfully. But it was never meant to be permanent. The goal was to use this as a way to discover what products might be missing that someone would need to build.

Here are a few products that I think are missing so far:

• Non-technical evals - The idea is kind of like Braintrust & Langsmith except for a non-technical audience. Why? Because every conversation around AI inevitably falls back to evals. What is the ROI? What eval set captures the tasks it can and can't perform today. Can we switch to an open source model? Depends on how it will perform on the evals. And once you start asking yourself those questions, you realize that all knowledge work needs this. If you believe that tokens are emerging as the 3rd pillar of business expense, then the only way to control your inference cost long-term is to own eval data. It's easy to see this as an infrastructure problem, but I think the most interesting challenges here actually lie in UX.
• Collaborative knowledge work - If you continue down the thread above, you will eventually realize that knowledge work can benefit from adopting all primitives we take for granted in software engineering for collaborative work. Version and source control management systems like Git. Staging vs production environments. Pull requests. The naive answer is to just throw everything in Git, but that doesn't work with existing mid-market ERPs and accounting systems. A lot of people are working on this, so I'm excited to see who can figure this out.
• Custom connectors - The first step for most companies when they buy Claude is to configure all the connectors. You see immediate gains, and then you start to realize how terrible the connectors are. Why? Because Google wants you to use Gemini and not Claude with the GDrive connector. Same thing for Microsoft. There are companies like Composio and Zapier that exist to broker this, but I still don't think it's a better experience than just building your own custom connectors. But the big idea here is that I don't believe vertical specific agents will win. In a world where the models are post-trained on specific harnesses, Anthropic and OpenAI's agents will continue to be the best suited for their models in the same way that their models are best suited for specific GPU hardware. So increasingly it feels like the winning setup is custom MCPs + frontier lab agent.
• Internal app garden - In most of my engagements, what people really want is a custom dashboard. Claude artifacts are great for this, but they are limited in many ways as well. The artifacts you create in Chat don't play well with Cowork. And in many cases they are read-only. But what's powerful is that you can leverage connectors to respect the RBAC policies of the source systems. You might be able to open an artifact built on Salesforce data, but the data won't load unless your OAuth token has access to those objects in Salesforce. I know companies like OpenAI have internal app gardens and maybe that's what the Lovable enterprise plan is for, but there's definitely a product idea in here somewhere.

The more I do this, the more I think that AI-native consulting might actually be the big idea. In a lot of cases, what these companies need is strategy more than anything. To figure out what are the right areas to apply AI to.

From talking to ex-Palantir & folks from Thrive Holdings, I'm increasingly bullish on the idea of an AI FDE agent. Something that can be deployed into your Slack workspace, learn how your org works, and just start executing on a backlog. In some ways, this is the final frontier of how an AI-native consulting firm becomes mega scalable. It's doing what Cognition did to software engineering with Devin, but with forward deployed engineering.

But from having done this myself, I can confidentally say that it's very hard. Unlike Devin who works within a codebase that is defined by syntax and patterns, this FDE agent would need to do user interviews. It would need to understand how to do discovery and navigate complicated incentives and people relationships. It would need to know that sometimes people forget, but sometimes they lie to protect what it is that they really do at a company.

If I can automate this though, I have no doubt this would be a super valuable company.

Closing thoughts

If any of this resonates and you are interested in working on something together, please reach out at [email protected].