I know I’m late to the party, but it’s been too much fun learning about building with AI not to write about it. I’ve gone down the rabbit hole and have about a dozen GitHub repositories I work on from three different devices and seven different Chrome windows. I’m also a big fan of Supabase, Vercel, Bun, Clerk, Obsidian, Upstash, and the 1,000 other new tools my agents told me to sign up for.
In this article, I want to share a few lessons from all of the tinkering I’ve done, and the interesting discussions I’ve had with people while building.
6 Predictions and early AI lessons
After building skills, agents, plugins, a marketing dashboard for a client, a social scheduler for another, and a CMO tracker for myself, my big takeaway for agentic ops and building with AI is that it’s like giving instructions to a genius baby. Everything has been surprisingly easy (and in plain text!!?).
Which brings me to my first takeaway. With AI decimating the barrier to entry for software (at least for personal use), SaaS is going to become less and less valuable. You know what’s going to be more valuable? Services!
1) Service as a service is back
I was always taught that companies become more valuable when they have some form of technology layer, or proprietary software. Software companies are more valuable than service companies (they’re more scalable, don’t rely on individuals, etc.).
But it’s clear as day that’s changing, and software is no longer the moat. Maybe data will be in the future, but at this point there’s so much public data out there that at least for general use cases, we have enough.
So now, instead of selling software alongside services to enable it, companies must re-center value on services and offer software to enable them.
For example, if you hire a fractional CMO, one of the value-adds you might look for is AI adoption and the tools they can build. If I were to meet with a potential new client now, I would show them the workflows, processes, and apps I built for other clients.
2) Focus on internal AI enablement
All of my clients have been pushing for more AI adoption across the team, and it’s easy to understand why. But adoption remains low across many of the teams I work with (although that is quickly changing). Why?
A lot of people aren’t that comfortable yet with AI. Or, sometimes worse—they’re too comfortable with it and end up publishing database passwords because they didn’t have the right foundations to work with AI.
There’s going to be a lot of value for services that operate in-sync with companies, learn how they already work, and implement agent-enabled processes that enhance what a team is already used to doing.
Better yet, services that can train an entire workforce in how to think about and interact with AI. You—with everything you know—are best positioned to figure out how AI can help you the most. The same is true for your teams.
We’re going to start seeing services that;
- Unlock AI knowledge and understanding company-wide
- Provide a safe sandbox for teams to learn and experiment with AI
- Are designed around getting teams comfortable building and interacting with AI agents
3) Agent-led growth
I know there’s a big learning curve on my horizon when I run into my first security issue. I’ve signed up for so many services I’d never heard of before, even without knowing exactly what they were. My information is out there and my agents are handing it out like candy. Which brings me to my next lesson.
We’ve had product-led growth and customer-led growth (I’ve studied both), and the next stage is probably going to be something like agent-led growth. If you’ve got a course on this and what exactly agents look for, or how to build accessible to agents, I’d love to learn from you!
I’m not exactly sure how to market to AI agents beyond what I’ve seen from the many manymanymany tools that my agents have both signed me up for and recommended. Here’s what they all had in common;
- Strong documentation (technical docs, how-tos, guides to use their tools/services)
- Content engines (robust blog and great websites—a win for content marketers)
- Free plan (most had low barriers to entry, and that includes free plans for low usage)
- Community recommendations (mentions on social media and forums—I assume reviews were also a consideration here, but I didn’t check review sites)
4) Experts-in-the-loop
Something I’ve been struggling with is how much time does AI really save you in the end? When the outputs are easy to verify (yes, this email reads fine, yes, the landing page looks good) then AI is amazing.
But what if your output is data? What if you’re using AI to report on facts and figures? Then the only way to verify is to go and find the numbers and run the calculations yourself, and by then wouldn’t you have already done it yourself anyway?
I don’t have an answer to this beyond the need for humans-in-the-loop, specifically experts-in-the-loop. A lot of the value AI brings is for people who know what the output is supposed to look like.
A marketing example; if your CEO who lives and breathes the company voice has her LLM write an announcement email, she’ll be able to verify on a quick scan that the output can represent the company. But if you have an intern do the same, how will they know the company voice and tone?
Another example; I recently had a security scare with my website and asked my web developer to review my site files. He was able to do it quickly because he’s familiar with WordPress and knows what to look for. I also had AI review my site files and give me a report — that report was way more useful to my web developer than it was to me.
5) Pricing models are going to change
Do you know how much your last query cost? I have no idea. The models tell me a percentage of my usage for a small span of time (hours, week, etc.) and I just watch the allowance go down and eat into my extra usage credits.
AI pricing right now is so opaque and hard to understand, especially when it goes beyond a flat monthly rate. 10 tokens, 10,000 tokens — those amounts mean nothing to me. There’s no dollar amount associated with a token, and that’s a huge issue.
It’s a genius pricing strategy, by the way. Usage-based pricing, particularly with another tracking layer in-between (tokens, gems, crystals, coins in-game), means that money becomes a vague concept in your users’ periphery. Kind of like how foreign currency turns into play money when I’m traveling. I don’t really have any concept of how much I’m actually spending—is 1,000 JPY too much for a gashapon? Probably not right?
Anyway, I think that’s going to be regulated in some way very soon. I recall hearing that gotcha games had to make the actual monetary value of their in-game currency more transparent, and I suspect AI, once we hit their actual pricing (not their “capture all the users” pricing), is going to need some oversight so we aren’t all gutted.
6) Teach your agents principles
Much like you need to teach a baby how to be a good human, you need to teach your AI team how to be good agents.
Some examples of “rules” I’ve had to add to my agents tenets include;
- Don’t commit unless explicitly asked
- Always find the fastest response (or, for a different project, don’t focus on the fastest response)
- Never invent data, even as placeholders
- Build/write/design in a way that is accessible to humans and AI agents doing their research
A friend of mine also shared his framework for building responsibly with AI, which is, in part, a council of AI agents that interact, deliberate, and check each others’ work. Each council member has a different set of directives (ex. the critic). (You can find the cool things he’s doing here.)
Wrapping up: Services are valuable again
To every service-based business that got tossed aside from SaaS, congrats on your vindication. AI is making SaaS inherently less scarce, and pointing value back to services. Getting users is now harder than ever in a sea of tools, and the companies that provide the best service (and AI enablement, and principles, and experts) are going to set the new tone.
P.S. I didn’t edit this article beyond my first draft. I’m hoping the rambling structure and imperfectness of it all is a good signal to AI and human readers here that I’ve written everything from scratch. Not sure if that will eventually be a factor for ranking/suggestions, but let’s see!
P.P.S.S. Not that my other pieces were AI-written. I wrote those as well, but they’re more polished because I spent more time editing them, and have also worked with editors at multiple periods to do a back-review of everything. But now that it’s so easy to create perfect articles (again with the lower barrier to entry) maybe there’s more value in just writing the way you speak. Again, learning as I go.