Last week, during office hours, I met the 10th AI startup in a brief period of time that pitched me this problem:
“I have substantial recurring revenue, but our clients need something so different, it feels like I’m running an agency. I look at Lovable’s growth, and our numbers are nothing compared to that”
Startups include: automatic screening of job applicants, AI powered single source of truth for construction, LLM powered processing of complex regulation documents, AI agent that lives in your company’s Slack environment. These are all the somewhat low-hanging fruits of the AI craze.
It resembles the early 2000s when all companies realised they needed a website1. Therefore, many like me who had the least programming skills started selling websites. AI is a bit like that these days, as most of those web-developing agencies weren’t scalable. Company, yes, startup, no.
The recipe
A stark contrast between aspiration and result. How do they get there? The recipe is as follows. A smart founder with tech capabilities does outreach to industry enterprises or SMEs, pitches that they can solve a problem with AI, the company has a relevant problem, they build something, get MRR, and move to the next.
After five such projects, they become either a prototype factory or an agency building and maintaining custom AI software, with related maintenance contract revenue. The overlap between those projects? Too limited to scale.
The core of the technology they built overlaps between 30% and 60%, but a lot of the work is manual configuration of agents, data preparation, process setup, and pipeline implementation. To make matters riskier, they rely heavily on external models.
In the best situation, they do something unique with that model, something others have a hard time replicating because the founders are truly smart and have an insight others don’t and therefore have a competitive advantage. If that’s the case, you have a shot at winning. In my understanding, competitive advantage is crucial if you don’t have deep marketing pockets.
In other cases, the founders are glorified LLM-wrapper implementers. Nothing wrong with that; I believe you can grow to €10M ARR, but it doesn’t scale like turnkey SaaS or at least some simple onboarding SaaS. In that case, very often, your competition is not another startup, but customers using Claude themselves, or worse: the in-house dev team that can build what you build in 2 to 3 months. Guess what they often settle for.
In that latter situation, I meet the founder, almost panicking, asking if they have missed something. They stress that they don’t want to build an agency, and what they are running now doesn’t resemble the fast-growing AI startup they had in mind. What to do next?
The antidote
There’s a third option, and that requires some proper analysis and scrutinising your GTM strategy, and being willing to kill darlings and clients.
Even though that just 30% might overlap, that could be your core. A unique JTBD that you can serve so well that any in-house dev team would rather just implement your approach than build it themselves.
Likely, this is a smaller project than the megalomaniac projects the corporate clients are currently asking you for. They just see you as cheap developers that they can order around to kill 10 AI projects on their backlog with one stone. Likely, you gonna have to start to say no to particular customers, even though you can definitely build it; it’s too far separated from your core.
The problem was not that you got there because your technology sucks. The problem is that AI is such a general technology that it has a very broad application potential. That is unique; not all technologies are like that. But that also invites this scattered projects approach, which hinders building a diversifying core.
Here are some questions to move forward in identifying a core that might turn into a startup:
Who are your happiest customers? Cluster them into groups.
Who ripped your product out of your hand? Who had a sluggish sales process? Focus on the smooth sellers.
What JTBD did they hire you for?
What alternative option didn’t they hire?
What can other founders with some tech skills also build with off the shelf solutions in a month or two? Scrap those ideas.
Which of these remaining things is scalable? Select that one and start selling before building.
I suggest that, for each cluster, you complete a market fit canvas and find where you have the strongest problem-solution fit and the best competitive advantage. Link here:
If you need help with your analysis, first 30-minute call is always free. Book here.
I forgot who made this comparison; somebody told me somewhere half a year ago over Zoom. If this was you, thanks, let me know!
I wonder if the dangerous bit is that the agency version can look successful for quite a while. You’ve got customers, recurring revenue and plenty of demand, so saying no to custom work can feel like deliberately slowing down a business that’s working.
This maps to what Palantir figured out with forward-deployed engineers: high-touch delivery can fund product discovery, but only if you're ruthless about extracting the repeatable 20% into an actual product. The job-screening and support-AI founders you describe seem stuck because they never made that extraction, the service just became the product. Did any of your 10 have a real roadmap to kill the customization layer, or were they just hoping growth would paper over it?