How Not to Startup / The AI Business Partner Trap
I've been a software engineer for 30+ years - Java (SE & EE), Android, C, C++, web, Javascript, SQL, Python.
I've had the privilege of being involved in some fascinating projects and delivering real value to truly discerning customers. I did run into some dead-ends along the way, usually due to someone being promoted well beyond their ability and specifying the wrong technology to be used in the wrong place. But that was a small minority - most projects delivered real value that made significant real-world difference.
I stopped working in March 2024 and did not touch a keyboard for months. But to be honest, I missed delivery! Knowing that someone needs something and delivering a product that meets and exceeds their expectations. Perhaps I'm strange but I like that whole (agile) requirements-to-delivery cycle. It's challenging but rewarding.
So I turned to my new best buddy - Artificial Intelligence, in the form of an LLM... ChatGPT, to be precise. Within minutes, I was fired up with a plan - roll out an Android app, gain users, build userbase, gather feedback, spot gaps in the market... and see where it takes us. Perfect! Organic, fluid, dynamic. No more hours than I want to commit. Users, demand, opportunity. And when the time is right, revenue, monetisation.
A year or so earlier, I had looked and failed to find an Android app that did what I wanted - a nice, clean, simple-to-use working hours logger, for keeping a record of my hours worked. If I had failed to find one that worked the way I needed then others would, too. But rather than go off half-cocked, I had a clear proposition in mind - a time logger aimed squarely at those who bill clients for their time - freelancers. They would be prepared to pay for a good app and I could draw in users with a free offering, only charging users once they wanted to charge their time to three or more clients.
Yeah.
I launched in February. As advised by AI, I promoted on the likes of LinkedIn, Facebook and Reddit... and here, on Indie Hackers. I knew the MVP was minimal (the clue's in the name, after all) and so would need extension and specialisation... but I was hopeful and optimistic.
The vagaries of analytics aside, user numbers grew to nearly 200 in March... but by April, the graphs were all headed in the wrong direction. Long story short, I now have 8 users.
I know the code is rock-solid and the app is sweet to use. And I had a 5-star review, too. Just a couple of minor issues in prod, quickly fixed. Growing features - monthly views as well as weekly, summary reports as well as session-based and data export. The design and development - my core skillset - has gone very well. It became more and more clear - the gap is clearly proposition and promotion/visibility.
And this is where the lesson was learned. Having been inspired and supported by AI, I looked to it for answers on how to get past this valley of death. I felt it was time to buy installs - with paid promotion. The AI said I would be wasting my money. It clearly had a very strong inclination to double-down on its earlier advice to promote harder and discover a few early adopters to help refine the proposition. I followed the advice and put a lot of effort into finding such users... but with little success.
When that approach was exhausted, the AI then started to contradict its own advice on which segment to target - it now said that my proposition was too narrow and I should widen my market by making the product more generic. Abandoning that original niche proposition was emotionally tough but you have to be flexible in this game. I re-positioned the app and all the surrounding collateral (store listing, graphics, etc.). I then spent further time promoting it to various new groups to try and find those early adopters who would help me build community and spot unmet need in the market.
When that also came to nothing, the AI then suggested that my proposition was too generic and that I should narrow it down... and one of the groups it suggested I target was the exact same one I had started with - freelancers. Right back to square one.
So my advice is to use AI... but use it wisely. In short, do not trust it. The one thing that LLMs like ChatGPT do very well is this: Illusory competence. Their grasp of English and their ability to corral supporting material and present supporting arguments is so good that it often takes an expert in the field to spot their weaknesses. If you are not an expert in the field then you will not know when it misdirects you.
But that's not the only thing they do well - they also do sycophancy extremely well. When I have challenged ChatGPT on this, it has told me that very few users notice its 'ingratiating manner'. But when I have asked other users, it is clear that they are well aware of its tendency to suck up. This combination of illusory competence with sycophancy can be an intoxicating mix.
I will continue to use ChatGPT for what I consider it to be best for - inspiration. It really is a good partner for batting ideas about and even brainstorming. But I'll be a bit more careful to test credibility before moving on from the 'Ideas' phase to the 'Action' phase.
I find it very interesting how you were able to gain 200 users. That is no small achievement for an app just entering the market. I'd be curious on where the users came from? What platform? What was their background?
It's interesting in the sense on how you couldn't get more from the same platform, even if it wasn't as many.
Hi @TheGauntletReport, that phrase 'The vagaries of analytics aside' in the article is doing some heavy, heavy lifting... as I explained in another IH article, early on I was overly trusting of analytics stats (or under-informed). It took some time to work out that every time I reinstalled in testing, that fired another first_open event. And what Firebase calls 'New users' is, in fact a count of first_open events. So my real user count was much lower; how much lower, I couldn't really say. In fact, I now have 8 active users and another 10 or so inactive. The majority of the real users that I know about and have had interactions with have been attracted via a LinkedIn post! Yeah, really. One gave a 5-star review and genuinely values the product despite not being in the target segment (freelancers). The other was from the target segment and gave some very useful feedback and suggestions for improvement. So it was really valuable.
I'm now thinking that I probably need to sweat LinkedIn some more...
This comment was deleted 22 days ago.
This is painfully relatable. LLMs are incredible for accelerating the build phase — they compress weeks of boilerplate into hours — but they are genuinely dangerous as strategic advisors because they have no skin in the game. They'll confidently tell you to pivot to enterprise one day and go solo-founder niche the next, with equal conviction both times.
The real pattern I've noticed: AI is best treated as a junior developer you can throw tasks at, not a co-founder you take direction from. It should accelerate execution, not shape strategy. The market feedback you got (8 users from 200) is worth 100x more than any AI-generated pivot suggestion.
On a related note, I've been building CostLLM, an OpenAI-compatible API gateway that helps small teams track usage and set budgets across different LLM providers. The irony of using AI to manage AI costs isn't lost on me, but it's a real problem indies face. Happy to share what I'm learning if useful.
No skin in the game. Indeed!! I didn't mention it in the article but in parallel with learning that ChatGPT's advice on app launch is 'unreliable', I also learned that its health and fitness advice is similarly dubious. It gave me a treatment regime for a musculo-skeletal issue which made the issue much worse. The world will learn not to trust LLMs... and that that lack of skin in the game is a major issue for those with skin in the game! As transformations shift from trial to production, this accountability issue is going to bite real hard.
I have been using ChatGPT for chatty launch-approach sessions and Gemini for coding assistance. I used Gemini since Android Studio ships with it pre-installed. I have found Gemini to be like a savant graduate - mostly very good and fast... but it does make silly errors and forgets stuff! It will also pick some awful architectural approaches, given a free hand. I generally review every commit as I've caught it doing too many silly things that would compile, pass cursory manual testing and ship... but destroy user experience.
One thing that resonated with me is that AI was giving you answers before you'd earned enough evidence to justify them.
When demand is still uncertain, the biggest risk isn't bad advice—it's becoming overconfident in explanations that haven't been tested yet. AI is incredibly good at making a strategy sound coherent. The market is still the only thing that decides whether it's correct.
You nailed the core issue: AI is phenomenal at coherence, not wisdom. It will confidently rationalize 10 different strategies because the skill it has perfected is 'making anything sound plausible.' The real lesson isn't 'don't use AI' but 'force every AI recommendation through real customer signals first.' Your 200 to 8 drop was painful but valuable data - way more useful than any pivot suggestion. Many founders won't get that clarity until much later (and after much more spend).
A useful guardrail is to force every AI strategy recommendation into a falsifiable experiment before acting on it: target segment, expected signal, budget, time window and a threshold that decides continue/stop. Keep that decision log immutable. Then the model cannot quietly rewrite the story after results arrive or send you from narrow to broad and back again without explaining which evidence changed. AI is good at generating hypotheses; market data has to own the decision.
I joined Ai developing industry its seems easy but due to Ai availability the competition is also very high
This comment was deleted 22 days ago.
Really, repositioning was driven mostly by lack of signal and lack of progress, from the target segment, at least... you could call it desperation. I did have someone with freelance experience provide really great encouragement, feedback and feature suggestions via comments on a LinkedIn post. But that was my only evidence that any freelancers had any interest in the product at all. My only review was a 5-star review from someone from a completely different market segment. She really values the product for completely understandable reasons. So the apparent lack of progress in the original target combined with clear applicability (and a dedicated user) in another prompted the pivot. So the initial idea to pivot came from myself but AI backed-up my intuition with compelling justification ("Great idea, sir! Here are lots of reasons why your idea is a really wise one"). Sycophancy mixed with illusory competence.
That makes sense. Lack of signal is a signal too, especially when another segment is showing even a small amount of pull.
The tricky part is separating encouragement from evidence. A detailed LinkedIn comment and one dedicated user are worth taking seriously, but probably not enough to let AI turn the pivot into certainty.
If I were testing this, I’d treat the new segment as a short validation sprint: 10 targeted conversations, one focused landing page, and a specific activation metric. Then AI can help structure the test instead of declaring the strategy.
“Sycophancy mixed with illusory competence” is a painfully accurate phrase.