I run an agency called OpenLoop. The tagline is "AI-first software products and agency." I put that in the description because it's accurate, and I want to be honest about what it actually means in practice — because the gap between what AI is marketed to do and what it does in a real business is significant, and I have spent the last year building inside that gap.
This is not a post about AI being overrated. It's a post about what changed, what didn't, and what I've learned running real client work with real AI tools in a city that doesn't have the luxury of paying for hype.
What actually changed
Code generation is real. I want to say that clearly before anything else. The ability to describe what I need in plain language and get back working code — not perfect code, but working code — is a genuine shift in what a small team can produce. At OpenLoop, three people are now covering what would have required six or seven two years ago. That's not a pitch. That's the actual number.
First drafts come faster. Whether it's a client proposal, a technical specification, or the boilerplate structure of a new feature — AI reduces the time from blank page to something workable. The blank page is the expensive part. Once something is on the page, editing is fast. AI has made the blank page cheaper.
For a team in Srinagar competing for clients who could hire agencies in Bangalore or Hyderabad, this matters more than it might elsewhere. The competitive advantage of a big-city agency used to be speed and output volume. AI compresses that gap significantly. We can move as fast as a team twice our size, which means we can price more competitively and take on more scope without burning out.
What didn't change
Client trust still requires a human. This is the thing the AI marketing consistently underestimates. When a client is deciding whether to trust a team with their business — their actual systems, their actual data, their actual livelihood — they are not reassured by hearing that you use AI. They want to know: is there a person behind this who understands my problem? AI cannot be that person. A human who uses AI well can be.
Complex requirements still need deep thinking. The cases where AI fails badly are the cases where the requirements are ambiguous, the system has significant constraints, or the problem is genuinely novel. Those cases are also the most valuable ones. Clients pay more for hard problems. If you let AI draft the solution to a hard problem without doing the hard thinking first, you get plausible-sounding code that doesn't actually work. Debugging that is expensive.
AI-generated code has bugs. This is not controversial — everyone who uses it seriously knows it. The question is how often and how severe. In my experience: often enough that you can't skip review, and occasionally severe enough that a missed review would have caused a real incident. The right mental model is a capable junior developer who works very fast and needs supervision. Not a replacement for developers.
The hype problem
Every agency now says AI-first. Some of them mean it. Many have added AI to their pitch deck without changing how the work actually gets done. The clients who've been through a cycle of this — who hired an "AI-first" agency, got mediocre output, and are now more skeptical than before — are a growing segment of the market.
This creates a real opportunity for teams that operate honestly. The pitch is not "we use AI so you'll pay less." The pitch is "we use AI to move faster without growing the team, and we're disciplined about where it helps and where it doesn't." That's a harder pitch. It requires specificity. But specific commitments are the only kind that can be kept.
The agencies that will survive this cycle are the ones where AI is a tool in the hands of people who know what they're doing — not a replacement for knowing what you're doing.
What it means to be AI-first in Kashmir
There's a version of "AI-first" that only works when you have unlimited runway, access to the best models, fast reliable internet, and the flexibility to iterate indefinitely on a client engagement. That version doesn't exist in most places.
The version that exists here: you use the tools you have, you're precise about where they add value, and you don't have the margin to let AI mistakes compound into expensive problems. When a project goes wrong in a large city, you have a buffer. You can absorb a bad week. We don't have that. So AI gets used where it genuinely helps — initial structure, boilerplate, documentation, first drafts — and humans take over where it doesn't.
What I've found is that this constraint actually makes the AI more useful, not less. When you're deliberate about where you apply it, you don't get into the failure modes that come from over-relying on it. You use it for the right 60% of the work. You do the other 40% yourself.
That ratio is probably the most honest summary of "AI-first" I can give: 60/40, with you deciding which 60.
The honest version
AI is a significant tool. It is not a transformation. The people treating it as a transformation — and building company positioning or client promises around that framing — are going to have a difficult year as clients get more experienced at evaluating what they're actually getting.
The durable advantage is not having AI. Everyone has AI now. The durable advantage is the judgment to know where it helps, the discipline to supervise it properly, and the honesty to tell clients what they're actually buying.
Those things were true before AI. They'll be true after whatever comes next. The tool changes. The work doesn't.
If you're building something with AI — or thinking about hiring a team that says they are — I'm happy to talk through what that actually looks like in practice. Reach me at me@mehranshahmiri.com.
Related: Why I'm Not Raising Money · On Building at Twenty-One