Founder notes
Why I stopped building Nila
I spent four months building an AI accountant. The product worked and one firm was about to try it. Here’s why I stopped anyway, and what I’d do differently.
I paused Nila on October 8, 2026.
By then the product was close to done. It could take a month of messy bills, read them, check the GST, put each one beside its source for review, then hand the accountant an Excel working and GSTR-1 file or post the entries straight into Tally. I’d written the build log and the extraction article about the hard parts. None of that was the problem.
What stopped me was what winning it would actually look like.
Why I picked it in the first place
After Kaguya, I was looking for the next thing to build, and I was scared of sales. I didn’t have any picture in my head of how cold outreach or selling to businesses actually worked, but I could tell it was the bottleneck everything else depended on. So I started with something close to home: an app for my parents’ business. They had three people reading emails, typing quotes and invoices into an Excel file with hundreds of tabs, and tracking purchase orders by hand. I thought I could automate most of that, and the first customer was already at home.
Then I read a YC request for startups. They were excited about AI-native companies that don’t sell software but just do the work, and “accounting, tax, and audit” was on their list. I thought, accounting is obviously something AI will automate, so I asked my dad how our auditor actually does it. I didn’t know the process, the terms, or even what Tally and an ERP were. I picked it all up gradually. I even had Zip2 in my head: Elon Musk saw the internet coming and sold to newspapers instead of every business. I’d sell to accountants instead of every business.
YC was asking for companies that do the work themselves. I ended up building software for the firms that do the work, basically because that’s what I thought I could do.
Looking back, it was a consensus bet: the idea came from what everyone agreed was coming, more than from anything pulling me. I picked it because it felt certain. In the end, that’s also why I stopped.
What the firms actually said
I approached 16 chartered-accountant firms and had around 25 conversations. Some were cold calls. Some came through my parents, whose auditors are two of these firms.
I wrote about those visits in more detail in What CA firms told me when I walked in. When I reviewed everything on September 29, here’s where things stood:
| Firms | |
|---|---|
| Approached | 16 |
| Real conversations about their workflow | 8 |
| Pain that matched what Nila did | 3 |
| Saw Nila demoed | 3 |
| Handed over real client files | 2 |
| Said they’d try it | 1 |
| Trial actually run | 0 |
The strongest lead was one of my parents’ auditors, a firm with about 120 monthly GST clients. Junior accountants typed scanned bills into Excel by hand, seniors checked them, and that Excel working became the firm’s record. Then they filed through Winman. Near deadlines people worked late into the night. I visited again and again, and they gave me real bills and the workings they’d made from them.
The second was my parents’ other auditor, with about 300 clients and 20 staff. I sat with one of their operators for two hours, watching them rebuild messy client spreadsheets inside Tally, fixing failed imports and re-entering what didn’t fit.
Both were friendly, gave me files and explained everything. And on September 9, when I went to show the first firm what I’d built, they said they’d try Nila in October next to their own process, to see how it compared. If it went well, the partner would introduce me to a friend whose CA office works with scrap dealers who send in 300-page bills. That’s exactly where Nila would have been most useful.
We set up a meeting for around October 7. By then I’d already decided to stop, and we never had it. That’s the part worth explaining.
How the firms actually work
Talking to firms, I found they mostly fall into two groups.
Old filing software plus Excel. Many firms have filed with the same decades-old software for years, like KDK, Winman or Computax. Every month they prepare an Excel working for each client and keep it as their record, so they have something to show when a notice comes.
Tally for everything. Others keep everything in Tally, and that gets awkward. Some clients send their sales in their own custom Excel format, which the GST offline tool can’t import. So the staff enter it into Tally, generate the return file from there, and then type the HSN summary into the GST portal by hand, one line at a time. The lady at one of these firms told me a single client can take a whole day. My parents’ second auditor works like this.
Nila already covered both: Excel workings and a GSTR-1 file for the first, and posting straight into Tally for the second. So the software side was handled. What I couldn’t change was how much any one firm needed it.
Why I stopped anyway
What I found in the offices was real, but none of it was the whole reason.
It was hard to target. Accountants are easy to find, because they have to be visible, but from outside you can’t tell how much GST work a firm does. Many had around 50 clients and could spread the work across a few staff, so it wasn’t a big deal for them. The one firm with a room full of fast Tally operators, where I think Nila could have taken over maybe a quarter of the work, came through my parents, not through Google Maps.
The pain wasn’t urgent. Some firms had a lot of manual work. Others finished a client’s sales in half an hour, or felt their existing software was fine. Annoying work they’d done for years didn’t feel like an emergency to them.
It was a lot of work for each sale. Even if Nila took an office’s cost from ₹1 lakh a month to ₹60,000, I could maybe charge ₹10,000 for it, and I’d still be selling one office at a time.
The real reason was something else. I could see the path where Nila works: win firms one by one, add more of their work after GST, and eventually build a real accounting business in India. I’d learned enough to believe I could do it. But it was a known path. Basis had already shown that AI accounting works. The questions were mostly answered, and I’d mainly be executing a known playbook. It was a reliability venture, not frontier work.
Hormozi has a line I keep coming back to: running a successful restaurant is hard, and running a successful software company is hard, so if the cost of hard is the same, you might as well go big. Even if Nila became big, I didn’t think it would feel like enough to spend my best years on.
The timing made it sharper. If this were the dot-com era, I think I’d have kept going, because I knew it could make money, and sunk cost would have done the rest. But AI is moving fast enough that a lot of white-collar work, maybe parts of accounting itself, could look very different in ten years. My guess is that what stays scarce is what AI can’t copy: atoms, capital, access, distribution. That made this feel like the wrong decade to spend on a known playbook. It’s just my hypothesis, but it’s the one I’m betting on.
So in early October I decided to stop. A great October trial wouldn’t have changed that.
I did spend a few days thinking about going after US clients instead, where the money is bigger. But that would have been starting over, in a market I knew even less.
When I told my dad, he tried to find a way to save it. Maybe we could use it in our own office. Or he could market it on the side and let it grow slowly, the way Kaguya had. He never asked whether it had all been for nothing, but I heard the question anyway.
What I got wrong
I built first and called it credibility. From the start, my dad told me to build everything first so firms would take me seriously. I think he was mostly encouraging me. By September I thought he was wrong, but for months I’d been telling myself the same thing: there wasn’t much point visiting firms without something to show. Discovery doesn’t work like that. If a firm truly has the problem, they’ll let you sit in their office before anything exists. If they won’t, a finished product won’t change their mind.
I didn’t talk to nearly enough firms. From May to August I mostly built, and by September 1 I’d really only talked to about four people. In September I finally went out properly and planned to work up to 20, then 50, then 100 firms, because that’s roughly what it takes to really see a market. I knew I could do two a day. I stopped at about 16. I still think stopping was right, because my real doubt was about what winning would mean. But next time I’d have that many conversations much earlier, before building most of the product.
I didn’t know what a design partner was. Kaguya taught me one way to find people: post on Reddit and see who shows up. That doesn’t work for accounting firms. I needed one firm that would let me come in every day, use Nila on their real work, and tell me what was wrong. I never set that up.
I went deep before anyone asked me to. I went through a GST course for a week, learned Tally, and then built a reliable product that handled GST edge cases most firms would never notice. All that detail is only worth it once a customer is pulling you toward it. I built it on my own guess.
Some of it I’m proud of. My first cold call took twenty minutes of staring at the phone. A few weeks later I was walking into firms on my bike, four in a day.
What I’d do now
If I started Nila again tomorrow, I’d first get inside an accounting office, through a job or a warm introduction, and do the work by hand for a while. Find out what’s manual, what’s slow, and what people actually complain about. Then ask the person who owns the firm whether fixing it is worth real money to them.
Only after someone says yes would I start building, and I’d build it in their office, with them.
The short version: prove the demand before you build. Let the market pull you.