Kinter: agents that run your accounting close proactively, not on demand

I started coding at 10 and working professionally at 17. Five years ago, I built Alloy Automation (YC W20), an integration platform backed by a16z and YC that connected SaaS tools and powered financial workflows for companies like Amazon.

Last year, I made a call that felt risky: walked away from a growing business because I could see the moat eroding. Integration infrastructure was getting commoditized. We pivoted to Kinter.ai.

Here's what we saw: The real problem in accounting is execution capacity, not data. Most finance teams close their books 14-21 days after month-end – not because they lack information, but because the work (reconciliation, accruals, journal entries) still happens manually. Every "AI" tool in the space is basically a co-pilot: you ask it something, it helps you do the work faster. The work still lands on a person.

What we built: AI agents that run accounting workflows proactively, inside existing ERPs (QuickBooks and NetSuite), without waiting to be prompted. One agent does payroll. Another monitors for accruals throughout the month and drafts proposals. Another reads invoice attachments to extract prepaid expense candidates. All actions go through a human approval step before touching the books. Full audit trail on everything.

https://www.youtube.com/watch?v=jhWdPmLRGOk

What surprised us building this:

1. The hardest part isn't the AI, it's the data. We had an early customer where we were silently missing 2,300 bills because of a pagination limit in the QBO API. No error, no warning. Just missing data. Reconciliation built on incomplete data isn't reconciliation.

2. Finance teams aren't afraid of AI, they're waiting for it. We’ve had over 600+ conversations with Controllers & CFOs. The reaction to "an AI that does your close" wasn't fear. It was a relief. People went into accounting to solve hard problems, not to copy numbers between spreadsheets at 11pm.

3. Trust is built incrementally. We shipped human-in-the-loop approval flows before broad autonomy. The confidence threshold to let an agent post a journal entry is higher than the threshold to let it flag an anomaly. We design for that explicitly.

4. The integration layer we built at Alloy turned out to be the prerequisite. You can't build a trustworthy accounting agent if you can't reliably read bank feeds and write back to ERPs. The connectivity problem isn't solved by the model.

See more at kinter.ai and read our manifesto here: kinter.ai/manifesto