Before Letting an Agent File Taxes, First Answer: What Happens When Payroll Stops Halfway?

Warp is a YC W23 company that does payroll, taxes, and HR. Founder and CEO Ayush Sharma gave a standard at YC Root Access: when building products, platforms, and architecture, every task should be executable by an agent, starting with deeply complex tax compliance workflows. Warp recently announced a $60 million Series B, serves more than 1,000 customers, processes $600 million in payroll annually, and plans to exceed $2 billion next year.

This standard can easily be heard as 'every company should add an agent.' Someone in the comments offered a harder test. Austin Amato said every company claims to be agent-native, but the test is simple: can a customer's problem be resolved without speaking to a human at all? The standard is blunt, but it gets close to the real issue in enterprise purchasing. What enterprises want when they buy software is for the problem to be fully handled; a demo only proves half of it.

Tax compliance in particular is not something to romanticize. Dusan mentioned that the German tax office still requires a human signature, no matter who did the work. ShadowAguy said state payroll tax compliance still involves faxing signatures to Delaware. These details are not cool, but they define the boundary of what an agent can do. Models can read forms, fill them out, and check them, but legal responsibility still lands on people, signatures, and audit records.

Pingu asked something more specific: what happens if payroll stops halfway through? Can the agent recover without paying people twice? This question is closer to production systems than 'can it fill out forms automatically?' Payroll is not a one-off Q&A; it is a workflow with state, batches, rollbacks, and reconciliation. If failure recovery is not done well, automation just amplifies errors.

Lincoln Powers's comment was a bit sarcastic: with agents doing tax compliance, it is the first time people are excited to hand over the part that gets audited. The sarcasm is useful. Tax and payroll are high-value because they are troublesome, rigid, and error-prone. Precisely for that reason, trust has to come from a clear audit trail; marketing cannot solve it. Eggs inSpace has run payroll for a small team for years and said the compliance grunt work is exactly what agents should take over, but trust can only be solved by a clear audit trail.

An Jiaoshou's judgment was more direct: starting from a hard workflow like tax compliance is far more substantive than wrapping a chat box around something. Warp's choice fits this logic too. It did not first build a general-purpose agent and then look for a use case; it chose unsexy but must-be-accurate workflows like payroll and taxes. For AI practitioners, the hard part is putting agents into systems that have permissions, state, audits, and human-signature nodes. For ordinary users, one way to judge whether an AI product is reliable is to see whether it dares to show how it handles failures.

Lunar Misfit's joke hit the mark: soon the paycheck will say, 'You're right, it should be higher.' It is a joke, but it also reminds us of something. Once agents enter payroll, sounding human is useless; the numbers have to reconcile, and responsibility has to be traceable.

The full discussion is in YC Root Access's 'How Warp Went From YC to a $60M Series B.' Video link: https://www.youtube.com/watch?v=80eZ7DjCGSo