Teamday
← Back

Basis: Why AI Makes You 100x Faster but Not Your Company

AI Engineer
agentsenterprisebusinessknowledge-workstartup

How Basis Builds Accounting Agents That Firms Trust End to End

Mitchell Troyanovsky is one of the two co-founders of Basis, a New York startup that builds AI agents for accounting firms. Basis reached a $1.15 billion valuation in February 2026, and its research question is narrow: how to get agents to do production-scale work reliably enough that accounting firms use them end to end. In this AI Engineer talk, Troyanovsky starts with a gap most builders feel. Personal productivity with Claude Code, Codex or Cowork has exploded, yet the average company is nowhere near 10x more productive. "I know my productivity has gone up maybe 100x. But if I were working inside of a company, it wouldn't be anywhere near that order of magnitude."

Companies already run on non-deterministic workers: Basis's starting point is that "humans are already used to working with non-deterministic systems. It's just those systems are normally their co-workers, not their computers." The head of an accounting firm has no idea what the most junior staff do day to day, yet trusts the output because processes are set from the top. Even reconciling cash has a process, such as using the approved software to pull bank data rather than letting anyone pick a PDF extractor.

Two gaps that stall agent adoption: Coding-style agents, where you tell an agent to go do something, fail organizations in two ways. The competency gap: they push complexity onto users, who must manage skills and memories and have a theory of mind about the agent, which most knowledge workers can't do. The process gap: if every user customizes how the agent works, the work is defined bottom up, which is the opposite of how firms scale. Basis's answer is an agent that feels like a co-worker and behaves consistently. Being perfect isn't the goal: "It's okay if you can't do something. But you should be consistently able to not do something."

Reliability comes from opinions: There are endless ways to read a PDF, from a vision model to OCR libraries to five models checking each other. An agent told only to "extract PDF" may get good outcomes, but not reliable ones. Some opinions come straight from expertise: a tax agent that reads the primary source before answering is more likely to be right, and an agent that screenshots its PowerPoint before returning it makes fewer formatting mistakes. Opinions also trade accuracy for latency, which is a product decision, not a model one. Troyanovsky's tip on context: treat everything the agent sees "as runtime training data. The word prompt, I think, trivializes it a bit."

Behavior specs scale supervision beyond evals: Outcome-based evals with ground truth cover hundreds or thousands of cases, while production produces hundreds of thousands of situations across ever longer trajectories. Basis writes a behavior spec, a product spec for the agent era, listing only the opinions that matter most. One example from bank reconciliation: when the numbers don't tie, the agent should explain the difference instead of posting a plug entry. The spec doubles as a rubric. For any trajectory, even one without ground truth, a judge asks whether the situation came up and whether the agent behaved as specified.

Forking and simulations to find out why: When a trajectory fails a behavior check, Basis forks it at the decision point, recreating the environment, and asks the agent why it chose what it did or changes the context to see if the behavior changes. The cheaper tool is simulation: ask a coding agent to spin up a subagent that pretends to be the production agent, with the same system prompt and skills, and tell it not to cheat. Troyanovsky used this for more than a year and a half before Basis built forking internally.

5 Takeaways from Basis on Production Accounting Agents

  • Personal productivity doesn't transfer to firms by itself - Organizations scale through processes set from the top, so agents must follow them rather than each user's own setup.
  • Hide the agent - Users shouldn't manage skills, memory or theory of mind; working with the agent should feel like working with a co-worker.
  • Consistency beats brilliance - An agent that is great one day and forgets basics the next isn't reliable, even if its average is high.
  • Write behavior specs - A short list of the opinions that matter aligns the team and becomes a rubric for trajectories that have no ground truth.
  • Debug behavior with forks and simulations - Forking a failed trajectory or simulating the production agent shows why it failed, which outcome scores can't.

Why Opinionated Agents Will Win in Accounting and Tax

Troyanovsky closes on the argument that agents left to find their own way might sometimes find a better one, and that this doesn't matter: "Nobody wants their agent to move 37 their tax return." Trust grows the way it does with a new hire, as the agent proves reliable on narrow work and is given more. For another view of how regulated firms earn trust in agents, see NatWest's evaluation stack for customer-facing agents and Compound AI's finance agents.