On a damp Tuesday in March, a partner at Meridian CPA looked at her second monitor for what she said felt like the first time in a year. On it, an agentic workflow had drafted thirty-one trial balances overnight — flagged eleven for review, deferred four to a human, and filed sixteen in a queue she had never seen before. She signed three before her coffee cooled.

"I expected a black box," she said later. "Instead I got a line item I could argue with."

"I expected a black box. Instead I got a line item I could argue with."

The system in question is not a chatbot. It is a small set of agentic routines, run nightly, with explicit checkpoints where a partner — never an analyst, never a marketing seat — has to put eyes on a draft before it leaves the building. Helpyr Technologies, the Nashville firm behind the work, calls this the review checkpoint, and treats it less like a feature and more like the actual product.

Across four engagements over the past fourteen months, the pattern has held. Throughput climbs. Errors fall, then keep falling. The partner's name on the deliverable is the asset. The agent is the instrument.

"We'd rather pick up the phone than schedule a meeting about scheduling a meeting."
— A house rule, since 2024.

The instinct, in a frontier market like agentic automation, is to sell removal. The pitch writes itself: fewer humans, more output, software does the rest. Helpyr's position is the inverse, and partners — the real ones, the ones whose signatures matter — have told us, repeatedly, that this is the difference between something they'll deploy and something they'll politely decline.

"The first phase of every engagement starts with a chair in the operations team's room — not a spec," reads an internal note from the firm, recently published in The Herald's engineering column. "The exceptions are where the work is."

See full engineering breakdown on B1 – B5; ledgers and unit economics on B4 – B5; the partner letter on A6.