The Accountant Who Arrived After the Work Was Done

By the first business day of the month, an accountant at Roofstock can open an Excel workbook that appears to have done the night shift. Reports have been pulled. Schedules are built. About 75 tie-outs have run. A broken number is marked on the precise cell, with a proposed fix and a link to the transaction beneath it. The accountant still decides what enters the ledger. But the long, clerical walk to that decision has already happened.

That is the sort of scene Qurrent wants to sell. It builds and operates custom AI workforces for companies whose essential work lives in inboxes, spreadsheets, ERPs, and approval queues. Its customers pay for processes to be completed and measured, rather than for another piece of software that someone must learn to operate. For Roofstock, the process was monthly close across 52 entities and 16 institutions. Every entity needed a package of 30 to 50 tabs. A perfectly ordinary month was a small factory.

The job that comes before the judgment

The Roofstock close reveals the company's most practical idea. The agents operate inside the accountants' existing workbooks. They pull roughly 11 reports per entity, build workpapers, roll prior-month balances forward, and test reconciliations. If a variance appears, the workbook points to the exact cell and the underlying ledger item. Qurrent says the package is ready with zero variance at the start of the first business day. Its case study reports about a quarter of accountant time returned to analysis after a 12-week deployment.

There is a line the system does not cross. Only Roofstock's accountants can write a number into the system of record. That separation is less theatrical than the usual promise of an autonomous office, and much more interesting. A finance team can accept machine speed when it also gets a visible trail of what happened and a clear answer to who approved the result.

Roofstock first tried Qurrent on a different chore: moving residents out of managed properties. The agents took on about 80% of routine move-out tasks, including checking records, sending messages, and coordinating follow-ups. Roofstock reports a service-cost reduction of more than 30% and $1.6 million saved within two quarters. The monthly close came after that first operation had made the company comfortable giving digital workers more responsibility.

The coding agent took a detour

In early 2023, Colin Wiel and August Rosedale built an AI system that could behave like a small software-development team. It could take an objective, edit files, run code, and correct errors. They launched Qurrent that May and built an operating system for agents. Wiel later explained the turn: writing software was a demanding test of the technology, but plenty of business work followed more stable rules. The founders saw a larger near-term opportunity in operating those processes reliably.

Wiel had spent years inside labor-heavy property businesses, including Waypoint Homes and Mynd. That history matters. Anyone who has run a service operation knows the trouble is rarely a single task. It is the handoff after the task, the exception that lands in the wrong inbox, and the audit question six months later. Qurrent's answer was to make the agents a managed workforce. Its people map the workflow, deploy the system, watch it in production, and update it when a rule changes.

“We’ve evolved our operating model and driven greater operational efficiency by embedding AI and automation across Yahoo.”Matt Sanchez, COO, Yahoo

The mechanics are deliberately mixed. Generative AI reads unstructured invoices and emails. Deterministic logic handles financial transactions. Actions are logged. When an agent meets a low-confidence exception, it pauses and asks a person. The answer can be added to its runbook so the same case needs less attention next time. In a field fond of agents that sound confident, a system with permission to stop may be the more useful colleague.

Twenty thousand invoices, two inboxes

A large insurer shows the pitch from another angle. Its accounts-payable team faced as many as 20,000 invoices a month, arriving through two inboxes and three payment channels. A manual item could take 10 to 20 minutes and more than 20 steps. An OCR tool cost over $100,000 a year, yet its errors still needed correction. The insurer considered building automation itself or licensing a platform. Qurrent's managed model won because the buyer could purchase the result rather than assemble a new engineering project.

The deployed agents classify requests, check for duplicates, validate fields, enter data into Workday, and route approvals. Missing information goes to an exceptions inbox with a short explanation. The human job has become three steps: review the exception, tell the agent what to do, and approve the request. Qurrent's case study reports a 90% reduction in processing time and a 350% rise in requests handled in the first weeks. Those are company-reported early results, not a universal rate to paste into another finance department's budget.

Qurrent also works outside finance. Pacaso's Oscar Concierge answers homeowner questions at any hour, escalating with context when a person is needed. Pacaso reports that 78% of contacts are handled without a live agent, while 29% are resolved fully autonomously. Spire says its workforce settles 80% of targeted email support tickets without a person. Second Life uses agents to sort duplicate bug reports and examine an archive of more than 27,000 requests. These are different jobs, but all punish a service provider that treats the difficult last few percent as somebody else's problem.

A contract for the unglamorous part

Qurrent's market sits between traditional business-process outsourcing and software a customer must build or run itself. It usually prices by outcome or volume, not by seats, and says it puts accuracy or turnaround commitments into service level agreements. It connects to existing systems through APIs where possible and secure interface automation where necessary. Its public site says most deployments go live in weeks; Roofstock's complex accounting deployment took 12. The first question for a buyer is therefore less “Can the agent do this?” than “Can both sides define done?”

In March 2026, Cervin Ventures led a $15 million Series A, with Streamlined Ventures participating. Qurrent said then that its agents had performed more than six million operational tasks in production. By June, as it launched a finance-focused AI BPO platform, the company said the tally had passed 13 million. Counts at that scale suggest repeated use, though the work itself still has to be judged one contract and one process at a time.

The model has limits that make it legible. A process needs rules that can be mapped, data the system can reach, and a team willing to own the decisions that should remain human. High-stakes work also needs permissions and an audit trail. Roofstock's accountant, arriving to a prepared workbook, is still the accountant. Qurrent's wager is that removing the night's assembly work makes that person more valuable in the morning.