跳到正文
原文
Accounting Today(美国会计行业媒体)· Chris Gaetano·· 14 天前精选站内评分72

Clark推出面向CPA事务所审计项目的Clerks AI

Clark launches Clerks AI, an audit engagement team member

导读

Clark推出面向CPA事务所审计项目的Clerks AI,作为审计团队成员持续跟进任务和资料状态。它可按事务所方法选取样本、执行实质性测试、编制并更新工作底稿,并将需要人工判断的事项升级给审计人员。

推荐理由

Clerks将资料催收、抽样测试、底稿更新和人工升级串联为可配置流程,为审计自动化工具的落地方式提供参照。

正文
  • Key insight: Uncover how autonomous AI assistants are being deployed directly onto CPA firm audit teams.
  • What's at stake: Putting startup capital on the line to financially guarantee the accuracy of automated audits.

Clark, an AI-driven accounting platform, announced the launch of Clerks, described as an AI coworker designed specifically for CPA firms that perform audit work.

Processing Content

"For as long as people have kept accounts, someone has had to check them. Today that checking happens once a year, long after the decisions it was meant to inform," said Enes Witwit, founder and CEO of Clark. "Our Clerks don't sleep. They stay on the engagement, chase what's missing and keep the work moving between the moments an auditor looks at it. Today they work alongside auditors. We are building toward fully autonomous audit employees, and toward an economy whose claims are verified as they happen, so fraud and waste are caught while the money is still there."

In a later email, Witwit said Clerks could be thought of as an AI employee on the engagement team. Like a staff accountant, it has a defined set of things it can do, it pays attention to what is happening on the engagement, and it acts when something needs doing, without waiting to be told. Generally, each Clerk is assigned to an engagement the way a staff auditor would be and works through audit areas from start to finish: reviewing documents, performing tests and preparing workpapers. When something is missing, it requests it from the client, tracks the response and picks the work back up when the information arrives, without anyone prompting it again. 

The software keeps track of the live state of the engagement at all times: the status of every PBC request: outstanding, requested, received, incomplete or overdue; the status of every workpaper: not started, in progress, prepared, in review, review notes open, signed off or stale; the status of every procedure in the audit program, and whether it is blocked, ready or done; and every document that arrives, every client email, and changes in the client's accounting data. When any of these change, Clark notices.

Its perception comes from being connected to the systems where the engagement lives. The firm defines the working environment for each engagement: which systems it can see, which it can act in, and what it may do in each. Typically this covers things like the PBC portal and the firm's document storage, the engagement email box, Clark itself (e.g., its audit program, checklists and workpapers), the client's accounting data, and third party sites where source documents live like vendor portals or bank sides. Where a system has an API, such as SharePoint, Box or QuickBooks, Clark uses it. Where it does not, Clark logs into the website the way a staff accountant would. Changes reach Clark either because the system reports them or because Clark checks on a schedule. The CEO noted that, because Clark can work through a login, firms do not have to wait for the company to build a connector before Clark can reach a system.

Clark is always aware of the engagement's state, including nights and weekends. If a client uploads a document at 11 p.m. on a Saturday, Clark sees it then, checks it, and if a procedure was waiting on it, starts the work. Routine updates, such as the arrival of a requested document or a matching bank confirmation, are handled quietly, and what deserves an auditor's attention is escalated with the evidence attached. 

In terms of executing tasks, Clark acts from a fixed set of predefined actions. It can request a missing document from the client, naming the specific item, period and reason; follow up on an overdue request; check an incoming document against what was requested: right entity, right period, complete and legible; file a document to the PBC request it answers and update that request's status; extract data from documents (PDFs, scans, Excel) into structured tables; log into websites and pull data from systems that have no API, such as invoices from vendor portals or statements from bank sites; select samples according to the firm's methodology and document how they were chosen; run substantive tests by tracing sampled items to invoices, bank statements, payroll registers and plan documents, and flag exceptions with evidence attached; prepare workpapers in the firm's own templates, with every figure cited to the exact spot in its source document; mark workpapers stale when the data underneath them changes; update checklist status as procedures are completed; and, finally, escalate to the auditor when something needs a human decision.

Clark does not act on every change. It asks whether a change affects the work or the path to sign-off. If not, it records the change and moves on. If so, it picks the right action from its toolset. Further, Clark acts only through its predefined actions, and each action has a permission level set by the firm. Calculations, sampling and statistics run as actual code, not as arithmetic inside the language model, and every result is tied back to its source.

Continuing with the employee analogy, Witwit in the email added that users train the AI as if it were an intern. During setup, the firm writes down how it wants things done, in plain language. For example: "Always tie payroll tests to the quarterly 941s," or "Never contact the client's CFO directly, go through the controller." Clark applies these standing instructions every time, on every engagement where they apply. 

During actual engagements, the AI works from the firm's own audit program, checklists and workpaper templates, not a generic methodology, making sure it follows the firm's procedures in the firm's order and produces work in the firm's format. When a reviewer edits Clark's work or leaves a review note, Clark records the correction and does the work the corrected way next time, the same way a staff member who got a note on Friday does not make that mistake again on Monday.

Over time Clark remembers what it learns about each client and engagement: materiality, prior-year issues, how the client organizes its records, who to contact for what. A returning engagement starts with that knowledge instead of from scratch. Meanwhile, when an instruction is ambiguous or two instructions conflict, Clark asks the auditor rather than picking one.

Aiding its memory is the fact that Clark keeps a live model of each engagement. It knows the PBC list and each item's status, the audit program, each workpaper and its status, and which procedures depend on which documents. When a document arrives, Clark knows what it unblocks. When data changes, Clark knows what it invalidates. 

While it runs 24/7, it only spends AI tokens when something changes. Clark's awareness of the engagement does not involve a language model: the connected systems report changes, or Clark checks them on a schedule. A model runs only when a change comes in. Every change first goes to a small, fast model that screens out the ones that clearly do not matter. Those are logged and nothing else happens. The rest go to a larger reasoning model, which compares the change against the engagement's state and decides whether and how to act. The larger model runs only for changes that might matter. A quiet engagement costs close to nothing. A busy one costs in proportion to how much is actually happening.

While built for audit, the same platform can also serve client accounting services teams. Clark pulls source documents, reconciles them against the books, and prepares entries and schedules for the accountant to review.

Clark also stands behind the work its AI performs. The company offers CPA firms a 100% accuracy guarantee and a peer review guarantee. 

"A promise of accuracy only means something when it has money behind it," Witwit said. "If AI is going to do the work, the company that built it should answer for it."

The announcement comes on the heels of an $11 million seed round led by Touring Capital, with participation from returning investors Marathon Management Partners, Rerail and Sofeon Ventures. Together with the $5 million round Marathon Management Partners led in September 2025, the financing brings Clark's total funding to $16 million. Clark will use the funding to expand its senior team across product, engineering and sales, bring Clerks to more firms and more types of engagements, help CPA firms extend their audit services into government, and build toward fully autonomous audit employees. The company is also working toward continuous monitoring that keeps findings current between audits, with each finding stating what it covers and when it was last checked. The goal is for a firm to tell its client, a regulator or the public not just what was true at year-end, but what is true now.

Gaetano-Chris

Technology Editor

Chris Gaetano is the technology editor for Accounting Today. He brings with him more than a decade of experience covering the accounting profession... Read full bio

来源:Accounting Today(美国会计行业媒体) · accountingtoday.com

审计热点 是聚合摘要和阅读索引,原文版权归各来源所有,引用请以官方原文为准。如果你是来源方,希望更正、下架或调整展示方式,可以通过反馈页联系我们。

来自微信公众号的内容,本站仅呈现摘要与原文链接,不转载文章全文;文章权利归属原作者及原公众号

关于使用条款隐私信源