Clark's AI Clerk Can Chase Audit Evidence at 11 P.M.

Picture the most ordinary audit delay: one missing bank statement stops a procedure on Friday afternoon.

The client uploads it at 11 p.m. Saturday. No staff member is watching the portal. The file sits there until Monday.

Clark wants its new Clerks AI to notice the upload, check whether the document is complete, attach it to the right request, and resume the blocked work before anyone asks.

That is a bigger promise than drafting a memo from a prompt. It turns AI from a tool someone opens into a worker that watches the engagement.

Each Clerk is assigned to an engagement much like a staff auditor.

The system keeps a live record of what is happening: every client request, workpaper, audit procedure, document, email, and change in the client's accounting data. A request can be overdue. A workpaper can be in review or stale. A procedure can be blocked, ready, or done.

Those labels create the map the AI uses to decide what happens next.

If a requested document arrives, the system can see which procedure was waiting for it. If underlying data changes, it can mark the related workpaper stale instead of leaving an old conclusion in place.

The important idea is not that the software works at night. It is that the software remembers what the night-time upload changes.

Clerks is designed to do more than monitor a dashboard.

Its listed actions include requesting missing documents, following up on overdue items, checking whether evidence covers the right entity and period, extracting tables from files, selecting samples under the firm's method, and tracing sampled items to invoices or bank statements.

It can also prepare workpapers in the firm's templates and cite each figure to the exact location in the source document.

Clark's product site describes the same basic control goal: evidence, workpapers, and review live in one engagement, with draft work tied back to sources.

That source trail matters. A reviewer needs to move from a number in the workpaper to the supporting evidence without trusting a polished answer on sight.

Clerks does not keep a large language model running all day.

Connected systems report a change, or Clark checks them on a schedule. A small, fast model screens out events that clearly do not matter. Only a change that may affect the engagement goes to a larger reasoning model.

A quiet engagement should therefore use very little model capacity. A busy engagement costs more because more events need decisions and actions.

That design also limits noise. The auditor should not get an alert every time a file moves. The system is supposed to escalate the changes that affect the work or the path to sign-off.

During setup, a firm gives Clerks standing instructions in plain language.

One rule might require every payroll test to tie to quarterly Forms 941. Another might forbid the AI from contacting the chief financial officer directly. The system then applies those instructions across the engagements where they belong.

The firm also defines which systems the AI can see, where it can act, and what permission level each action receives. Clark can use an application programming interface when one exists or work through a website login when it does not.

Calculations, sampling, and statistics run as code rather than mental arithmetic inside the language model. When instructions conflict or remain unclear, the system is supposed to ask the auditor.

Those details are the difference between automation and improvisation.

Clerks can prepare, test, trace, and follow up. That does not make it the engagement partner.

Auditors still have to evaluate whether the evidence is sufficient, resolve exceptions, challenge unusual results, and make the judgments that lead to an opinion. The product's value depends on whether it clears routine work without hiding the basis for what it did.

Clark says the system learns from review notes. If a reviewer corrects a workpaper, the AI records the change and applies the corrected method next time. That could reduce repeated coaching, but it also makes change control important. Firms will need to know which instruction changed, who approved it, and which engagements inherited it.

Clark offers CPA firms a 100% accuracy guarantee and a peer review guarantee for the AI's work.

The company's guarantee page and main site emphasize traceability and peer review support. Its site also lists SOC 1 Type 2, SOC 2 Type 2, and ISO 27001 credentials.

A guarantee is useful only when the contract is clear. Firms should examine what work is covered, what counts as an error, how losses are measured, which exclusions apply, and how quickly a claim must be made.

The promise does not transfer the auditor's professional responsibility. It does give buyers a concrete place to test how much risk the vendor is willing to accept.

The company recently raised $11 million, bringing its total funding to $16 million. It plans to expand Clerks, reach more engagement types, and build toward continuous monitoring.

That vision changes the timing of assurance. Instead of learning at year-end that a control or balance went wrong months earlier, a firm could keep selected findings current as data changes.

There is a long road between watching an engagement and delivering continuous assurance. Data access, client permissions, audit standards, evidence quality, and human review all have to keep up.

The first test is simpler. Give the system one real engagement, one controlled set of permissions, and one pile of messy evidence. Then see whether the reviewer can trace every action, understand every exception, and tell exactly when the AI stopped and asked for judgment.

Working at 11 p.m. is the flashy part. Producing evidence another auditor can trust on Monday morning is the part that matters.

For more plain-English accounting news, follow Ledger Lowdown.