Industry

AI in Accounting and Audit

Classification, reconciliation and workpaper drafting on infrastructure you control.

Accounting has more structured, rule-bound, high-volume work than almost any other profession, which makes it unusually well suited to automation. The blocker is rarely capability. It is that client financial data sits under engagement-letter confidentiality terms that a public API endpoint does not satisfy. Senteras deploys the models inside the firm.

Book a free consultation

What makes this hard

  • Busy season capacity is a hard ceiling: the work arrives in a ten-week window and hiring cannot flex to match it
  • Transaction classification and reconciliation consume junior hours that clients increasingly refuse to pay for
  • Client financial data under engagement-letter confidentiality cannot be sent to a third-party model without a conversation most firms do not want to have
  • Audit documentation standards require you to explain the basis for a conclusion, which rules out tools that cannot show their work

Where AI earns its place

Transaction Classification

Categorise general ledger activity against the client's own chart of accounts and prior-year treatment, rather than a generic taxonomy that gets it wrong on anything unusual.

Most of the coding work done before a human opens the file

Reconciliation and Exception Surfacing

Match across systems and surface only genuine breaks, with the matching logic visible so a reviewer can see why two items were paired.

Review time spent on exceptions, not matches

Audit Sampling and Testing Support

Risk-weighted sample selection with documented rationale, plus first-pass tie-out of supporting documentation to the population.

Defensible selection rationale in the workpaper

Workpaper and Memo Drafting

Draft technical memos and workpaper narratives from the firm's own prior conclusions on comparable facts.

First drafts, reviewed rather than written

Figures are drawn from Senteras engagements and are illustrative of typical results. Outcomes vary by data quality, infrastructure and scope.

The rules that shape the build

These are the constraints that decide the architecture, usually before anyone has picked a model. This is general information about the regulatory landscape, not legal advice on your obligations.

AICPA Code of Professional Conduct, confidential client information

Disclosing client information to an outside service requires the client's consent. A self-hosted model is not an outside service, which is why the deployment model matters more here than the model choice.

PCAOB and AICPA documentation standards

The workpaper must let an experienced reviewer who was not on the engagement understand the basis for conclusions. A model output with no traceable inputs fails that test.

IRS Circular 230

Due diligence obligations on tax positions do not transfer to a tool. Anything the model produces on a filing position is a draft for a practitioner to verify.

How we approach it

The model never writes the conclusion

In every accounting deployment Senteras builds, the model assembles evidence and drafts narrative but does not reach the conclusion. That is a deliberate design constraint: professional standards require the engagement team to own the judgment, and a system that quietly makes the judgment for them creates a documentation problem that only surfaces during peer review.

Where this applies

The same core systems, with the differences that matter in each setting.

Public accounting firms
Busy-season capacity is the driver; the payback is measured against contractor and overtime spend, not headcount.
Tax preparation practices
Source document extraction from the client organizer, where the input is a shoebox of PDFs and photos.
Internal audit departments
Continuous control monitoring rather than periodic sampling.
Bookkeeping and outsourced accounting
Transaction coding at volume; margin depends almost entirely on how much stays automated.

Common questions

Can we use this on client data under our engagement letter?

If the model runs on firm infrastructure, no outside service receives the data and the confidentiality question is straightforward. If you are considering a cloud service, the AICPA confidentiality rule points at needing client consent, which most firms would rather not request one client at a time.

Will a reviewer accept AI-assisted workpapers?

Only if the workpaper documents the basis for the conclusion the way any other workpaper would. Our builds record what evidence the model assembled and what it produced, and the engagement team reaches the conclusion. A system that reaches conclusions creates a documentation problem that surfaces at peer review.

Does this help with busy season capacity?

That is usually the business case. The comparison to make is against contractor and overtime spend in the ten-week window, not against headcount, because the constraint is capacity in a fixed period rather than annual cost.

What about tax positions?

Circular 230 due diligence does not transfer to a tool. Anything the model produces on a filing position is a draft for a practitioner to verify, and we build it to present as one.

How we build it

Local & On-Prem LLM Deployment

The most powerful AI models, running entirely on your hardware.

Custom AI Agents & Automation

AI that doesn't just answer questions. It gets things done.

Model Fine-Tuning & Integration

Models that speak your industry's language, trained on your data.

Document Processing

Extraction, classification and validation across formats you do not control.

Invoice Processing

Reads any layout, and tells you when it is unsure.

AP Automation

Handles the invoice formats your template-based OCR cannot.

Start with a conversation, not a proposal

Thirty minutes. We will tell you what we would change first, and whether you need us at all.

Book a call

The firm behind the firm