Tax preparation and accounting: AI on the most sensitive financial workflows

Tax and accounting work is one of the longest-running professional services categories where AI has been quietly transforming the workflow without much industry fanfare. The largest firms have been investing in AI for years. The midmarket firms are catching up fast. The smaller firms are starting to feel the productivity pressure from competitors who have already adopted.

The work involves the most personal financial data in existence — every dollar earned, every dollar spent, every transaction made — flowing through systems that are increasingly AI-augmented. The data sensitivity, the regulatory framework, and the brand-trust requirements all favor local inference.

The work

Tax and accounting AI workloads include:

Tax return preparation: extracting structured data from source documents (W-2s, 1099s, K-1s, brokerage statements), classifying income and deduction items, identifying optimization opportunities, drafting return narratives. Volume scales with client count and complexity.

Document review and source-document analysis: reading client-provided documents at intake, organizing them, identifying missing items, generating intake summaries.

Audit and assurance support: reviewing client books and records, identifying control issues, drafting workpapers, producing client communications.

Tax research and planning: drafting research memos, summarizing tax law positions, generating planning recommendations, producing client deliverables.

Client communications: drafting status updates, organizer follow-ups, planning recommendations, year-end summaries.

Compliance documentation: drafting representation letters, generating engagement documentation, producing the documentation that PCAOB and state board examinations expect.

Practice management: drafting time entries, generating client billing, producing internal reports.

The math

A representative midsize accounting firm — say, fifty professionals serving a few thousand clients — generates a substantial AI workload across these functions.

Tax season alone produces tens of thousands of AI operations across return preparation. Year-round audit, advisory, and tax planning work add ongoing volume.

Aggregate AI workload at a midsize firm: in the millions of tokens per month, weighted toward operations with substantive token counts. At frontier pricing, the bill is in the low five figures per year for a midmarket firm.

For larger firms — top regional firms, national firms, the Big Four — the numbers scale dramatically. Major accounting firms with thousands of professionals and tens of thousands of clients have AI bills in the seven figures per year.

These numbers are growing fast as firms expand AI integration into more parts of their workflows.

Why tax and accounting is structurally a local-SLM case

The standard properties are present, with several at the extreme.

The work is narrow within each firm. Each firm has its specific client base, its specific tax planning approaches, its specific audit methodologies, its specific brand voice. A model fine-tuned on the firm's corpus outperforms a general model on the firm's specific work.

The work is repetitive in structure. Returns follow predictable structures. Workpapers follow predictable formats. Client communications follow predictable templates. Specialization compounds.

The privacy story is at maximum. Tax data is among the most sensitive personal financial information. Income, wealth, business operations, family structure, transactions — all of it on the return. For business clients, the return reveals operational details, customer relationships, profitability, and competitive positioning. Sending all of this to a third-party cloud LLM creates exposure that the firm's professional liability insurance and its professional standards both have positions on.

The professional responsibility framework matters. Tax preparers operate under Circular 230 in the US, equivalent frameworks elsewhere. CPAs operate under state board rules. Auditors operate under PCAOB standards (for public company audits) and AICPA standards (for everything else). Each framework has specific positions on client confidentiality and how client data can be handled.

The brand-voice and quality story matters. Client deliverables — research memos, planning recommendations, audit reports — need to match the firm's specific voice and quality standards. A general model produces generic accounting language; a fine-tuned model produces deliverables that match the firm's brand.

The audit trail matters for the firm's own professional defense. Workpapers, review documentation, and supervisory records are the firm's evidence of how engagements were conducted. AI augmentation needs to produce audit trails that map onto professional standards expectations.

What changes with local inference

A tax and accounting AI workflow on a local SLM looks like this.

A model is fine-tuned on the firm's corpus — historical returns (anonymized for training), audit workpapers, research memos, client deliverables. The fine-tune captures the firm's specific approach and voice.

The model runs on infrastructure the firm controls — typically in the firm's existing technology environment with appropriate access controls. The deployment respects the professional responsibility framework and the firm's professional liability insurance carrier's expectations.

Client work flows through the inference pipeline within the firm's controlled environment. Returns get prepared, workpapers get drafted, research gets summarized, communications get generated — all without crossing the security boundary.

The cost flips from per-engagement to fixed. Client growth doesn't scale the AI bill.

The professional responsibility story is preserved. Client confidentiality is protected by the architecture, not negotiated with third-party vendors.

The brand voice is preserved across all client deliverables.

What's specific about tax season

Tax season has a unique workflow shape that makes the local-SLM case especially strong.

Volume spikes dramatically from January through April (in the US — equivalent timing elsewhere). A firm's AI workload in those months is several times its baseline. With cloud-LLM pricing, the bill spikes proportionally — and the firm has to forecast and budget for the spike.

With local inference, the volume spike has no cost impact. The infrastructure is sized for peak; off-season, it sits idle. The economics flip from variable to fixed in exactly the season when variability hurts most.

The privacy story matters even more during peak season. Volume is high, attention is stretched, mistakes are more likely. Architectural choices that reduce attack surface (like keeping data local rather than routing through third-party LLMs) reduce the probability of a tax-season data incident, with all the reputational and regulatory consequences such an incident would have.

Where the cloud LLM is still acceptable

A few cases.

For research workflows operating on public tax authority materials — IRS publications, court decisions, regulatory commentary — without crossing into client work.

For internal training and CPE content that doesn't touch client data.

For practice management workflows that don't touch client work — vendor management, internal communications, generic operations.

For client work at any meaningful scale, the local-SLM case is structural and the privacy story is closer to mandatory.

The pattern, in the professional services workflow

Avery NXR is not a tax tool. It scaffolds Next.js applications. The architectural pattern repeats, with the privacy and professional responsibility dimensions making the case unusually strong.

Tax and accounting AI is a narrow, repetitive, volume-seasonal, extreme-privacy, professional-responsibility-relevant, brand-voice-sensitive workload. Every dimension favors local inference. The seasonal volume pattern is a unique argument specific to this category.

The tax and accounting AI vendors that build on local infrastructure — with appropriate fine-tuning, integration with tax software (CCH, Lacerte, ProSeries, Drake, UltraTax) and audit software (CaseWare, AuditFile, Inflo), and evidence packages that satisfy professional standards — will own the institutional tax and accounting AI market. The cloud-LLM-default products will hold pockets but face structural friction with professional responsibility and privacy expectations.

The pattern continues. Tax and accounting is one of the workflows where the architectural shift to local inference is being driven primarily by professional responsibility and privacy concerns, with cost as a secondary factor. Firms that move first will be ahead on professional standing and on cost simultaneously.