Tax Authorities and Preparers Get New AI Risk Frameworks

Tax Authorities and Preparers Get New AI Risk Frameworks

As artificial intelligence tools proliferate across the tax industry, a new set of risk frameworks aims to help both government tax authorities and private tax preparers navigate deployment decisions more systematically.

The frameworks, detailed in The Tax Adviser, establish separate risk registers for the two groups—identifying 20 distinct risk areas for tax authorities and 15 for tax preparers. Each register organizes risks into categories and provides illustrative scenarios and controls to guide responsible AI adoption.

Why it matters

Tax is structurally well-suited for AI because it operates on deterministic logic, relies on massive labeled datasets, and is anchored to codified law. But failure modes carry high stakes: incorrect guidance can cause financial harm at scale, and bias in audit selection can undermine public trust. Without domain-specific frameworks, organizations lack a common language to assess their own readiness and conduct vendor due diligence. These registers provide that starting point.

Four risk categories for tax authorities

The tax authority framework groups 20 risks into four quadrants:

Information integrity covers hallucination (confidently incorrect outputs), drift (gradual accuracy decay), explainability failures, over-trust by staff or taxpayers, and unintended policy changes embedded in AI outputs.

Fairness and legitimacy addresses bias in outcomes, disparate audit targeting, trust erosion from visible failures, and public perception of mass surveillance.

Security and data includes prompt injection attacks, data leakage to vendors, re-identification of anonymized records, adversarial manipulation of detection models, and risks from autonomous AI agents taking irreversible actions.

Institutional capacity warns of workforce expertise erosion, scope creep beyond approved use cases, control dilution as AI scales, lack of auditability, regulatory noncompliance, and governance gaps as AI gains autonomy incrementally.

The framework acknowledges that risks often conflict: making AI more transparent reduces explainability risk but can increase adversarial manipulation risk by revealing how detection models work. A three-tier priority system helps organizations determine which risks must be addressed before launch versus during operations.

Five risk categories for tax preparers

The preparer framework organizes 15 risks into five categories: technical (hallucinated advice, client data leakage, incorrect document extraction), practice (failure of professional judgment, misplaced authority), legal and regulatory (increased malpractice exposure, regulatory sanctions), business (erosion of client trust), and workforce and strategic (loss of expertise needed to practice responsibly).

Both frameworks emerged from literature reviews, established enterprise risk models, discussions across the tax community, and direct experience leading a tax authority, according to The Tax Adviser. The publication notes that responsible deployment begins with choosing the simplest AI method suited to the task, as "AI" encompasses generative models, machine learning, natural language processing, and other tools with distinct risk profiles.

The frameworks include an eight-step methodology for applying the registers to specific deployment decisions, covering scope definition, risk identification, prioritization, and control assignment.

The detailed risk frameworks and methodologies were first reported by The Tax Adviser.