AI can make tax compliance materially more efficient by helping people and systems search regulations, summarize changes, classify products, map data, reconcile transactions, detect anomalies, prepare filings and automate repetitive workflows. The limit is that tax compliance often requires a correct result, not merely a plausible one. Generative AI should therefore be treated as an assistive and orchestration layer around authoritative tax content, validated rules, deterministic calculations, human approval and audit evidence rather than as the final source of tax truth.

Darko Pavic’s practical rule Use AI aggressively where it can propose, prioritize, explain and prepare. Use validated knowledge and deterministic controls where the system must decide, calculate, report, sign or prove compliance. AI proposes. Validated knowledge governs.

---

Table of contents

• What is AI in tax compliance?
• What AI in tax compliance includes — and what it does not
• Why AI matters for tax compliance now
• The Four-Level AI Tax Compliance Control Model
• Where AI can create real value in tax compliance
• What tax-technology providers are doing with AI
• How a controlled AI tax-compliance architecture works
• Impact on retail and POS systems
• AI compared with adjacent tax-compliance technologies
• International and jurisdictional variation
• Common misconceptions about AI in tax compliance
• The hard limits: why tax compliance cannot simply trust an LLM
• Can AI reach 100% accuracy in tax compliance?
• Darko Pavic’s perspective
• Research agenda: making AI safer for tax compliance
• Implementation considerations for retailers and solution providers
• Frequently asked questions
• Source list

---

What is AI in tax compliance?

AI in tax compliance is the use of generative AI, machine learning and related intelligent systems to support tax research, classification, data mapping, reconciliation, risk detection, filing preparation and compliance workflows. In compliance-critical processes, AI should complement rather than replace authoritative tax content, validated rules and deterministic controls.

What AI in tax compliance includes — and what it does not

“AI in tax compliance” is an umbrella term, not a single technology. The useful distinction is between the kind of intelligence being applied and the legal consequence of its output. Generative AI is strong at language and semantic search. Machine learning is strong at pattern detection, prediction and classification. Rules engines are strong at repeatable logic. Workflow automation is strong at moving data and tasks. A mature tax architecture combines them instead of asking one model to do everything.

Capability	Best fit	Main risk	Appropriate role
Generative AI / LLM	Search, summarize, compare, explain, draft, extract candidate structures, natural-language interfaces.	Hallucination, stale context, source mismatch, weak applicability reasoning.	Assistive unless output is independently validated.
Machine learning / predictive AI	Anomaly detection, risk scoring, classification, prioritization, pattern recognition.	Bias, drift, false positives/negatives, weak explainability.	Use thresholds, monitoring and review.
Deterministic rules / tax engine	Tax calculation, explicit eligibility logic, mandatory validations, fiscal sequence rules.	Wrong rule modelling, stale rules, configuration errors.	Primary control for repeatable compliance-critical logic.
RPA / workflow automation	Move files, collect data, submit forms, route approvals, create tasks.	Automates a bad process consistently if controls are weak.	Useful when inputs and actions are well defined.
Agentic AI	Orchestrate multi-step work, call APIs, investigate exceptions, prepare or execute actions.	Autonomy amplifies errors and permission risk.	Needs policy boundaries, tool permissions, validation, logs and rollback.

AI in tax compliance does not mean that tax law itself becomes probabilistic. A VAT rate, fiscal signature requirement, invoice format rule or statutory filing deadline remains a legal or technical rule even if AI is used to find, explain or implement it. The technology should make compliance work easier without silently changing the authority of the underlying rule.

Why AI matters for tax compliance now

Three changes are happening at the same time. Tax rules are becoming more digital and more granular, tax authorities are receiving more transaction-level data, and businesses are introducing AI into the systems that create and process those transactions. The result is a compliance environment in which speed and data quality matter much earlier in the transaction lifecycle.

The OECD reports a rapid rise in AI use inside tax administrations. In Tax Administration 2025, 69% of administrations reported AI implemented and in use in 2023, with another 24% implementing it; 2024 data show AI widely used for risk assessment and fraud detection. The OECD also describes AI being used for taxpayer services, correspondence support and decision assistance. [S1] [S2] [S3]

This creates an asymmetry for businesses. Tax authorities can increasingly compare large data sets, identify patterns and target inconsistencies quickly, while many company tax functions still reconcile POS, ERP, invoice and return data in separate tools. AI can help close that operational gap — but only if it is connected to reliable source data and a governed compliance model.

The Four-Level AI Tax Compliance Control Model

A useful way to decide whether AI belongs in a tax process is to classify the use case by the consequence of an error. The more directly a model can change tax, invoice, fiscal or reporting outcomes, the less freedom the model should have to improvise.

Level	Typical use	Error consequence	Control principle	Retail example
Level 1 — Assist	Search, summarize, translate, explain, draft, developer support.	Low to moderate; output is advisory.	AI may generate freely, but important facts should be source-linked and verified before external use.	Tax research assistant; API documentation assistant.
Level 2 — Recommend	Product classification, account mapping, risk scoring, anomaly prioritization, change-impact suggestions.	Moderate; bad recommendation can create downstream exposure.	Confidence score, evidence, thresholds, human review for exceptions or high-impact cases.	Propose a taxability category for a new SKU.
Level 3 — Prepare & orchestrate	Prepare filings, build mappings, configure rules, generate test cases, remediate errors, create submissions.	High; output can alter a compliance process.	Deterministic validation, approval gates, version control, environment controls and audit trail before execution.	Generate an e-invoice mapping then validate against schema/business rules.
Level 4 — Execute	Calculate tax, sign fiscal transactions, submit legally relevant data, accept/reject compliance-critical actions.	Very high; an error becomes a legal or transaction error.	AI may orchestrate, but validated rules and deterministic controls govern the decision. System must be able to abstain or fall back safely.	POS tax/fiscal execution in a live transaction.

Architecture principle The closer AI gets to production execution, the more it should behave like a controlled operator of validated compliance capabilities rather than an author of tax logic.

Where AI can create real value in tax compliance

The highest-value use cases are not limited to chatbots. For retailers and solution providers, AI can improve nearly every stage before and around the final deterministic compliance decision.

1. Regulatory research and knowledge access

Use case	Practical value
Source-grounded Q&A	Ask natural-language questions across laws, authority guidance and technical specifications while returning source passages, jurisdiction and effective date.
Summarization	Condense long consultation papers, technical specifications and rulings into operationally relevant points.
Regulatory comparison	Compare requirements between countries, versions or channels and identify meaningful differences.
Change detection	Diff new and old regulatory documents and propose which obligations may have changed.
Multilingual assistance	Translate and align concepts across languages while preserving the original source as the authority.

2. Product, service and master-data classification

Use case	Practical value
Product tax categorization	Interpret product descriptions and attributes and propose taxability categories. Vertex and Sovos publicly describe AI-based approaches to this problem. [S15] [S16]
HS/customs classification	Propose tariff classifications and flag uncertain or restricted items; Avalara publicly includes cross-border classification in its agentic compliance portfolio. [S11]
Tax-code and account mapping	Map ERP, POS or product-master fields to tax engine categories or filing schemas.
Entity and registration data quality	Detect inconsistent tax IDs, legal entities, registrations, addresses or master-data relationships before they affect transactions.

3. Transaction analytics, reconciliation and risk detection

Use case	Practical value
Anomaly detection	Find unusual tax rates, missing invoice numbers, unexpected zero-tax transactions, outlier stores or unusual return patterns.
Cross-system reconciliation	Compare POS, ERP, e-invoice, fiscalization and return data to identify gaps before a tax authority does.
Risk prioritization	Rank exceptions so scarce tax expertise is focused on the cases most likely to matter.
Pattern detection	Detect repeated error signatures, configuration drift or clusters of transactions associated with compliance incidents.
Mirror visibility	Build a view of what tax authorities are likely to see from transaction-level and declarative data; Sovos markets this concept through its compliance intelligence/reconciliation layer. [S14]

4. Document intelligence and operational case handling

Use case	Practical value
Invoice and notice extraction	Read tax notices, invoices, certificates and registration documents, extract key facts and route them to the right workflow.
Exemption certificate handling	Extract, validate and monitor certificate data, with exceptions routed for review.
Notice triage	Classify authority correspondence, identify deadlines and prepare a proposed response or task list.
Audit preparation	Assemble source records, mappings, explanations and evidence packages for review.

5. Filing, e-invoicing and regulatory workflow automation

Use case	Practical value
Data mapping	Map source ERP/POS data to filing or e-invoice structures and suggest corrections when formats change.
Return preparation	Prepare calculations, reconciliations and draft filing data while keeping final validations explicit.
E-invoice validation assistance	Explain validation errors, suggest mapping changes and classify rejection causes.
Registration/onboarding	Guide users through tax registrations, data collection and configuration; Avalara publicly describes AI-guided onboarding and rule setup. [S11]
Error diagnosis and remediation	Identify likely root causes and propose fixes; Sovos describes AI remediation actions that can execute after user approval. [S13]

6. Engineering, testing and support for compliance software

Use case	Practical value
Developer assistant	Answer API questions, generate example payloads and explain errors. fiskaly publicly offers an AI assistant trained on its documentation; Avalara describes AI-powered API support. [S18] [S11]
Test generation	Generate positive, negative, boundary and exception test scenarios from validated requirements.
Schema and mapping review	Compare interface payloads against technical specifications and flag missing fields or incompatible versions.
Incident analysis	Summarize logs and transaction traces, correlate failures and propose root-cause hypotheses.
Synthetic test data	Generate non-production scenarios that cover country rules and edge cases without exposing customer data.

7. Agentic tax compliance

Use case	Practical value
Workflow orchestration	Agents can call tax services, gather evidence, open cases, prepare returns or manage exceptions across systems.
Proactive monitoring	Agents can continuously watch for expiring certificates, failed submissions, regulatory changes or reconciliation breaks.
Business-system integration	MCP and A2A patterns are beginning to expose tax capabilities to enterprise agents; Avalara and Sovos have publicly described such interfaces. [S12] [S13]
Bounded execution	An agent may execute an approved task only if the compliance package, permissions and deterministic validations all pass.

What tax-technology providers are doing with AI

The table below summarizes public vendor materials reviewed for this page. These are vendor-described capabilities, not independent accuracy certifications. The purpose is to show where the market is investing, not to rank products.

Provider	Publicly described AI/compliance direction	What it illustrates
Avalara	Agentic Tax and Compliance; tax insights; AI-guided onboarding and rule configuration; returns; exemption certificate validation; generative tax research; cross-border classification; notices; e-invoicing; MCP servers.	AI embedded into ERP, POS, e-commerce and compliance workflows; agents observe, advise and execute. [S11] [S12]
Sovos	Ask Sovi; problem diagnosis and remediation; business-specific recommendations; proactive issue detection; regulatory intelligence; classification; VAT data mapping; natural-language analytics; reconciliation; MCP/A2A foundation.	Strong emphasis on governed AI, approval before corrective actions and audit transparency. [S13] [S14] [S15]
Vertex	Smart Categorization for U.S. retail sales-tax product mapping; confidence scores; human-in-the-loop approval; audit packages; public strategy also references certificate AI, assistants and determination workflow support.	A good example of using generative AI to propose classification while preserving review and audit history. [S16] [S17]
fiskaly	Pursuing an “AI-first” strategy through internal AI hackathons, AI internships and prototype development. Public examples include natural-language interfaces for customer onboarding and API configuration, workflow automation and data analysis. Its current public AI activity is primarily focused on improving product interaction and internal processes rather than positioning AI itself as the authoritative tax-compliance decision layer. AI assistant trained on developer documentation; core compliance stack for signing, reporting, export, archiving and e-invoicing remains API-driven.	Shows a useful separation between AI-assisted developer experience and deterministic fiscal execution. [S18]
efsta	efsta is exploring AI both as a knowledge-access tool and as a way to derive intelligence from large volumes of fiscal transaction data. Its public work includes the “Olaf” chatbot for fiscalization questions, while efsta has also discussed using AI to identify patterns and irregularities across transactions rather than looking at each transaction only in isolation. This points toward AI-supported monitoring and anomaly detection as an additional layer on top of fiscalization infrastructure.	Uses AI for fiscal knowledge access and is exploring transaction-pattern analysis to identify anomalies and compliance-relevant signals across large volumes of fiscal data. [S19]

How a controlled AI tax-compliance architecture works

The safest architecture separates the probabilistic strengths of AI from the authoritative controls of tax compliance. AI can interpret language, identify patterns and orchestrate work, but it should not silently become the legal source of truth.

1. Authoritative sources are collected with jurisdiction, source type, publication date, effective date, version and legal status.

2. AI retrieves, compares and extracts candidate facts, obligations, mappings or classifications. At this stage, the output is a proposal, not a production rule.

3. Compliance Intelligence connects the proposal to source evidence, business context, applicability, system impact and an explicit review state. [S26]

4. Human experts approve interpretations where judgment is required. Deterministic validators check fields, thresholds, schemas, rates, sequence rules, dates and other testable constraints.

5. The Compliance Compiler or an equivalent rule-management process can turn approved knowledge into implementation requirements, policy packs, configurations and tests. [S27]

6. Production systems — POS, ERP, e-commerce, e-invoicing, fiscal middleware and agents — consume only approved rules and interfaces appropriate to their task.

7. Every compliance-critical action records evidence: source/rule version, input facts, decision, execution result, exceptions and approval history.

8. If authoritative evidence is missing, contradictory or stale, the system abstains, falls back or escalates rather than inventing an answer.

The design goal Do not ask: “Can the AI answer this tax question?” Ask: “What is the maximum autonomy this use case can safely have, and what must be true before its output is allowed to affect a legal result?”

Impact on retail and POS systems

Retail is one of the hardest environments for AI tax compliance because the system must make decisions at scale, under latency constraints, across stores, e-commerce, self-checkout and returns, often while country-specific fiscalization or e-invoicing obligations run in parallel. A model that is occasionally wrong in a research workflow is inconvenient; the same error repeated across millions of live transactions can become material.

Where AI fits well in retail

• Before checkout: product categorization, master-data enrichment, jurisdiction-quality checks, tax-code mapping and configuration recommendations.

• Around checkout: anomaly monitoring, diagnostic assistance, support for configuration selection and detection of unexpected transaction patterns.

• After checkout: reconciliation between POS, ERP, fiscal records, invoices, payments and filings; notice analysis; audit evidence and incident investigation.

• During development: API support, test generation, technical-specification comparison, mapping assistance and regression analysis.

• Across countries: regulatory-change comparison, multilingual retrieval, impact analysis and reusable compliance knowledge.

Where AI should be constrained

A general-purpose LLM should not be the sole inline authority deciding the legal tax result of a live retail transaction. At checkout, latency, availability, reproducibility and auditability matter alongside legal correctness. A safer pattern is to use AI upstream to classify products or propose configuration, validate and approve the result, and then let a deterministic tax or fiscal engine execute the approved logic at runtime.

The same principle applies to offline operation. If the store loses access to an AI service, the retailer still needs a legally defined behavior. Offline and degraded-mode rules must therefore be explicit. An AI assistant may help diagnose the situation, but the fallback path should not depend on the model improvising a legal answer.

This distinction is visible in current market products. Vertex uses generative AI to propose product tax categories and then exposes confidence, review and audit history. fiskaly’s public stack uses AI for documentation assistance while fiscal transaction signing and reporting remain explicit API workflows. These are different products, but both illustrate a broader pattern: AI can improve the work around tax rules without replacing the deterministic core. [S16] [S18]

Example: a new retail product appears in 12 countries

This example is hypothetical. A retailer adds a new plant-based beverage to its global product catalog. Product descriptions differ by market, local tax categories differ, and the product will be sold through stores and e-commerce.

1. AI reads the global product description, ingredients, packaging and existing category mappings and proposes likely country-specific tax categories.

2. The system returns a confidence score and the source or tax-content basis for each proposal. Low-confidence or high-impact countries go to an expert queue.

3. Approved mappings are written into a governed product-tax master with an effective date and change history.

4. The POS and e-commerce engines use the approved mapping deterministically at transaction time. The LLM is not called for every sale.

5. AI monitors transaction data and later flags stores where the observed tax treatment differs from the approved product rule.

6. A change in official guidance triggers impact analysis against affected products, countries, tests and configurations before any new rule is promoted.

The value of AI is real: less manual research, faster catalog onboarding and better exception targeting. The legal result remains controlled because the live transaction consumes approved tax data rather than an unconstrained model response.

AI compared with adjacent tax-compliance technologies

Dimension	AI / ML / LLM	Deterministic tax or rules engine	RPA / workflow	Compliance Intelligence
Primary purpose	Understand, classify, predict, recommend or orchestrate.	Execute explicit tax logic consistently.	Automate repetitive steps between systems.	Turn authority into validated, contextual, implementation-ready knowledge.
Typical input	Text, documents, transactions, product attributes, history.	Structured transaction facts and configured rules.	Structured files, events and workflow states.	Authoritative sources + structured business context.
Output	Probabilistic answer, score, classification or action proposal.	Repeatable calculation/validation result.	Process action or data movement.	Approved obligation, applicability, requirement, test and evidence link.
Main strength	Handles ambiguity, scale and unstructured information.	Predictability and repeatability.	Efficiency and integration.	Traceability from rule to system action.
Main weakness	Can hallucinate, drift or misclassify.	Only as correct as the modelled rules and configuration.	Can automate bad inputs/processes.	Requires governance and expert maintenance.
Best role in retail	Research, classification, anomaly detection, diagnostics, orchestration.	Runtime tax/fiscal/e-invoice validations.	Filing and data-transfer workflows.	Constraint and knowledge layer for people, software and AI.

International and jurisdictional variation

AI is global, but tax compliance is not. The same model can support research across countries, yet the allowed automation, source hierarchy, reporting timing, data requirements and transaction controls remain jurisdiction-specific. A global retailer therefore needs a common AI operating model with country-specific compliance profiles rather than one global prompt.

In U.S. sales and use tax, product taxability and jurisdiction complexity make product categorization a natural AI use case; Vertex’s Smart Categorization is explicitly built for U.S. retail product mapping. [S16] In VAT and e-invoicing environments, structured invoice data and continuous reporting create strong use cases for field mapping, validation, reconciliation and error remediation; Sovos and Avalara both market AI in these workflows. [S11] [S13] [S14]

In fiscalized POS environments, legal requirements may involve cryptographic signing, transaction sequence controls, certified components, authority communication or immutable audit records. These are typically poor candidates for free-form generative decision-making at runtime. AI can support development, configuration, support, testing and anomaly detection around the fiscal core while the legal transaction mechanism remains explicit. fiskaly and efsta public materials illustrate the continuing importance of deterministic fiscalization services. [S18] [S19]

AI governance law can also vary independently from tax law. In the European Union, Regulation (EU) 2024/1689 applies a risk-based framework to AI, with human-oversight and logging requirements for systems that fall within its high-risk categories and transparency obligations for certain AI interactions and generated content. A tax-compliance AI tool is not automatically “high-risk” merely because it is used for tax; classification depends on intended purpose and the legal criteria. The AI Act is therefore a governance consideration, not a blanket tax-software classification. [S9] [S10]

Common misconceptions about AI in tax compliance

“If the answer cites a source, it is safe to use.” A citation can be irrelevant, outdated or misapplied. Retrieval proves that text was found, not that the rule applies to the transaction. Peer-reviewed legal-AI research found material hallucination even in retrieval-based systems. [S7]

“A better prompt can make the model 100% accurate.” Prompts improve behavior but do not transform a probabilistic language model into an authoritative tax rule engine. High-confidence language is not a legal control.

“AI should calculate tax because it is smarter than rules.” Tax calculation needs repeatability, traceability and stable legal logic. AI is often more valuable upstream in classification, setup and exception handling.

“Automation and AI are the same thing.” A deterministic API workflow or rules engine may be highly automated without using AI. For compliance-critical work, that is often an advantage.

“Human in the loop means the risk is solved.” A human approval step is useful only if the reviewer sees the source, context, version, confidence, impact and evidence needed to make a real decision.

“One global AI can normalize country differences.” AI can help compare and map concepts, but legal differences cannot be averaged away. Country-specific applicability and effective dates must remain explicit.

“Agentic AI means removing humans from compliance.” Agentic AI is more useful when routine actions are automated and humans retain authority over ambiguous, high-impact or exception cases.

The hard limits: why tax compliance cannot simply trust an LLM

The central limitation is structural. A large language model is optimized to generate likely language given context; tax compliance often requires a result that must remain correct under a defined legal source, date, transaction and jurisdiction. Those goals overlap, but they are not identical.

Risk	Failure mode	Control
Hallucination / confabulation	A fluent answer invents or misstates a rule.	Source grounding, verification, abstention, test sets. [S6] [S7]
Applicability error	Correct rule applied to wrong entity, transaction, date or country.	Structured context and explicit applicability model.
Temporal error	Current rule used historically or obsolete rule used today.	Effective-date versioning and historical snapshots.
Source hierarchy error	Blog/guidance/internal memo treated like binding law.	Source type, legal status and approval metadata.
Data-quality error	Wrong product, address, tax ID or transaction data produces a wrong result.	Master-data controls, validation and anomaly monitoring.
Model drift	Model update changes behavior without tax rule change.	Regression testing, model/version controls, stable deterministic boundaries.
Automation bias	User accepts confident recommendation without review.	Meaningful approval UX, confidence/evidence, training and escalation.
Permission leakage	Agent accesses or acts on data outside its authorized scope.	Least privilege, tenant separation, tool permissions, logs and approval gates.
Availability / latency	AI service is slow or unavailable during checkout.	Keep compliance-critical runtime logic deterministic, cached or locally resilient.

Can AI reach 100% accuracy in tax compliance?

Not as a general claim for open-ended tax reasoning. Tax law contains ambiguity, exceptions, changing guidance, business facts and jurisdiction-specific interpretation. Current LLMs also remain capable of unsupported generation. It would therefore be misleading to promise that a general AI system can answer every tax-compliance question with 100% accuracy. [S6] [S7] [S8]

A more realistic engineering target is narrower and stronger: for a bounded set of approved, testable tax rules, the production system should achieve deterministic conformance or refuse to execute. The AI can help discover the rule, draft the model, classify cases and generate tests, but the final production behavior should be validated against an authoritative rule package.

A better research target than “100% accurate AI” 100% of compliance-critical production actions should be traceable to approved knowledge and pass required deterministic controls — or the system should abstain and escalate. This does not make legal interpretation infallible, but it prevents an unvalidated model guess from becoming a legal system action.

Darko Pavic’s perspective

In my view, tax compliance is one of the clearest examples of both the power and the limit of generative AI. We should use AI much more than we do today. It can remove enormous amounts of manual searching, document comparison, mapping, testing and repetitive support work. But we should not confuse speed with authority.

Based on my work in international retail technology and fiscalization, the practical problem appears when a regulation leaves the tax department and enters a real system. A POS application, e-commerce platform, ERP, e-invoicing service or fiscal middleware cannot “probably” comply. At some point the legal requirement must become a controlled technical behavior.

I therefore distinguish between AI assistance and compliance authority. AI should be able to propose a product category, identify a regulatory change, explain an error, draft an implementation requirement or even orchestrate an approved workflow. The authority to govern a tax-critical action should come from validated compliance knowledge, not from the confidence of the language model.

This is the reason I see Compliance Intelligence as the missing layer between regulatory information and autonomous software. Compliance Intelligence connects sources, interpretation, applicability, implementation, tests and evidence. The Compliance Compiler is a related research direction: how to transform validated regulatory knowledge into machine-readable rules, implementation artefacts and verification without hiding ambiguity or professional responsibility. [S26] [S27]

The longer-term shift may be from prompt engineering to rule engineering. If AI agents increasingly decide how business tasks are executed, the valuable asset will not be the cleverest prompt. It will be the validated set of obligations, prohibitions, permissions, context and tests that defines what the agent is allowed to do.

Research agenda: making AI safer for tax compliance

Research stream	Core question
Source-grounded retrieval	How can systems guarantee that answers use the correct jurisdiction, source hierarchy and effective date?
Machine-readable regulation	Which tax and fiscal rules can be represented as authoritative machine-consumable logic, and how should ambiguity remain visible? [S4] [S5]
Compliance ontologies	How should countries, actors, transactions, obligations, exceptions, systems and evidence be represented without flattening legal differences?
Validated rule compilation	How can an approved interpretation become executable rules, configurations, contracts and test cases without semantic drift?
Formal verification and testing	Which properties of tax calculations, fiscal workflows and e-invoice validations can be proven or exhaustively tested?
Controlled abstention	How should AI recognize insufficient evidence, uncertainty or conflict and escalate instead of answering?
Temporal knowledge	How can current and historical rules coexist without applying the wrong version to a transaction?
Agent authorization	How should permissions, tool access, spending/action limits and legal constraints be attached to autonomous tax agents?
Evaluation benchmarks	What expert-approved benchmark sets can measure factual accuracy, applicability, citation fidelity, completeness and change sensitivity?
Evidence by design	How can every recommendation and action preserve source, model, rule, approval and execution history for audit?

The OECD’s current Law as Code consultation is particularly relevant because it frames authoritative machine-executable legal logic as shared digital infrastructure and explicitly connects the problem to AI. The consultation also preserves an essential boundary: authoritative legal text remains binding, and interpretive or discretionary elements should not be silently converted into deterministic code. [S4]

Implementation considerations for retailers and solution providers

The checklist below is an implementation aid, not legal advice. The right control level depends on the tax domain, country, transaction, data, system and consequence of error.

Use-case and risk

• What decision or action will AI influence?

• What is the consequence if the AI is wrong once, and what happens if the same error is repeated across millions of transactions?

• Can the use case remain advisory, or does it affect calculation, reporting, signing, invoicing or filing?

• What maximum autonomy level from the Four-Level AI Tax Compliance Control Model is acceptable?

Sources and knowledge

• Which sources are authoritative, and how are guidance, industry practice and internal interpretation distinguished?

• Are jurisdiction, publication date, effective date, version and language stored?

• Can every important conclusion be traced to a source passage?

• What happens when sources conflict or no sufficient evidence is found?

Data and classification

• Are product, customer, legal-entity, address and registration data reliable enough for AI use?

• Are confidence thresholds defined for classifications?

• Who reviews low-confidence or high-impact cases?

• How are approved mappings versioned and reused across channels?

Architecture and integration

• Is the AI in the checkout-critical path, and if so, what are the latency and availability guarantees?

• Can production execution fall back to deterministic behavior when AI is unavailable?

• Which systems consume AI output: POS, ERP, e-commerce, tax engine, e-invoicing, fiscal middleware, filing or agents?

• Are AI recommendations separated from approved production configuration?

Validation and testing

• What expert-approved benchmark questions and transaction scenarios exist?

• Are positive, negative, boundary, exception and historical-version cases tested?

• Are model changes regression-tested separately from tax-rule changes?

• Can the system prove that a production action passed deterministic checks?

Governance and security

• Who is accountable for the output?

• Which models and tools are approved, and how is shadow AI controlled?

• Are customer and tenant data separated and excluded from unintended model training?

• Are permissions least-privilege, logged and reviewable?

Operations and evidence

• How are AI recommendations, human approvals, rule versions and system actions logged?

• How are exceptions and abstentions monitored?

• Who owns regulatory change management and model lifecycle management?

• Can an auditor reconstruct why the system behaved as it did?

Frequently asked questions

What is the best use of AI in tax compliance?

The best early use cases are high-volume tasks where AI can save time without becoming the final legal authority: research, document review, product classification, data mapping, anomaly detection, reconciliation, support and test generation.

Can AI calculate VAT or sales tax?

AI can assist the configuration and classification work that feeds a tax engine, but live tax calculation should generally remain deterministic and based on validated tax content. An AI agent may call a tax-calculation service without being the source of the tax rule.

Can an AI chatbot replace a tax expert?

No. A chatbot can make tax knowledge easier to search and explain, but it does not remove the need to validate source authority, applicability, effective dates, business facts and high-impact interpretations.

Does retrieval-augmented generation eliminate hallucinations?

No. Retrieval improves grounding, but a 2025 peer-reviewed study of leading legal RAG tools still found hallucinations in 17% to 33% of tested queries. The study is not a tax benchmark, but it shows why retrieved citations are not a guarantee of correctness. [S7]

How can AI help POS software providers?

AI can accelerate country research, developer support, product and transaction mapping, test generation, log analysis and compliance-change impact analysis. It should not silently replace certified or deterministic fiscal controls.

How can AI detect tax compliance problems?

Machine-learning and analytics systems can identify anomalies, unusual patterns, mismatches between transaction and filing data, missing records and clusters of behavior associated with known problems. Tax administrations already use AI widely for risk assessment and fraud detection. [S1] [S3]

Can agentic AI submit returns or e-invoices?

Technically yes, if the agent is given tool access. The safer architecture is to let the agent orchestrate an approved workflow while deterministic validations, permissions, approval gates and evidence govern the legally relevant action.

What would “100% accuracy” mean in tax AI?

It should not mean that an LLM is infallible. A defensible target is that every compliance-critical production action is supported by approved knowledge and passes required deterministic checks, or is blocked and escalated.

Source list

Official and international-organization sources

[S1] OECD. Tax Administration 2025: Comparative Information on OECD and other Advanced and Emerging Economies — 17 November 2025; DOI 10.1787/cc015ce8-en Direct source

Why used: Primary comparative source for tax-administration AI adoption, risk assessment, fraud detection and service use.

[S2] OECD. Tax Administration Digitalisation and Digital Transformation Initiatives — 17 June 2025; DOI 10.1787/c076d776-en Direct source

Why used: Primary OECD source on AI, data, APIs, prefilling and digital transformation in tax administration.

[S3] OECD. Governing with Artificial Intelligence: AI in tax administration — 18 September 2025; DOI 10.1787/795de142-en Direct source

Why used: Detailed OECD analysis of AI for fraud detection, risk assessment, taxpayer services, data quality and trustworthy governance.

[S4] OECD. Consultation on the digital provision of law: Towards a shared reference framework for Law as Code — Consultation opened 29 July 2026 Direct source

Why used: Current authoritative policy work on state-authorised machine-executable law, provenance, interpretation boundaries and AI.

[S5] Mohun, James & Alex Roberts. Cracking the code: Rulemaking for humans and machines — 12 October 2020; DOI 10.1787/3afe6ba5-en Direct source

Why used: Foundational OECD Rules as Code work on machine-consumable official rules and their limitations.

[S9] European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text — Consolidated 27 July 2026 Direct source

Why used: Primary legal source for risk-based AI governance, human oversight and logging requirements where applicable.

[S10] European Commission. Guidelines on transparency obligations for providers and deployers of AI systems — 20 July 2026 Direct source

Why used: Current EU guidance on transparency obligations applicable from 2 August 2026.

Standards, AI-risk and academic sources

[S6] NIST. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — 26 July 2024; updated 8 April 2026; DOI 10.6028/NIST.AI.600-1 Direct source

Why used: Authoritative cross-sector guidance on generative-AI risks, including confabulation and governance.

[S7] Magesh, Varun et al. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools — Journal of Empirical Legal Studies, 2025, 22:216–242; DOI 10.1111/jels.12413 Direct source

Why used: Peer-reviewed evaluation showing that retrieval-based legal AI reduced but did not eliminate hallucinations; used only as evidence about high-stakes legal reasoning, not as a direct tax benchmark.

[S8] Dahl, Matthew et al. Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models — Journal of Legal Analysis, 2024, 16(1):64–93; DOI 10.1093/jla/laae003 Direct source

Why used: Peer-reviewed evidence on hallucinations and uncertainty in general-purpose LLMs on legal tasks.

Industry implementation sources

[S11] Avalara. Agentic Tax and Compliance — Current product page reviewed 23 August 2026 Direct source

Why used: Vendor source for publicly described agentic tax use cases across calculations, filings, classification, research, notices, e-invoicing, onboarding and enterprise systems.

[S12] Avalara. Inside Avalara MCP servers: A developer’s primer for agentic tax and compliance — 27 May 2026 Direct source

Why used: Vendor technical source for MCP access enabling enterprise AI agents to interact with tax and compliance services.

[S13] Sovos. Sovos Expands Sovi AI with New Capabilities for Tax Guidance, Intelligent Automation and Agent Interoperability — 24 March 2026 Direct source

Why used: Vendor source for Ask Sovi, diagnosis/remediation, recommendations, proactive issue detection, governance and MCP/A2A direction.

[S14] Sovos. The Expansion of Sovos Intelligence Closes the Global Compliance Reconciliation Gap — 7 July 2026 Direct source

Why used: Vendor source for reconciliation across indirect-tax data, natural-language analytics and audit evidence.

[S15] Sovos. Sovi AI for Product Tax Code Classification and Data Mapping — 11 November 2025 Direct source

Why used: Vendor source for AI-assisted product tax classification and VAT data mapping.

[S16] Vertex. Vertex Smart Categorization — Current product page reviewed 23 August 2026 Direct source

Why used: Vendor source for generative-AI retail tax categorization, confidence scores, human review, mapping history and audit packages.

[S17] Vertex. 2026 Growth Stock Conference / AI Product Strategy presentation — June 2026 Direct source

Why used: Investor source used only for Vertex’s publicly stated broader AI roadmap, including certificate management, assistants/agents and human-in-the-loop categorization.

[S18] fiskaly Workspace — Current developer site reviewed 23 August 2026 Direct source

Why used: The current fiskaly Workspace genuinely supports the statement that they have an AI assistant trained on their documentation. And there is a public evidence of its AI-first work and hackathons.

[S19] efsta AI related cooperation with university  — Linkedin post 23 August 2026 Direct source

Why used: Vendor source for efsta’s public focus on transaction recording, manipulation protection, archiving and country-specific fiscalization; no AI-specific public capability was inferred beyond what was verifiable. There is actually an excellent official efsta source specifically about Olaf, so use that rather than relying primarily on LinkedIn.

[S20] Thomson Reuters Institute. 2026 AI in Professional Services Report — 2026 Direct source

Why used: Industry research source on professional AI adoption and tax/accounting use cases such as research, document summarization and review.

Darko Pavic’s related work and author sources

[S22] Forbes Councils. Darko Pavic — Founder / CEO, Fiscal Solutions — Profile reviewed August 2026 Direct source

Why used: Public biography supporting the author’s relevant retail technology and fiscalization background.

[S23] The Fiscalization Compliance Maturity Model — Website reviewed August 2026 Direct source

Why used: Author’s published fiscalization framework and book; used for related work and authority context.

[S24] Darko Pavic. From Shoppers To Systems: Inside The Agentic Commerce Revolution — Forbes Technology Council, 24 July 2025 Direct source

Why used: Related author analysis on autonomous commerce agents and retail system architecture.

[S25] Fiscal Solutions. Fiscal Compliance in the Age of Agentic Commerce — Current white-paper page reviewed August 2026 Direct source

Why used: Related industry white paper on AI-driven commerce, fiscalization, e-invoicing and auditability.

[S26] Darko Pavic. What Is Compliance Intelligence? From Regulatory Knowledge to Trusted System Action — Author-provided authority page, 23 August 2026 Direct source

Why used: Author framework used for the distinction between AI assistance and validated compliance knowledge.

[S27] Darko Pavic. Compliance Compiler and Machine-Readable Law — Current authority page Direct source

Why used: Author framework used for the research direction from regulation to structured obligations, implementation rules, tests and evidence.

---

How this page was prepared

This page combines OECD publications on tax administration and machine-readable rules, EU and NIST AI-governance material, peer-reviewed legal-AI research, public product documentation and announcements from tax-technology providers, and Darko Pavic’s published and author-provided work on retail compliance, Compliance Intelligence and the Compliance Compiler. Official evidence, vendor claims and professional interpretation are deliberately distinguished.

Generative AI materially supported source discovery, comparison, structuring and drafting. Important external claims were checked against the cited sources available on 23 August 2026. The named author reviewed the final wording, professional interpretations and links before publication and remains responsible for the published content.

First published: 23 August 2026 | Last substantively reviewed: 23 August 2026 | Author: Darko Pavic

Author bio

Darko Pavic is the founder and CEO of Fiscal Solutions and has more than 28 years of experience in international retail technology and fiscalization. His work focuses on POS architecture, global fiscal compliance, e-invoicing and the systems that translate regulatory requirements into scalable retail operations. He is a member of the Forbes Technology Council and the author of The Fiscalization Compliance Maturity Model. His current research interests include Compliance Intelligence, machine-readable regulation, AI governance and the use of validated compliance knowledge to constrain autonomous software. [S22] [S23]

---

Suggested citation

Pavic, Darko. How AI Can Be Used in Tax Compliance: Where It Works, Where It Fails, and How to Control It. 23 August 2026.

---

Follow the development of AI in tax compliance Subscribe to Future of the High Street for continuing analysis of fiscalization, e-invoicing, retail technology, Compliance Intelligence and the role of AI in compliance-critical systems.

Subscribe to Future of the High Street