How AI Tax Research Handles Multistate Nexus Questions

AI tax research handles multistate nexus by checking sales thresholds and linking what a business does to the authority that applies in each jurisdiction. For citation-backed federal and state research, Marble grounds answers in primary authority with every claim cited, while Thomson Reuters CoCounsel Tax uses defined, validated research content and expert-built templates. These systems accelerate synthesis across layered rules, but practitioners supply the verified facts, legal judgment, and responsibility that make a nexus conclusion defensible.

Why multistate nexus analysis has outgrown revenue threshold lookups

A business can change its state tax exposure through an operational decision that never reaches the tax team. A remote hire, leased asset, installation visit, or bundled service offering can create a filing obligation without a conspicuous change in sales. Nexus analysis in 2026 therefore requires detecting activity across the business. The rules for property, payroll, temporary presence, and digital activity differ from one jurisdiction to the next. A threshold monitor captures only part of that footprint.

Operations or sales teams may arrange installations, training, repairs, or consulting visits without knowing what they could mean for taxes. Yet these short-term activities can create physical nexus. For leased assets, many states tax the lease payments as they recur. Research must therefore establish both what the business does and how each state treats that activity.

Meanwhile, states increasingly use third-party data, interagency coordination, and automated matching to identify filing obligations. An overlooked trigger may surface only when an audit notice arrives, with back taxes, penalties, and remediation already in play. Each new state has its own tax bases, filing thresholds, and apportionment mechanics, so it becomes harder to keep positions consistent across states. Teams that handle engagements in several states need a way to keep track of all these facts and rules and coordinate their work across jurisdictions. Manual triage becomes untenable when the task is to connect changing operational facts with every relevant state rule.

What AI actually does when it encounters a layered nexus question

Two categories of tools address different parts of this problem. Purpose-built compliance automation, including Avalara, Anrok, Vertex, and Sovos, handles tax calculation, nexus threshold monitoring, filing, and exemption certificate management across jurisdictions. Machine learning adds a predictive layer that anticipates when a business will cross thresholds in new states. Natural language processing models scan state and local tax bulletins to find changes in rates, nexus thresholds, and incentives before they take effect. These functions support monitoring and execution rather than resolving every legal implication of a fact pattern.

Professional tax research platforms address whether a particular combination of activities creates nexus. Retrieval-augmented generation, or RAG, anchors the answer to a defined body of verified primary authority. This architecture gives the practitioner a basis for tracing an answer to the governing statutes and guidance. The distinction matters when a single remote employee or leased asset can change an engagement’s posture. The value is not simply an answer produced quickly, but an answer whose legal basis can be examined.

Thomson Reuters CoCounsel Tax uses expert-built templates for complex corporate income tax nexus questions. Its 86-272 Template incorporates detailed state-by-state analysis of the internet activities targeted by the Multistate Tax Commission, mapped by the Checkpoint Catalyst team. Thomson Reuters also uses prompting guardrails that restrict responses to a defined set of validated information from products such as Westlaw, Practical Law, and Checkpoint. In August 2026, Thomson Reuters disclosed that its proprietary model “Thomson” starts from a strong open-source foundation and is trained on decades of authoritative content. These details distinguish the source collection and research framework from the generated response itself.

Blue J’s Tax Foresight generates outcome predictions with confidence percentages for specific fact patterns, showing how courts or the IRS are likely to rule. Adjusting factors in real time reveals which variables drive the assessment. Marble operates in the practitioner research category, combining primary-authority research with intake, document review, and compliance checks. Its research can remain within a client or engagement, connecting the question to the relevant facts and subsequent work.

At the apportionment stage, AI also handles volume work by ingesting sales data, applying nexus rules across jurisdictions, flagging sourcing or allocation anomalies, and generating schedules at scale. That work differs from determining which legal interpretation a practitioner can support.

How the P.L. 86-272 fracture created a new class of nexus questions AI must navigate

The treatment of internet activities under P.L. 86-272 shows why nexus research cannot rely on a stable table of rules. Since the Multistate Tax Commission’s 2021 revised guidance, more states have treated common website activities as grounds for income tax nexus outside the law’s protection. Activities that businesses regarded as ordinary digital operations now require state-specific analysis. Interpreting the activity requires identifying how a state has adopted the position, when it became effective, and what courts have permitted.

New Jersey formalized its position through regulations effective June 16, 2025. Massachusetts amended its corporate excise tax regulations effective October 10, 2025. The Massachusetts regulation specifies that cookies placed on in-state users’ devices can constitute activity beyond mere solicitation when the collected data serves purposes such as adjusting production schedules and inventory amounts, developing new products, or identifying new items to offer for sale. The relevant distinction is thus the use of the data. A research answer that omits that purpose can miss the operative fact.

Adoption has not followed a uniform path. New Jersey, New York, and Massachusetts have formally adopted the MTC’s position through regulations. California attempted adoption through guidance that was voided on procedural grounds, although auditors continue to apply its principles informally. Many states say nothing about the MTC stance until they bring it up in an audit.

The courts have added further distinctions. On April 28, 2025, the New York Supreme Court in Albany County upheld the constitutionality of internet activities provisions under P.L. 86-272, finding no federal preemption, but invalidated retroactive application before promulgation. On October 1, 2026, the Tax Court of New Jersey ruled against the American Catalog Mailers Association in its challenge to the Division of Taxation’s 2025 regulations. But the New Jersey court chose not to decide whether AI-powered chatbots are interactive customer support or just automated decision trees. That leaves a classification question without an authoritative resolution. The irony is precise because AI is being asked to research a legal boundary that may turn on the behavior of AI itself.

Federal legislation did not close the gap. Congress introduced the Interstate Commerce Simplification Act of 2025 (H.R. 427) in January 2025 to expand “solicitation” to protect digital activities that facilitate solicitation. The language appeared in the House-passed reconciliation bill (OBBBA), but the Senate removed it and the final enacted law excluded it. AI research must navigate the resulting fracture rather than present the MTC’s position as a uniformly settled national rule.

Why AI also faces its own multistate nexus exposure as a taxable product

AI adds a separate state tax question when it is the product being sold. Many states classify AI solutions under existing SaaS or data-processing categories, while some are moving in different directions. This taxability question is distinct from whether an AI-enabled website activity exceeds P.L. 86-272 protection. A practitioner who advises a seller of AI services may still need to answer both questions. The product classification and the seller’s activities belong to different layers of the same engagement.

The Illinois Department of Revenue clarified in a September 2025 ruling that online AI tools and services accessed through a browser or app without a software download do not incur Illinois state sales tax. Illinois treats these services like SaaS, generally as services rather than tangible personal property, although the City of Chicago imposes its separate lease tax. Indiana’s Department of Revenue stated in July 2025 that generative AI services accessed through web browsers or APIs are services not subject to sales/use tax. A general label such as “AI service” does not establish how a state taxes the service. Research must use current, jurisdiction-specific classifications.

Where AI research breaks down on the questions that matter most

The strongest objection to AI nexus research concerns its failure modes. State and multistate coverage can be thin and inconsistent, precisely where synthesis across jurisdictions appears most useful. An answer can read well even when the authority it draws on is incomplete. That is a structural weakness because the answer depends on what the system can retrieve and how that material applies to the question. Better prompting alone cannot add missing law or bring an old set of sources up to date.

The most consequential errors need not look like fabricated cases or invented statutes. They can be a threshold from a prior tax year, a superseded ruling described as active, a code section cited slightly out of context, or a sunset date that has passed. Each error can look right in an answer that otherwise makes sense. A citation establishes where a proposition came from, but not necessarily that it controls the stated facts and period. Citation-backed research is therefore necessary for traceability without being sufficient for correctness.

The P.L. 86-272 developments make this risk concrete. A research system drawing on material from only months before the Massachusetts or New Jersey regulations took effect could confidently describe an outdated compliance position. RAG anchored to verified primary authority significantly raises the floor by narrowing the source base. It does not eliminate an outdated document within that base or a jurisdiction the platform has not mapped. Nor can it retrieve an authoritative answer to the unresolved New Jersey chatbot question when none exists.

These limitations are most dangerous when the generated answer compresses differences in effective dates, legal status, and unresolved interpretation into a single conclusion. A practitioner needs those distinctions because they determine what the answer can support. The professional standard is consequently broader than producing a plausible result with citations attached. Defending how the facts, sources, and interpretation led to the result is part of the job. The methodology must withstand scrutiny as well as the conclusion.

What practitioner judgment must supply that no AI platform replaces

A defensible nexus position is a professional act. It requires an accurate account of the client’s activity, a reasoned application of law, and responsibility for the conclusion. AI contributes to that work by organizing and synthesizing information, but it does not establish the truth of the operational facts supplied to it. The practitioner must confirm those facts before treating them as the basis of a determination. That role makes the practitioner the source of the conclusion.

Judgment becomes decisive where authority stops. An unresolved chatbot classification or a state that remains silent until audit does not become settled because an AI response sounds confident. The practitioner must reach a reasoned position within that uncertainty and document its basis. A professional also signs off on the nexus determination, while an AI platform does not assume liability. These responsibilities remain distinct from the quality of the research assistance.

Audit readiness requires that the reasoning remain accessible after the immediate question is answered. Nexus determinations, sourcing logic, exemption support, and lease treatment need documentation that explains how and why the decision was made. When that information is fragmented or difficult to retrieve, the audit becomes more disruptive, expensive, and difficult to manage. New remote hires, out-of-state sales, and expansion make this harder because decisions made outside the tax team often create duties to file or pay tax.

That division of labor gives purpose-built practitioner tools, including Marble, a concrete role in clearing routine intake, document review, and compliance checks from the analytical workload. Automating that work frees time for the jurisdiction-specific decisions that cannot be delegated to retrieval or generation. This serves to direct senior judgment toward the facts and legal uncertainties that determine the client’s position.