TL;DR
Indirect tax automation in 2026 is two different problems wearing one name. The first problem, tax determination, means calculating the right rate for a given transaction in a given jurisdiction. This is what tax engines like Avalara, Vertex, and Sovos do, and they do it well. The second problem, tax operations, is the judgment-heavy work that wraps the calculation: monitoring economic nexus thresholds across roughly 12,000 US jurisdictions, managing exemption and resale certificates, reconciling return data across multiple ERPs and channels, responding to jurisdiction notices, and maintaining an audit trail that can reconstruct any decision years later. The second problem is where tax teams actually spend their time, and it is largely unautomated.
Agentic AI fits the second problem, not the first. It does not replace the tax engine; it sits around it, handling the reasoning and exception work the calculation engine was never designed to do. A deterministic, agentic platform can read an exemption certificate, decide whether it is valid and applies to a given sale, flag the ones that are expired or mismatched, and explain its reasoning in plain language an auditor can read. It can monitor sales against nexus thresholds and reason about when registration is triggered. It can reconcile the data feeding a return and surface the line items that do not tie out. Crucially, it can do all of this with a reconstructable audit trail, which matters because indirect tax is one of the most audit-exposed functions in the enterprise.
The reason this distinction matters for buyers: a tax engine and an agentic operations layer are complementary, not competing, purchases. The mistake is expecting the calculation engine to solve the operations problem, then concluding “the software does not work” when the exemption certificates are still a mess and the auditor still has questions. This post explains why indirect tax is harder than it looks, the four places the judgment work concentrates, and how agentic AI handles each — with the honest boundaries of where it fits and where the tax engine remains essential.
For the broader pattern of where agentic AI fits in finance operations, see How to Automate Data Extraction with Agentic AI and The 7 Places Generative AI Quietly Fails in Accounts Payable.
Why indirect tax is harder than it looks
The calculation is the visible part of indirect tax, and the part vendors demo. It is also the part that has been effectively solved. A tax engine looks up the jurisdiction, applies the rate, and returns a number in milliseconds. If indirect tax were only calculation, it would be a solved problem and tax teams would be small.
They are not small, because four things make indirect tax genuinely hard, and none of them are the calculation.
The jurisdictional surface is enormous and constantly moving. The US alone has roughly 12,000 tax jurisdictions, and their rates and rules change continuously. Static ERP tax tables go stale almost immediately, which is why manual maintenance is a documented audit risk. But keeping rates current is still the tractable part; the harder part is reasoning about which jurisdiction’s rules even apply to a transaction that touches several.
Nexus is a judgment call, not a lookup. Economic nexus rules mean a business can owe tax in a state where it has no physical presence, typically once sales cross a threshold like $100,000 or 200 transactions. Marketplace facilitator rules complicate this further: marketplace sales may handle remittance but still count toward the threshold that triggers registration obligations elsewhere. Deciding when an obligation is triggered, in which jurisdictions, and what to do about it is ongoing reasoning, not a one-time setup.
Exemptions require document judgment. B2B sellers must collect, validate, and apply resale and exemption certificates. A certificate can be expired, issued for the wrong jurisdiction, mismatched to the product category, or simply missing. Collecting tax on a genuinely exempt sale is an error; failing to collect on a non-exempt sale is a liability. Getting this right means reading documents and exercising judgment about whether each one is valid and applies, at the volume of every B2B transaction.
The audit exposure is severe and retrospective. Indirect tax is among the most frequently audited finance functions, and audits look backward by years. When a jurisdiction questions a return, the team must reconstruct why each treatment was applied, which certificate justified each exemption, and how the numbers were derived. If that reasoning lives only in a calculation engine’s outputs and a spreadsheet, the reconstruction is painful and the audit position is weak.
None of these four is a calculation problem. All four are reasoning, document, and reconstruction problems. That is precisely the shape of work agentic AI is suited to, and precisely the shape of work a rate-calculation engine was not built for.
The two layers of the indirect tax stack
The clearest way to think about indirect tax software in 2026 is as two distinct layers.
The determination layer calculates the correct tax for a transaction. This is the tax engine: Avalara (AvaTax), Vertex, Sovos, and newer AI-native entrants like Kintsugi and Anrok for specific segments. These platforms maintain rate databases across 190+ countries, integrate deeply with ERPs and ecommerce systems, and return accurate calculations at transaction speed. For multinational VAT and GST, continuous transaction controls, and e-invoicing mandates, these engines are essential and not something to replace. This layer is mature and well served.
The operations layer is everything required to be compliant around the calculation: nexus monitoring and registration decisions, exemption and resale certificate management, return preparation and the reconciliation behind it, jurisdiction notice handling, and audit defense. This layer is far less automated. It is where tax teams spend their time, where errors become liabilities, and where the audit exposure concentrates. Much of it is still done in spreadsheets, shared inboxes, and certificate folders, stitched together by human judgment.
Agentic AI is an operations-layer technology. The confusion in the market, and the reason some tax automation projects disappoint, is the expectation that buying a determination-layer engine will solve operations-layer problems. It will not, because they are different problems. The engine calculates correctly and the exemption certificates are still expired, the nexus thresholds are still tracked in a spreadsheet, and the auditor still asks questions the engine’s outputs cannot answer.
Where agentic AI fits: four operations-layer jobs
Agentic AI earns its place in the indirect tax stack by handling the four judgment-heavy jobs the determination engine does not. In each, the distinguishing requirement is the same: read or reason about something ambiguous, decide, and explain the decision in a way that survives an audit.
1. Nexus monitoring and registration reasoning
The job: continuously watch sales activity against economic nexus thresholds across every relevant jurisdiction, account for how marketplace-facilitated sales count toward those thresholds, and reason about when a registration obligation is triggered and where.
Why it needs reasoning, not just calculation: thresholds differ by jurisdiction, the rules around what counts vary, and marketplace facilitator law adds a layer where some sales are remitted by the marketplace but still count toward your nexus elsewhere. This is an ongoing judgment about obligation, not a rate lookup.
How agentic AI handles it: the platform monitors the sales data, applies the nexus rules expressed in plain language, and flags approaching and crossed thresholds with the reasoning attached, so a tax professional sees not just “you crossed a threshold” but which sales drove it and which obligation it triggers. Because the logic is explicit and readable, the tax team can adjust it as rules change without waiting on a developer.
2. Exemption and resale certificate management
The job: collect exemption and resale certificates, validate that each is current and issued for the correct jurisdiction, match it to the right customer and product category, apply it to the right transactions, and flag the ones that are expired, mismatched, or missing before they become an audit finding.
Why it needs reasoning, not just calculation: a certificate is a document that must be read and judged. Validity is not a single field; it is a combination of expiration, jurisdiction, entity match, and product applicability. This is document understanding plus judgment, at transaction volume.
How agentic AI handles it: the platform reads each certificate, extracts the relevant attributes, decides whether it is valid and applies to a given sale, and explains that decision. Expired and mismatched certificates surface as exceptions with the reason stated plainly. This is the same document-plus-judgment pattern that agentic AI handles across finance operations, applied to the specific artifact that drives indirect tax exemptions.
3. Return preparation and reconciliation
The job: assemble the data that feeds each return, often from multiple ERPs and sales channels, reconcile it, and surface the line items that do not tie out before the return is filed.
Why it needs reasoning, not just calculation: the determination engine calculates tax per transaction, but the return aggregates across systems that do not always agree. The hard part is finding and explaining the discrepancies — the same reconciliation reasoning that makes month-end close difficult, applied to tax data.
How agentic AI handles it: the platform reconciles the feeds, identifies the mismatches, and explains each one in plain language so a human resolves it quickly rather than hunting for it. The reconciliation logic is auditable, so the path from source data to filed return is reconstructable later.
4. Notice handling and audit defense
The job: when a jurisdiction sends a notice or opens an audit, respond with the reasoning and evidence behind the returns in question, often for periods years in the past.
Why it needs reasoning, not just calculation: audit defense is reconstruction. It requires showing why each treatment was applied, which certificate justified each exemption, and how each number was derived. A calculation engine stores outputs; it does not necessarily store the reconstructable reasoning an auditor wants.
How agentic AI handles it: because a deterministic agentic platform logs every decision with its inputs, the specific rule applied, and the plain-language reasoning, the audit response becomes retrieval rather than reconstruction. This is the same audit-trail standard that applies across regulated finance work, and it is the single highest-value property agentic AI brings to indirect tax, because indirect tax is so audit-exposed. See the AI Audit Trail Requirements checklist for the field-level standard.
Why deterministic execution matters specifically for tax
Not all AI is suited to tax work, and the distinction is consequential. Tax is a domain where the same facts must produce the same treatment every time, where the reasoning must be inspectable, and where “the model was fairly confident” is not an acceptable answer to an auditor.
A probabilistic AI system that produces a plausible tax treatment most of the time, with occasional variation on identical inputs, is a liability in this domain. Tax authorities expect consistency and defensibility. A deterministic agentic platform, where the same transaction and the same rules always produce the same treatment and the specific rule applied is recorded in plain language, aligns with how tax actually has to work. See also When Confidence Scores Lie for why “94% confident” is not an audit trail in any regulated finance domain.
This is why the architecture matters more in tax than in many other functions. The properties that make a platform suitable here are deterministic execution (identical inputs yield identical, reproducible treatments), reasoning expressed in readable policy rather than buried in model weights (so a tax professional can verify and adjust it, and an auditor can read it), and an audit trail that reconstructs any decision end to end. Kognitos is built around these properties, which is why deterministic, English-as-code, audit-native agentic AI fits the indirect tax operations layer specifically. The point is not the brand; it is that the architecture has to match the domain’s demand for consistency and defensibility, and probabilistic systems do not.
A note on honest scope: Kognitos is not a tax determination engine and does not replace Avalara, Vertex, or Sovos. The rate databases, the 190-plus-country coverage, the e-invoicing and continuous transaction control capabilities of those engines remain essential. Agentic AI sits alongside the engine, handling the operations-layer reasoning the engine was not designed for. The right architecture is usually both: the determination engine for calculation, the agentic platform for the judgment work around it. For a broader controller-office perspective, see Top AI Automation Tools for Controllers and Accounting Operations Teams.
Book a working session with a Kognitos solutions engineer → Or try Kognitos free →
What the strongest indirect tax operations share in 2026
Across the indirect tax functions that run well in 2026, a few patterns recur. They keep a clear separation between the determination layer and the operations layer, buying the right tool for each rather than expecting one to do both. They treat exemption certificate management as a document-judgment problem to be handled continuously, not a folder to be audited in a panic when a notice arrives. They monitor nexus as an ongoing reasoning task rather than a quarterly spreadsheet review. And they treat audit defensibility as a property of the system, captured at the moment each decision is made, rather than a reconstruction project undertaken years later when a jurisdiction asks.
The common thread is that the hard parts of indirect tax are reasoning, document, and reconstruction problems, and the strongest operations equip those parts with tooling suited to judgment work, while leaving the calculation to the engines that have already solved it. For a 90-day evaluation framework that applies cleanly to a tax-operations pilot, see How to Score an Agentic AI Pilot: The 90-Day Evaluation Framework. For the audit conversation that follows, see What Your SOX Auditor Will Ask About Your AI Automation.