Deloitte Calls It the Continuously Operating Tax Function. Here's What That Actually Requires.
Deloitte's October 2026 report makes the case for agentic AI in tax. The infrastructure gap between a quarterly close and a continuously operating function is larger than any vendor admits.
Editorial source review — November 1, 2026. This machine-generated article has been reviewed, but is not certified accurate. Source-backed corrections are incorporated where established. Other legal, survey and product claims may remain unverified; check applicable primary authorities before relying on a deadline or tax position.
Findings in the original version: Deloitte's own article supports its continuously operating tax vision. It does not make all tax filings continuous or establish a globally uniform real-time mandate across “more than 60 jurisdictions”.
Deloitte published a piece this week from their Southeast Asia practice titled "Where Proof Meets Possibility: The AI-Enabled Continuously Operating Tax Function." The argument is sound. The destination they describe — a tax function that operates continuously rather than in quarterly sprints, using agentic AI to manage risk in real time — is the right destination. But Deloitte's piece, like most vendor and firm thought leadership on this topic, stops well short of identifying the actual prerequisite. The infrastructure gap between where most tax functions sit today and a continuously operating model is not a software gap. It is a data architecture gap.
I spent years as a Managing Director at Deloitte building tax technology practices. I have seen every version of this conversation. The destination is correct. The path is harder than it looks.
What Deloitte Is Actually Arguing
The Deloitte piece, published October 9, 2026, argues that the corporate tax function is shifting from a model of periodic compliance — quarterly provisions, annual filings, periodic audits — to a model of continuous operation. In the continuously operating model, agentic AI monitors transactions in real time, flags anomalies before the books close, maps new legislation to current transaction patterns, and pre-audits filings before submission.
Deloitte ties the shift to two external forces. First, tax authorities are themselves deploying real-time analytics, which means the window between a taxable event and an authority inquiry is compressing. A compliance function built on quarterly cadences does not respond fast enough to a world where authorities have flagged your transaction-level data before your provision team has pulled their first journal entry.
Second, e-invoicing mandates are creating real-time data flows between companies and tax authorities in more than 60 jurisdictions today. When Brazil's SEFAZ, Mexico's SAT, and the EU's ViDA framework all receive transaction-level data at the point of invoice issuance, the concept of a "filing" as a discrete event becomes structurally obsolete. You are filing in continuous time, whether your compliance function knows it or not.
The Deloitte conclusion: tax teams should use agentic AI to pre-audit their own data before authorities do. The warning they embed — and this is worth quoting — is that automating tax work on weak data foundations does not solve compliance problems. It accelerates them.
Where the Argument Stops Short
Deloitte is right that agentic AI is the operating model. They are right that data quality is the prerequisite. But they stop before identifying the specific architectural requirement that makes the continuously operating model possible: a tax knowledge graph.
Here is the gap. An agentic AI system that monitors transactions in real time needs to apply tax rules to those transactions consistently, in real time, across every jurisdiction where the company operates. To do that, the rules have to be represented in a machine-readable, structured form — not as PDFs in a document management system, not as spreadsheet logic in a regional provision template, not as institutional knowledge in the heads of senior tax managers who took a call at 11pm during a Q3 close.
The machine needs a taxonomy of every relevant tax concept — income category, deduction type, exemption rule, treaty provision, timing difference — and it needs those concepts connected to each other, to the transactions they govern, and to the jurisdictional rules that apply to them. That is a knowledge graph. Without it, an agentic AI cannot apply tax rules. It can only describe tax rules, which is a language model's function, not a compliance engine's.
This distinction matters because the compliance function requires deterministic outputs. When your provision shows a $47 million deferred tax liability, every number in that computation has to be defensible to a PCAOB auditor. A language model that summarizes tax rules does not produce defensible computations. A deterministic engine built on a structured knowledge graph does. The AI layer sits on top of the deterministic layer — it handles pattern recognition, anomaly flagging, and narrative generation. It does not replace the deterministic core.
The Four Infrastructure Requirements for Continuous Operation
A tax function that operates continuously requires four infrastructure components that most corporate tax departments do not have today.
Structured transaction data at the point of origination. The continuously operating model requires that every transaction entering the tax function is already tagged with the attributes the tax system needs: legal entity, jurisdiction, transaction type, counterparty type, currency, and the relevant tax classification. If tax teams are manually re-classifying data that was coded incorrectly in the ERP, continuous operation is not possible. The ERP integration has to produce clean, tax-attributable data before it hits the tax system.
A machine-readable rule set organized as a knowledge graph. Every tax rule applied to every transaction has to be represented in a structured form. The knowledge graph connects the rule to the transaction type it governs, the jurisdiction where it applies, the effective date range, and the authority source. This is the infrastructure that Deloitte's agentic AI needs but does not describe building.
A deterministic computation engine. The knowledge graph and the transaction data feed a rule engine that produces deterministic outputs — the same answer every time for the same inputs. This is the provision engine, the ETR calculator, the GIR allocator. It is not AI. It is deterministic logic. The AI layer — the agentic system Deloitte describes — operates on the outputs of this engine, not in place of it.
A continuous monitoring layer with defined escalation protocols. The agentic AI monitors the outputs of the deterministic engine against defined thresholds — ETR anomalies, book-tax difference spikes, jurisdiction-level outliers against prior periods — and escalates to human review when thresholds are exceeded. The escalation protocols have to be defined by tax professionals, not inferred by the AI. The AI flags. The human decides.
Without all four of these, a "continuously operating tax function" is a marketing slide, not an operating model.
What This Means for How You Buy Technology
The current generation of tax technology vendors is selling agentic AI features to tax departments that have not yet built the knowledge graph or the deterministic engine that those features require. The result is what Deloitte correctly warns against: automating on weak foundations accelerates problems rather than solving them.
When you buy a tax automation platform, ask three questions. First: does the platform represent tax rules in a structured, machine-readable knowledge graph, or does it rely on human configuration of spreadsheet-equivalent logic? Second: is the computation engine deterministic — does it produce the same output for the same inputs every time, and can you trace every output back to the specific rule that produced it? Third: where does the AI layer sit in the architecture — on top of a deterministic engine, or in place of one?
If the answers are no, no, and in place of, you are buying a language model wrapper around an existing data export. You are not buying infrastructure for a continuously operating tax function.
At Clarity Tax Technologies, we built the knowledge graph first. The deterministic engine second. The AI layer third. The sequence is not an architectural preference. It is the only sequence that produces auditable compliance outputs at the scale and speed that a continuously operating tax function requires.
What to Do Monday Morning
1. Assess your ERP integration for tax-attributable data quality. Pull a sample of 100 transactions from the last quarter. For each one, how many required manual re-classification before the provision team could use them? If the answer is more than 20, your continuous operation prerequisites are not in place. 2. Identify where your tax rules live today. PDFs? Spreadsheet logic? Senior manager knowledge? Map the rule storage to the transaction types that depend on it. The gaps in that map are your knowledge graph requirements. 3. Read Deloitte's piece with the infrastructure lens. For each capability they describe — real-time anomaly detection, pre-audit filing review, legislative mapping to transactions — ask which of the four infrastructure requirements it depends on. Score your current state against each requirement. 4. Require a determinism demonstration from every AI tax vendor you evaluate. Ask them to run the same provision scenario twice with identical inputs and show you the outputs. If the outputs differ, the system is not deterministic. Do not buy it for provision work. 5. Set a 90-day goal to document your tax ontology. Start with the 20 transaction types that produce 80% of your book-tax differences. Define each concept, its relationships to other concepts, and the rules that govern it. That is the foundation of your knowledge graph. You do not need a technology platform to start. You need a structured taxonomy.
The continuously operating tax function is not a future state. It is an architectural decision that companies make or do not make in the next 18 months. The companies that build the infrastructure now will have a durable compliance advantage. The ones that buy AI features without the infrastructure will spend the next five years reconstructing positions under audit.