Your Tax Authority Has Better AI Than You Do

Revenue authorities are deploying AI for audit selection faster than taxpayers are deploying it for compliance. What the authorities have actually bought, what algorithmic selection changes, and what it means for your risk posture.

Insight article | 30 July 2026

---

In May 2026, HMRC signed a reported £175 million contract with Quantexa to scan tax returns, cross-reference financial data, link shell companies and trace hidden beneficial ownership (IFC Review). Two months later, Thomson Reuters published survey data showing that 26% of indirect tax teams still reconcile their data entirely by hand, and only 16% have fully automated reconciliation (2026 Indirect Tax Report, 290 professionals across the US, Canada, Mexico and Brazil, published 29 July 2026).

That is the whole story in two numbers. The people who examine your filings are industrializing. The people who prepare them are, in a quarter of cases, using a spreadsheet and a highlighter.

This is not a piece about whether AI is coming for tax. It is about which side of the table got there first, and what a growing business should do about it in the next twelve months.

One sourcing note. The HMRC contract value is corroborated across outlets but not confirmed against a primary contract notice, so treat it as well-reported rather than verified. We flag the same caution wherever it applies below — this article is about documentation discipline, and it would be poor form not to practice it.

---

What the authorities have actually bought

United Kingdom. The Quantexa engagement, announced 14 May 2026, is for contextual decision intelligence — entity resolution and graph-linking rather than simple risk scoring. Risk scoring asks "does this return look odd?" Graph-linking asks "who is this entity connected to, and does the network make sense?" HMRC had been testing AI at smaller scale for two years beforehand. Context for the spend: a UK tax gap of roughly £36 billion annually.

United States, federal. The IRS reported 126 AI use cases as of June 2025, up from 10 in August 2022 — 77 (61%) still in development, a further 65 withheld as too sensitive or R&D-exempt. AI spending exceeded $58 million in FY2024, with $28M in FY2025 and $32M projected for FY2026 for a single unit (GAO-26-107522, 24 March 2026). A reported $1.8 million Palantir contract is testing three AI models specifically in case selection (Money — secondary).

United States, state level. This is the one most growing businesses will actually meet. South Carolina's Department of Revenue signed a $3.5 million, five-year contract with FAST Enterprises in December 2025, launching in 2026. The FAS audit-review module plugs into the state's GenTax platform and flags anomalies — the reported example is businesses with atypical sales tax patterns compared to their peers (Post and Courier). GenTax already runs many state revenue departments, so the distribution channel for state-level AI audit selection is pre-built.

Elsewhere. Malta reports €400M in additional tax raised in 2024 using predictive AI; the Australian Taxation Office $530M+ in unpaid tax identified and $2.5B in fraudulent claims prevented; Italy's VeRa algorithm more than a million high-risk cases found by cross-matching filings, earnings, property, bank accounts and e-payments; Poland's STIR moved carousel-fraud detection from a two-month lag to near-real-time (vatcalc — a secondary aggregator; verify individual figures before quoting them).

---

What AI-driven selection actually changes

The instinctive fear is "more audits." That is not what the data shows, at least not in the US. Volume has collapsed: 205,000 audits opened in FY2025, against 335,000 in FY2024 and more than a million around 2017, after the loss of nearly a third of audit staff (Money — secondary). The IRS lost more than 25% of total staff, and its Research, Applied Analytics and Statistics group lost 63 employees working on AI (Accounting Today).

So the change is not volume. It is three other things.

Selection gets sharper. Fewer examinations, better targeted. If you are selected now, it is more likely because a model found something specific than because your return came up in a rotation. The base rate of "audited but fine" falls; the base rate of "audited and it hurts" rises.

Peer comparison becomes the default lens. The South Carolina description is explicit: businesses flagged for atypical patterns versus peers. A perfectly defensible margin, effective rate or exemption ratio can look anomalous against a peer set you never chose and cannot see.

Networks, not just returns. Graph-linking evaluates the entity alongside its relationships — shared directors, shared addresses, intercompany flows, ownership chains. Structures previously visible only if a human examiner went looking are now visible by default.

The corollary is uncomfortable: "why was I selected?" currently has no good answer — for the authority or the taxpayer. That gap makes taxpayer-rights litigation over algorithmic selection a realistic 2027–28 development, and it is why every authority now cites the same two warnings: the Netherlands childcare benefits scandal, where an algorithmic risk system falsely labeled thousands of families as fraudsters and the Dutch government collapsed over it in 2021, and Australia's "Robodebt" (2016–2019), where automated welfare debt-raising issued false demands against vulnerable citizens and allegedly contributed to suicides. Both appear explicitly as cautionary references in the South Carolina reporting — to whose credit the design responds directly: human review at every stage, and a closed loop keeping taxpayer data out of public AI services. That should become the standard.

---

Meanwhile, inside the corporate tax function

Here is the asymmetry stated plainly.

| The authority side | The taxpayer side |
|---|---|
| HMRC: reported £175M graph-analytics platform (2026) | 26% of indirect tax teams reconcile entirely manually |
| IRS: 126 AI use cases, >$58M spent in FY2024 | 87% of indirect tax teams run multiple disconnected systems |
| South Carolina DOR: $3.5M / 5 years, launching 2026 | Only 16% have fully automated reconciliation |
| Italy VeRa: 1M+ high-risk cases identified | Fewer than 25% have reached agentic AI automation |

Sources as cited above, plus the Thomson Reuters 2026 Indirect Tax Report for the right-hand column.

It is not that tax departments are refusing to move. They are not funded to. Thomson Reuters' 2026 Corporate Tax Department Technology Report (170 tax leaders, 16 March 2026) found AI usage in corporate tax departments remains relatively low, fewer than half of departments received a budget increase last year, many saw cuts, and a majority of tax leaders still rate upgrading tax technology as a relatively low priority — a widening "frustration gap."

Deloitte's AI-enabled Tax Transformation research fills in the barriers: 45% budget constraints, 36% limited AI expertise, 33% no clear AI strategy, 30% data security and privacy concerns, 30% difficulty integrating tax data across the organization, 28% limited technology and data management expertise, 28% insufficient leadership support.

That last cluster is the real problem, and it is not an AI problem.

---

The vulnerability is your data, not your software

Every AI use case above — on both sides of the table — runs on the same fuel: structured, reconciled, traceable transaction data. The authority assembles that fuel from your filings plus third-party sources. Many businesses cannot assemble it from their own systems.

Consider what a real examination now asks for. Not "show us the return," but: show us the transactions behind this line; show us how the ERP figure became the filed figure; show us the source document for this deduction; show us why this contract was treated this way at the time, not the rationalization you constructed last week.

The Deloitte finding that 77% of tax leaders require 90% or higher accuracy before entrusting AI with tax processes is usually read as caution about AI. Read it the other way. If your own data only supports 90% confidence, the problem is not the model.

This is the underrated bottleneck across the whole field: normalizing data from PDFs, scanned invoices, K-1s, trial balances and multiple ERP instances into a tax-ready model before any reasoning — human or machine — can occur. More than half of indirect tax teams name data reconciliation as their top automation priority (Thomson Reuters). They are right, and they are behind.

The authorities have the same disease. GAO found more than 25% of IRS AI use cases lacked any information on how the use case was to benefit the agency, several tools used in criminal case development were missing from the inventory entirely, and no single entity manages AI investments across the agency. That does not help you. An under-governed model still generates a letter.

---

What a growing business should actually do

Nothing here requires an AI budget. All of it requires discipline.

1. Documentation discipline. Every material position gets a written rationale at the time the position is taken, naming the authority relied on and the facts assumed. If the rationale is written after the enquiry letter arrives, it is worth less — and increasingly, it looks it.

2. Contemporaneous evidence. Contracts, board minutes, calculations, correspondence, timesheets, project records for R&D claims. Dated, stored, findable. The advisory market around this is already forming: EmpowerRD markets R&D claim preparation explicitly against HMRC's new AI (EmpowerRD).

3. Data lineage. For every number on every filing, trace source system → transformation → adjustment → filed figure. If that trace lives only in one person's head or one undocumented spreadsheet, you have a single point of failure that an examiner will find faster than you will.

4. Reconciliations that survive scrutiny. Statutory accounts to tax computation. ERP to filed VAT/GST returns. Payroll to filings. Intercompany balances both ways. On a schedule, signed off, with breaks explained rather than plugged.

5. Know your own peer profile. If selection runs on peer comparison, understand where you sit. Unusual is not wrong — but unusual and unexplained is an invitation.

6. Fix the source documents before you buy the tool. AI on top of disconnected systems reproduces the mess faster. Data first, automation second.

---

Readiness checklist for a growing business

Score each item: Green (documented and current), Amber (exists but stale or informal), Red (does not exist).

| # | Item | Test |
|---|---|---|
| 1 | Position log | Dated written rationale for every material position of the last three years? |
| 2 | Source-document repository | Underlying document for any filed figure retrievable within one working day? |
| 3 | Data lineage map | Written trace from source system to filed figure, for each major tax? |
| 4 | Reconciliation calendar | Statutory-to-tax, ERP-to-return and payroll-to-filing reconciliations scheduled and signed off? |
| 5 | Break register | Differences logged and explained, rather than plugged? |
| 6 | Single-point-of-failure test | If your longest-serving finance person left tomorrow, could anyone reproduce last year's filings? |
| 7 | Peer profile awareness | Do you know how your margins, effective rate and exemption ratios compare to your sector? |
| 8 | Entity and ownership map | Group structure, directorships and intercompany flows documented and current? |
| 9 | R&D and incentive claims | Supported by contemporaneous project records, not retrospective narrative? |
| 10 | E-invoicing exposure | Do you know which national mandates and ViDA dates apply to you? |
| 11 | Enquiry playbook | Named owner, response protocol and disclosure position agreed before a letter arrives? |
| 12 | Advisor alignment | Does your external advisor hold the same documentation set you do? |

Any Red on items 1–5 is where to start. A Red on item 6 will hurt most and usually gets fixed last.

---

The uncomfortable summary

Your tax authority has better AI than you do. It also has better data about you than you may have about yourself, because it can join your filings to third-party sources you cannot see. You will not out-spend it and you do not need to. You need to be the taxpayer whose numbers reconcile, whose documents exist, and whose positions were written down on the day they were taken.

That was always good practice. It is now the difference between a routine enquiry and an expensive one.

---

General information, not tax or legal advice. Figures are reproduced from the sources cited; secondary items are flagged and should be verified before public reliance.

EZ Consulting offers an Algorithmic Audit Readiness Review — a fixed-scope engagement that stress-tests your business against the checklist above and delivers a prioritized remediation plan.

---

Sources