The Nexairi Dispatch · Wednesday, May 13, 2026 · Issue #14
Who watches the AI that watches your taxes?
126 systems and no governance plan. The IRS's AI audit expansion has outrun its oversight structure.
Good morning, friends. The IRS now runs 126 AI systems to flag audit candidates — and still doesn't have a plan for who trains the people overseeing them. PwC is deploying AI audit tools at scale while the PCAOB is still researching whether to write standards for it. And CFOs shopping for AI compliance tools are finding a category that hasn't agreed on what compliance means yet. The governance is catching up. The deployments aren't waiting.
🏛️ TAX COMPLIANCE — IRS runs 126 AI systems, still lacks a governance plan
What happened: The GAO released a report finding the IRS expanded its AI inventory from 10 systems in 2022 to 126 in 2026. The agency lacks a workforce plan for AI skills gaps and has poor inventory tracking of the tools it already deployed. AI-assisted audit selection is now routine, not special.
Why it matters: The IRS agent still reviews and approves before sending an audit notice — that's policy. But there's no framework for training the people overseeing those AI decisions. The practical response for practitioners is stronger pre-filing documentation on high-risk returns: large partnerships, high-income self-employed, S-corps with unusual deductions and crypto activity.
What to watch: PCAOB-style governance standards for IRS AI are not coming soon. Until they do, documentation is your defense.
💼 CFO TOOLS — AI compliance tools aren't one category. A buying guide.
What happened: Three types of AI compliance products now reach CFOs under the same label. Governance platforms catalog which AI tools a company uses. GRC systems map AI risk to legal and audit controls. Finance-specific workflow layers document who signed off on an AI-assisted output before it touched a report. Each solves a different problem.
Why it matters: CFOs buying without a clear problem statement are picking the wrong layer. The right question before buying: where is your evidence gap? Most auditors will ask for human-review documentation on AI-assisted finance outputs before they ask for an AI governance platform. Start with the documentation layer.
What to watch: Vendors in all three categories are expanding their feature sets. The CFOs who clarify what they need first will spend less and move through auditor questions faster.
🔍 AUDIT STANDARDS — PwC deploys AI audit tools. Nobody has validated them.
What happened: PwC announced an AI Center of Excellence on May 11 to deploy audit AI at scale. Five days earlier, the PCAOB listed AI audit validation as a research priority, meaning formal standards are 18 to 24 months out. In the gap, firms are validating their own tools using internal protocols they write themselves.
Why it matters: There's no independent body, no PCAOB certification and no industry testing framework for audit AI right now. The firms deploying tools today are also writing the playbooks for how those tools get validated. When standards arrive, firms that skipped documentation will face retroactive compliance work.
What to watch: Start documenting tool selection decisions and testing protocols now. The firms with records will have the easier conversation when the PCAOB standard lands.
🏢 ACCOUNTING FIRMS — Grant Thornton built its own audit AI. Small firms need a plan.
What happened: Grant Thornton launched gtap, a proprietary audit platform that treats AI as core firm infrastructure rather than a vendor add-on. The firm built its own stack instead of buying from a third party. For large firms, that approach is becoming table stakes. For smaller ones, it's a signal that infrastructure decisions are arriving.
Why it matters: Smaller firms can't build what Grant Thornton built. But deciding not to respond is also a strategy — and probably the wrong one. The practical options are to build in-house, buy from specialized vendors, join a network with shared infrastructure, or define a niche practice where deep AI infrastructure isn't required.
What to watch: Platform-level decisions are happening among smaller firms right now. Step one is auditing your own tool stack to understand what client data flows where.
Outside Nexairi
Finance teams are buying AI faster than leaders can govern it — MIT Technology Review
Employee adoption is outpacing the governance frameworks most finance departments have in place. Capital One outlines how centering AI decisions on customer outcomes speeds up deployment without losing oversight.
Acemoglu says watch AI agents, not AI announcements — MIT Technology Review
Nobel economist Daron Acemoglu identifies three signals worth tracking: whether AI agents are actually deployed at scale, how many economists tech firms are hiring and whether practical applications are materializing. His read on near-term productivity gains is measured.
ChatGPT's Q1 growth came from an unexpected demographic — OpenAI
OpenAI's Q1 2026 update shows adoption growth came primarily from older demographics and more balanced gender representation, suggesting the tool is past its early-adopter moment and into mainstream professional use.
Tool Worth Knowing: Hopper (producthunt.com/products/hopper-4)
Hopper is the first agentic development environment built for mainframe and COBOL code. Most of the world's banking transactions still run on COBOL systems, and Hopper is an early answer to how those institutions get AI-assisted development workflows without rewriting 40 years of infrastructure.
Deeper Read
How enterprises are scaling AI — OpenAI
OpenAI's enterprise guide maps the move from pilot to production, focusing on governance structure, workflow optimization and quality assurance as the steps most organizations skip on the way to real adoption.
World Models: 10 Things That Matter in AI Right Now — MIT Technology Review
MIT Technology Review's editors break down world models — AI systems that reason about physical reality rather than just text — as one of ten developments most likely to shape how AI gets applied over the next two years.
Quick Hits
- OpenAI DeployCo: a new company whose only job is enterprise AI deployment
- Year-round tax planning cuts liability 15–25% vs April-only filing
- Karbon's The Loft: AI training benchmarks for accounting firms
- Touchless payroll vendor claims: a three-question checklist
- OpenAI: how Codex runs safely with sandboxing and approval workflows
- GPT-5.5 with trusted access opens to verified cyber defenders
- AWS + HuggingFace: infrastructure building blocks for foundation model training
- display.dev: publish agent-generated HTML behind company authentication
- MY AI Agent: one sentence assembles a 3–10 agent team