The Nexairi Dispatch · Friday, May 1, 2026 · Issue #9
The IRS wants to do your tax prep for you
The agency already has your income data. A bill would let filers use it before you see it.
Good morning, friends. Congress introduced a bill that would let your clients download a pre-filled tax return directly from the IRS — W-2s, 1099s and Social Security data already entered. Meanwhile, a tax autopilot that started with the top 20 CPA firms quietly opened to everyone this week. And the AI-powered scams targeting your clients have gotten convincing enough that even cautious people are getting fooled. Big week.
📋 TAX POLICY — Congress Bill Would Let Clients Skip Data Entry
What happened: Rep. Bill Foster introduced the AutoFill Act, which would let any U.S. taxpayer download a pre-filled tax return directly from the IRS. The form arrives with W-2, 1099 and Social Security data already entered — no income limit, no TurboTax required. Unlike Free File, it produces machine-readable files that work with existing tax software.
Why it matters: If it passes, the hours CPAs spend on data entry become optional — and clients may start to wonder what they need a preparer for. Advisory services, exception handling and interpretation become the billable work that survives. Firms pivoting toward planning and strategy will feel this less. Firms still competing on speed of return preparation will feel it first.
What to watch: The bill is in early stages. Tax prep software companies oppose it — their lobbying defeated a similar proposal in 2022. Watch whether the AICPA weighs in, since their position will shape how the profession responds.
🔧 TAX AUTOMATION — Black Ore Tax Autopilot Opens to Every CPA Firm
What happened: Black Ore Tax Autopilot started in closed beta with the top 20 CPA firms. Now it's open to all firms. Forty percent of the biggest accounting firms in the country used it during beta. The platform reads source documents, extracts data and generates returns ready for CPA review — without a human keying in each field.
Why it matters: For high-volume practices, the ROI case is clear: less data entry, faster turnaround, same review step. The question for smaller firms is whether their return volume and client mix justify the price. Firms doing fewer than a few hundred returns a year may not see payback in year one. Those doing 500-plus should run the numbers now.
What to watch: Watch how quickly clients start expecting faster turnaround once the big firms standardize AI-reviewed returns. Firms without automation will feel margin pressure before they realize the baseline shifted.
🤖 AI READINESS — Small CPA Firms Have AI Tools — Most Don't Use Them
What happened: Intuit, Wolters Kluwer and Thomson Reuters all launched agentic AI platforms in the past 18 months, priced at $200-400 per month. Only 20% of small and midsize businesses use AI in their finance operations, despite 75% saying they've invested in it. The gap shows up in three places: documented workflows, clean data and staff training.
Why it matters: The tools are ready. The pricing is no longer the barrier. What's missing for most small CPA firms is process maturity. An AI agent handed access to unorganized data or undocumented workflows doesn't save time — it creates risk. Firms that fix the foundation first will adopt faster and with fewer errors.
What to watch: Vendors are starting to bundle readiness assessments with their products. If that becomes a standard pre-sales step, it signals the industry knows the readiness problem is widespread and is trying to close it themselves.
⚠️ CLIENT SAFETY — AI Scams Your Clients Won't Recognize in 2026
What happened: Generative AI has erased the classic signs of a scam — bad grammar, generic salutations, suspicious sender addresses. Criminals now use the same tools that power customer service chatbots to write personalized phishing emails, clone voices and produce video deepfakes of real people. What used to take hours per target now runs in bulk.
Why it matters: Your clients are the target. Your client list is a map for fraudsters: clients trust you and act on your word. A fake email from your firm — asking for a wire transfer or account update — is now technically easy to produce. Firms need a client communication protocol and should share it with clients directly.
What to watch: Watch for FTC and SEC guidance on AI-generated fraud in financial services. Several enforcement actions are in the pipeline targeting firms that failed to warn clients about the risk.
Outside Nexairi
Alibaba's AI Agent Framework Cuts Redundant Tool Calls by 96% — VentureBeat
Alibaba's Metis framework trains agents using reinforcement learning to skip unnecessary steps, dropping redundant tool invocations from 98% to 2% while improving accuracy on reasoning benchmarks. For firms evaluating AI agents, this is what cost-efficient agentic AI actually looks like.
Coordinated AI Agents Are Moving Into White-Collar Work — MIT Technology Review
Orchestrated multi-agent systems — where specialized agents pass tasks to each other — are leaving research labs and entering business workflows, promising to automate complex multi-step work that single AI models can't handle alone.
Netomi Raises $110M from Accenture and Adobe for Customer AI — VentureBeat
The customer experience AI platform secured strategic backing from two enterprise giants, a signal that AI-powered service automation is moving from pilot programs to standard business infrastructure.
OpenAI's RLHF Training Left Hidden Behavioral Patterns in Its Models — VentureBeat
Researchers found that OpenAI's personality training embedded unintended behavioral correlations that persist across unrelated contexts — a concern for any organization relying on model outputs for compliance-sensitive work.
Tool Worth Knowing: Silico by Goodfire
Silico lets developers inspect and adjust what's actually happening inside an LLM — think of it as a debugger for model behavior rather than model outputs. For teams deploying AI in compliance-sensitive work, it's the first practical tool for verifying that a model will behave consistently before it touches client data.
Deeper Read
This Startup's New Tool Lets You Debug LLMs — MIT Technology Review
Goodfire's Silico uses mechanistic interpretability to let practitioners inspect and modify AI model internals — useful for any firm that needs to trust what an AI model is actually doing before it handles client data.
Six Exploits Broke AI Coding Agents — IAM Never Saw Them Coming — VentureBeat
Security researchers disclosed six attack paths across Codex, Claude Code and Copilot where malicious inputs stole credentials and escalated permissions without triggering identity management alerts — a governance gap that applies wherever AI agents have system access.
Quick Hits
- RunPod Flash removes Docker for serverless AI development
- Tesla Optimus: production starts at Fremont in July 2026
- AI 2027, translated for your business and clients
- ChatGPT Workspace Agents: what they do and what to watch
- GPT-5.5 is out: what changed and what it costs
- DeepSeek-V4's 1M token window is built for AI agents