Is Customer Data Safe with AI Tools? A Guide for Accounting

Accountants Are Already Heavy AI Users — Whether They Planned to Be or Not

If your firm hasn't formally rolled out AI, it may not matter — your staff probably already has. Professional services, the sector that includes accounting, leads U.S. AI adoption at roughly 33% of firms, ahead of even financial services at about 30%, according to Federal Reserve data.

The worker-level numbers are just as striking. 62% of professional-services workers report using generative AI at work, and 12% use it daily. That means client tax records, payroll data, and financial statements are likely flowing through AI tools right now — often without anyone checking where that data goes.

This is the exposure problem in a nutshell. Most AI use in firms is ad hoc: an associate pastes a client's trial balance into a chatbot to draft an explanation, or a manager feeds contract language into a summarizer. Each act seems harmless, but collectively they create an unmanaged data pipeline with no policy, no vetting, and no audit trail.

The risks span the entire AI lifecycle. In joint guidance released in May 2025, CISA, the NSA, the FBI, and international partners emphasized "the critical role of data security in ensuring the accuracy, integrity, and trustworthiness of AI outcomes" — covering development, testing, deployment, and operation alike.

For accounting specifically, the stakes are amplified by the nature of the work:

• Client data is confidential by professional obligation, not just by contract.
• Financial data is highly attractive for identity theft and fraud.
• Cognitive, analytical tasks — exactly where adoption is most prevalent — involve the most sensitive inputs.
• Small and medium practices often lack dedicated IT security staff.

Trust turns out to be the decisive variable. A peer-reviewed study of 238 accounting professionals found that trust and vendor support positively influence AI adoption intention, while perceived cost is a barrier. Firms that skip the trust-building step — vetting how a tool handles data before staff adopt it — effectively let individual employees make security decisions on the firm's behalf.

The profession is responding. EFAA has published an AI in Accounting Guide aimed squarely at small and medium practices, with what EFAA's Paula Franco calls "a strong focus on data protection and ethical best practices."

So the core question isn't whether accountants will use AI — they already do, at rates few industries match. The question is whether client data is safe in those tools, and whether the firm is managing that use or leaving it to chance. Ad hoc adoption is the real risk, and it's fixable with a deliberate policy.

One more layer is coming: agentic AI, where coding agents and automated workflows touch client data directly. Published research found roughly 26% of community-published agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans, version-pins, and sandbox-benchmarks every skill before you equip an agent with it. The same vendor-trust discipline applies there as anywhere else.

What Actually Makes AI Tools Risky with Client Data (and What the Evidence Says)

Your firm's client files are already flowing through AI tools — the question isn't whether that creates risk, but whether anyone is checking. Accountants sit in the sector with the highest AI adoption in the economy: roughly 33% of professional services firms now use AI, and work-related generative AI use among those professionals has reached 62%.

The most authoritative picture of what can go wrong comes from the U.S. government itself. In May 2025, CISA, the NSA, and the FBI jointly published guidance on securing data used to train and operate AI systems. Their core warning: data risks don't live in one place. They exist across the entire AI lifecycle.

That lifecycle framing matters for accounting firms because client data can touch an AI tool at multiple points:

• Development and testing — data used to train or fine-tune a model may be retained, reused, or exposed
• Deployment — configuration errors and integration gaps can leak inputs between clients
• Operation — every prompt, document, and ledger extract a staff member pastes becomes data the vendor may process or store

The agencies' position is blunt: data security is what guarantees "the accuracy, integrity, and trustworthiness of AI outcomes." A model that handles client data insecurely isn't just a compliance problem — it can quietly produce wrong answers.

Here's the uncomfortable gap. A peer-reviewed study of 238 accounting professionals found that trust and vendor support are among the strongest proven drivers of AI adoption. Yet the research says little about firms actually verifying what vendors do with data before adopting. Trust, in practice, is often assumed rather than earned through vetting.

That gap is widening as adoption accelerates — overall firm AI adoption grew 68% in the year ending September 2025, and over 20% of firms expected to begin using AI in the first half of 2026. Firms extending AI into automated workflows face the same discipline problem at the component level: agent skills and connectors need the same security scrutiny as the tools themselves, which is the verification approach SkillzBench is building into its free skills directory.

The honest bottom line: no source tracks AI-specific breaches at accounting firms yet. The risk is structural, not statistical — and the firms that vet vendors before trusting them with client data are the ones writing the safer playbook.

The Solution: Anchor Your Firm's AI Security on Official Guidance, Not Vendor Marketing

When a firm considers an AI tool, the most reliable question isn't "what does the vendor's website promise?" but "what do the official security authorities say?" Vendor marketing is written to sell; government and professional-body guidance is written to protect you. Anchor your evaluation on the latter.

Two frameworks deserve a place on every accounting firm's shelf. First, on May 22, 2025, CISA, the NSA, the FBI, and international partners jointly released "AI Data Security: Best Practices for Securing Data Used to Train & Operate AI Systems," emphasizing "the critical role of data security in ensuring the accuracy, integrity, and trustworthiness of AI outcomes." The guidance covers risks across the entire AI lifecycle — development, testing, deployment, and operation — not just the moment you sign up for a tool.

Second, the European Federation of Accountants and Auditors for SMEs published its first "AI in Accounting Guide," built specifically for small- and medium-sized practices, with what EFAA's Paula Franco calls "a strong focus on data protection and ethical best practices." Monica Foerster notes that SMPs "will face many challenges navigating the transformative impact of AI," and the guide offers "timely, actionable insights" — available in French, Dutch, Spanish, and German.

The urgency is real. Roughly 33% of professional services firms — a category that includes accountants — have already adopted AI, and a peer-reviewed study of 238 accounting professionals found that trust and vendor support are decisive factors in whether AI gets adopted at all. Cost is a barrier; trust is a gate. That means vetting vendors isn't optional paperwork — it's the mechanism through which adoption succeeds safely.

Turn both frameworks into a practical vetting checklist before any tool touches client data:

• Security capabilities: Does the vendor document how it protects data across the AI lifecycle, consistent with CISA's guidance?
• Data handling: Where does client data go, how long is it retained, and is it used to train models?
• Support quality: The research links vendor support directly to adoption outcomes — test responsiveness before you commit.
• Trust signals: Look for independent verification, transparent documentation, and evidence rather than slogans.

The same discipline applies as firms extend AI into agentic workflows — coding agents and assistants that automate accounting tasks. Agent skills (instruction packages that extend what these agents can do) vary wildly in quality, and published research suggests roughly a quarter carry at least one vulnerability. That's why SkillzBench is building a free, independent directory that security-scans, version-pins, and sandbox-benchmarks agent skills, so firms can equip agents only with vetted tools — applying the same trust-first standard to skills that they'd apply to any AI vendor.

Official guidance gives you the standard. A disciplined checklist makes it operational.

A Practical Playbook: Securing Client Data Before Your Next AI Prompt

The gap between firms using AI and firms securing AI use is widening fast. Professional services — accounting included — leads every industry in AI adoption at roughly 33% of firms, and more than 20% of all U.S. firms expect to put AI to work in the first half of 2026, according to Federal Reserve survey data. If your firm doesn't have a data security policy yet, the time to write one is before the next prompt, not after.

Start with a written firm AI policy. It doesn't need to be long, but it should name which tools are approved, who may use them, and what happens when someone is tempted to paste a client's tax file into a chatbot. Ad hoc, unmanaged use is where client data leaks first. Anchor the policy on authoritative guidance rather than guesswork: CISA, the NSA, and the FBI jointly published best practices for securing data used to train and operate AI systems, covering risks across development, testing, deployment, and operation.

Next, classify your data. Create a simple tier list of what may never touch an AI tool — client identifiers, tax returns, payroll records, engagement files — and what may be used only in approved, vetted tools. Make the "never" list explicit and post it where staff will actually see it.

Then vet vendors before adopting anything. A peer-reviewed study of 238 accounting professionals found that trust and vendor support are decisive factors in AI adoption, while cost is the main barrier. That finding cuts both ways: vendors who can't clearly explain their data-handling practices haven't earned your trust, no matter how attractive the price.

• Adopt a written firm AI policy now, before adoption pressure makes the decision for you.
• Classify client data into "never touches AI" and "approved tools only" tiers.
• Vet every vendor against the CISA/NSA/FBI data security guidance.
• For small and medium practices, use the EFAA AI in Accounting Guide as a profession-specific starting point.

Finally, extend the same discipline to agentic workflows. If you're equipping coding agents or chat-app agents with skills to automate accounting tasks, only install skills that have been security-scanned and pinned to a specific commit — published research found roughly 26% of agent skills carry at least one vulnerability. That's exactly the problem SkillzBench is building for: a free directory where every skill is version-pinned, license-checked, security-scanned, and benchmarked in a sandbox, plus a free MCP connector so your agent only sees the skills you deliberately chose.

Trust is earned through verification, not vendor promises. If you want to go further, browse and test agent skills free on SkillzBench — and join the waitlist to be first in line when the directory and MCP tool go live.