5 tips to protect your firm's AI use

AI is becoming ubiquitous in the workplace, from asking ChatGPT for advice to incorporating enterprise-level software to automate workflows. Patricia Thaine, co-founder and chairwoman of Private AI, shares how accounting firms can utilize these tools while realistically implementing guardrails to protect sensitive client information.

If you enjoy this Q&A, let us know! Reply to this email and share why it was relevant.  

—Interview by Lauren Ward, edited by Bianca Prieto

Many employees use AI informally to save time, sometimes without realizing the privacy implications. What kinds of risks do you see most often in how AI tools are used in accounting firms today?

Accounting firms are sitting on some of the most sensitive data that exists. Social Security numbers, bank account details, tax records, financial statements. And what we're seeing is that among enterprise employees who use AI tools, 77% paste data directly into those tools, with 82% of that activity happening through personal accounts that bypass enterprise controls entirely.  

In January 2025, Sage had to temporarily suspend its Copilot AI assistant after a customer reported that when they asked for a list of recent invoices, the AI pulled data from other customer accounts. The especially tricky part for accountants is the "messy data problem." 

According to Gartner, 80% to 90% of enterprise data is unstructured: tax documents, client communications and financial statements where personal information isn't neatly labeled in database fields. Someone might say, "My account number is a five, three, no, sorry, actually that was a five-two" in a call transcript. You need AI to even identify what's PII in that context, but you also need protection before that data touches external AI systems.

What this means is that firms need a privacy layer that can identify and redact sensitive information before data ever leaves your environment, not after the fact when it's already too late, along with appropriate and strict data access controls.

You've talked about the need for governance to mitigate AI data exposure risk. What is the minimum level of guardrails CPA firms need to have in place, both for employees and the company as a whole?

First, inventory what AI tools your people are actually using. You can't protect what you don't know exists. 

Second, create an acceptable use policy that classifies uses as green light, yellow light or red light according to the AI tools being used (where they send the data to, what they do with the data) and the types of data being used (do they contain client information? Corporate confidential information?).  

Third, and this is where thinking about fundamentals really pays off, require data classification before any AI use. Make it a habit: before you prompt an AI, you automatically check "what sensitivity level are these data?" That single checkpoint prevents many disasters.

Fourth, deploy enterprise AI platforms only, because the difference between ChatGPT free and Azure OpenAI is enormous: enterprise versions don't use your data for training, offer SOC 2 compliance, provide audit trails and give you actual administrative control. 

Finally, if you embed privacy at the very beginning, it'll save you a lot of headaches. Consider deploying PII redaction technology that strips sensitive information before data reaches any AI system. The goal isn't to block AI use; it's to enable it responsibly.

Once guidelines are in place, how can firm owners ensure they're being followed in day-to-day employee behavior?

At the technical level, deploy DLP (data loss prevention) that can inspect prompts before they're submitted to AI systems. That way, you can identify when someone's about to send a Social Security number or client financial data to an AI provider and either block them, seamlessly replace them or alert the user in real time. 

Now, a trickier part is having visibility on which AI tools are being accessed, by whom, and how often. In theory, CASB tools (cloud access security brokers) give you visibility into this and can block access, but in practice, some providers and websites may sneak AI and data ingestion without the CASB tools being fully able to pick those up. Nevertheless, deciding what to do regarding employee access to the most popular AI tools and providing a viable alternative to those you wish to block are key. 

The most effective control is actually making doing the right thing the easy thing. If your approved AI tools are clunky, slow or hard to access, people will route around them. If you provide genuinely useful enterprise AI platforms with good training on how to use them effectively, and you deploy PII redaction and appropriate data access controls that let people work with real data safely and effectively, compliance becomes the path of least resistance.

In addition to implementing usage guidelines, what other steps should accounting firms take to protect sensitive client data before using AI tools?

The core challenge with corporate data is that roughly 80% to 90% of enterprise information is unstructured: emails, PDFs, scanned documents, conversation transcripts. Just think about trying to identify personal information in a tax return from a scanned document that has been processed into text with errors, or a client email where they're sending their account number without context. And then you want to run AI on top of that data.

Now imagine if your accountants can take client documents, run them through a reliable system to automatically identify and redact sensitive information without the data leaving their environment, and then use AI tools for analysis, research or drafting. If you embed privacy in the very beginning, it'll save you a whole lot of headaches down the line.

The Net Gains is your one-stop shop for fresh, FREE accounting insights. You can reach the newsletter team at thenetgains@mynewsletter.co. We enjoy hearing from you.

Interested in advertising? Email us at newslettersales@mvfglobal.com

 If you've been enjoying the newsletter, don't keep it a secret. Share it with an industry colleague. (Copy the link here.)

The Net Gains is curated and written by Lauren Ward and edited by Bianca Prieto.