|
⬡ Shadow AI & Building Your Policy · Issue 3 of 5
|
|
|
Wednesday · July 8, 2026 · Issue #038
Before you can write a policy, you need to know what you're governing.
Most firms that decide to address AI governance skip straight to the policy document — and end up writing rules around tools they think their team uses, for workflows they assume are the risky ones, based on data they've never actually collected. The policy looks fine on paper. It governs a business that doesn't quite exist.
Today's issue is the audit that precedes the policy. Five questions. Run them this week. The answers tell you exactly what your policy needs to cover — and exactly where your firm's actual exposure sits right now.
|
|
⬡ Before You Start · How to Run This Audit Without Making It Adversarial
The framing of this conversation matters as much as the questions. People underreport when they think the goal is to catch them doing something wrong. They're more forthcoming when the goal is clearly to help them do things better.
The opening line that works: "We're building out our firm's AI toolkit and I want to understand what's actually useful to people. Can you tell me what AI tools you've used in the last 30 days — for anything work-related, including things you signed up for yourself?"
The goal is an accurate inventory, not a compliance investigation. Once you have the inventory, the questions below tell you what to do with it.
|
|
|
⬡ The 5-Question Shadow AI Audit
|
|
|
1
|
Which tools are on free consumer accounts versus paid business plans?
|
This is the most important question in the audit. Free consumer tiers of most AI tools have significantly different data practices than their business or enterprise counterparts. Whether inputs are used for model training, how long data is retained, and what data protections apply often differ substantially between account tiers.
I want to be careful here because these terms change frequently and vary by provider. Rather than cite specific policies that may be outdated, I'd recommend checking directly: for ChatGPT, look at OpenAI's enterprise privacy page and the specific settings in each account. For other tools, look for the terms of service section covering data usage for model training.
|
What to do with this answer
Any AI tool being used for work-related tasks on a free consumer account needs to either be moved to a business plan with appropriate data controls, replaced with an approved alternative, or explicitly prohibited for work use. The policy you write Thursday and Friday governs this transition.
|
|
|
|
2
|
What types of data are people actually putting into AI tools?
|
Ask this directly and specifically. Not "do you ever put sensitive data into AI?" — that question gets a reflexive "no." Instead: "Walk me through the last time you used an AI tool at work. What did you type in?"
The categories to listen for: client names or identifiers, financial figures, contract language, personnel information, health-related data, internal strategy or projections, and anything marked confidential. Each of these represents a different risk level depending on your industry and the tool's data practices.
|
What to do with this answer
This answer builds your data classification matrix — the foundation of your AI policy. What data types are in use, which should never enter an AI tool, and which require governance controls rather than outright prohibition.
|
|
|
|
3
|
Is AI output going to clients, regulators, or official records — without a documented review step?
|
AI-generated content that reaches clients, courts, regulators, or official records without human review is where the biggest professional liability exposure lives. This isn't about whether people are checking AI output — it's about whether the review step is documented, consistent, and impossible to skip under deadline pressure.
For legal firms: the Mata v. Avianca case — where AI-generated legal citations were submitted to federal court without verification — remains the clearest public example of what happens when review is treated as optional. The attorneys were sanctioned. The AI didn't bear any consequences.
|
What to do with this answer
Every AI-assisted output that leaves the firm needs a mandatory review checkpoint in the workflow — not a policy statement that review is expected, but an actual step in the process that documents who reviewed it and when. Your AI policy needs to specify this for each output type.
|
|
|
|
4
|
Does your existing client contract language cover AI tool usage — and do your clients' contracts cover it from their end?
|
Most professional services agreements written before 2023 have confidentiality clauses that almost certainly cover AI tool usage in ways the parties never discussed — because AI tools of this kind didn't exist when those agreements were written. "Disclosure to a third party" typically includes sending data to an AI platform. "Confidential information" typically includes client business information regardless of form.
Additionally, many enterprise clients now include AI governance requirements in their vendor contracts. If you serve larger organizations, check whether any current contracts require you to maintain an AI policy or limit AI tool usage with their data.
I'd recommend reviewing this with your legal counsel rather than relying on my read of it — contract interpretation is highly specific to exact language and jurisdiction.
|
What to do with this answer
At minimum, read your current client agreements for confidentiality and data processing language. Consider whether new agreements should include explicit AI usage provisions. If you're unsure, ask your attorney — this is a 15-minute conversation worth having.
|
|
|
|
5
|
If a client or regulator asked you today: "What is your AI governance policy?" — what would you say?
|
This isn't a rhetorical question. It's becoming a real one. Enterprise procurement teams are adding AI governance requirements to vendor questionnaires. Regulatory bodies in healthcare, financial services, and legal are beginning to ask about AI policies in the same way they've historically asked about data security policies.
If your answer is "we're working on it" or "we don't have one," that's the gap we're closing Thursday and Friday. If your answer is "we have something but it's vague and hasn't been reviewed in a while," that's also what we're addressing.
|
What to do with this answer
If the answer is "nothing" — tomorrow's issues give you the framework and the template to change that by Friday. The goal isn't a perfect document. It's a real one that your team can follow and that you can show a client or regulator with confidence.
|
|
|
|
⬡ Jordan · AI Solutions Director · thepromptory.com
Free · No account required · No sales call after
|
|
N
|
I ran those five questions with my team this morning. Found out we have people using 7 different AI tools across the firm. Three of them I'd never heard of. Two are definitely on free personal accounts. I don't know where to start — do I shut things down first or build the policy first?
|
|
J
|
Don't shut down before you've built the alternative — that just drives usage underground and you lose visibility. The sequence: this week, identify the two free-tier tools being used for anything involving client or sensitive data, and move those specific use cases to an approved alternative immediately. Simultaneously, build the policy framework this week. The policy legitimizes the approved tools and gives people a clear path forward. Shutting things down without a replacement is how you create a compliance-through-ignorance situation instead of a compliance-through-governance one. Which of the three tools you don't recognize are concerning you most?
Jordan · thepromptory.com →
|
|
|
Run your audit and then bring what you find to Jordan → thepromptory.com
|
|
|
|
📬 Tomorrow
Thursday: Building Your AI Policy, Part 1. The four sections every AI policy needs. Why most firms get section two wrong. How to write the data classification section so it's actually usable — not just defensible on paper.
Run the audit this week. Bring what you find here: thepromptory.com →
|
|
|