All Blogs
Here is a scene that plays out in professional services firms every day.
A paralegal is drafting a client brief. She has a deadline. She knows a good AI tool could get her a first draft in twenty minutes instead of two hours. She opens a free version of ChatGPT on her work computer, pastes in the relevant case notes, and gets her draft. She is not trying to create a problem. She is trying to do her job.
Three offices down, a financial advisor is preparing for a client review. He copies a summary of the client’s portfolio into an AI tool to help him frame the conversation. He gets a useful outline. He does not think twice about what he just put into a system he has no contract with, no terms reviewed, and no governance over.
Neither of those employees is doing something they believe is wrong. Neither of them has been told not to. And in both cases, sensitive client information has just left your organization’s control.
That is shadow AI. And the question is not whether it is happening at your firm. It almost certainly is. The question is whether you have a policy that addresses it.
Midnight Blue implements Microsoft 365 Copilot, which is one of the governed AI solutions this article discusses. We have a financial interest in how you think about this. What follows is the same conversation we have with clients in SBRs right now — including the parts that acknowledge Copilot is not the only answer and that some firms are not yet ready to deploy it. If your situation calls for a different approach, we will tell you that. The goal of this article is to help you understand the actual risk, not to sell you a specific product.
Your employees are using AI tools right now. A policy does not stop that — nothing stops that completely. What a policy does is define which tools are sanctioned, what data can and cannot be used in AI prompts, how AI-generated outputs are reviewed before they go to clients, and what the consequences are for using tools outside those boundaries. Without it, you are not avoiding AI risk. You are just not aware of the exposure you already have.
The term “shadow AI” sounds like something from an IT security briefing. The reality is much more ordinary, and that is what makes it hard to address.
Shadow AI is not a rogue employee trying to circumvent controls. It is a capable, well-intentioned employee using the fastest tool available to get their work done. In 2026, that tool is often a free or consumer-grade AI application, used without any organizational oversight.
Here are the patterns we see most often in the professional services environments we work with:
An employee uses a free AI tool to draft a client-facing document. To get a useful output, they paste in background information — client name, matter details, financial information, relevant case history. The AI tool generates a usable draft. The employee edits it, sends it, and moves on. The pasted data has now been submitted to a third-party service with its own data handling policies that the employee has not read and the firm has not reviewed.
Even if a firm has blocked certain AI tools on the corporate network, an employee picks up their personal phone, opens a free app, photographs or transcribes relevant information, and gets the answer they need. There is only so much that network controls can do. As Larry has noted in client conversations: even if your only approved product is Copilot, employees are still using ChatGPT on their personal phone and uploading sensitive documents — and not thinking twice about it.
AI tool usage spreads informally. One person on the team discovers a useful tool, mentions it to a colleague, and within a few weeks five people are using it. Nobody decided to adopt it. Nobody evaluated it. It simply became part of how work gets done, and by the time leadership is aware of it, months of data have already flowed through a system the firm has no governance over.
“There’s a lot of policy that needs to be in place. What’s the AI acceptable use policy? It doesn’t stop everything, but together it’s important.”
— Larry Schwartz, CEO & Founder, Midnight Blue Technology Services
For firms in regulated industries — legal, financial services, insurance, healthcare — the stakes are higher than data hygiene. Client confidentiality is a professional obligation, not just a preference. And shadow AI creates real exposure against that obligation.
When an employee submits client information to a consumer AI tool, that data may be used to train the AI model, stored on servers outside the firm’s control, or accessible to the service provider under the tool’s terms of service. Most free AI tools have terms that grant broad rights to submitted content. Most employees have not read those terms. Most firms have not either.
That is not a technical failure. It is a policy and awareness failure. And in a legal or financial context, it can constitute a breach of client confidentiality regardless of whether any harm results.
Even for governed AI tools like Microsoft 365 Copilot, the security exposure depends entirely on what your permissions structure looks like before you activate it. Copilot accesses everything your employees can access. If your Microsoft 365 permissions have never been properly structured, Copilot will surface things that should not be surfaced.
The specific example we use with clients: an employee goes into the HR area to look up their health benefits. They click further because the AI surfaces something nearby. Now they are in payroll data they technically should not have access to — not because Copilot failed, but because the permissions setup was never right, and the AI made the gap visible faster than anything else would have.
In regulated industries, you may be asked to demonstrate data handling practices during an audit, a client review, or a matter involving litigation. “We didn’t have a policy” is not a defense. It is an admission. A documented AI acceptable use policy, combined with evidence of employee communication and a governed technical environment, is what creates a defensible position.
This is where most conversations stall. Firms know they should have a policy. They are not sure what that means in practice. Here is the honest version.
A credible AI policy is not a blanket ban on AI tools, which does not work and drives usage underground. It is not a one-page statement that says “use AI responsibly,” which says nothing actionable. It is a document that gives employees clear, specific answers to the questions they are actually facing.
| What the Policy Addresses | What That Looks Like in Practice |
| Which AI tools are sanctioned for use | A named list of approved tools (e.g. Microsoft 365 Copilot for organizational data tasks) and a clear statement that unlisted tools require IT review before use with any client or business data |
| What data can and cannot go into an AI prompt | Specific categories: client names and matter details are restricted; general drafting and research without client-specific data may be permitted with approved tools. Employees need a yes/no answer, not a judgment call |
| How AI-generated outputs are reviewed before client delivery | Who reviews it, what they are checking for, and what the sign-off process looks like. Especially critical for client-facing documents in legal and financial contexts |
| Personal device and personal account usage | Clear guidance on whether employees can use personal AI accounts for work-related tasks and on what basis. Vague policies here are as good as no policy |
| Incident reporting | What to do if an employee realizes they submitted something they should not have. The goal is to know about it quickly, not to punish disclosure |
| Annual review cadence | AI tools and their capabilities change fast. A policy written in 2024 may not be current in 2026. Build in a review date |
The standard set of policies MBTS helps clients build out for AI starts with the AI acceptable use policy and the incident response plan as the two most immediately necessary documents. From there, depending on the firm’s size and regulatory environment, it typically expands to cover BYOD, data handling, and remote access in the context of AI usage.
None of this is exotic. Most of it is the same policy work that was already overdue for social media, email, and cloud storage. AI just made it urgent.
I wrote another article that shows the full cost picture for Microsoft 365 Copilot deployment, including the governance and permissions work that makes secure adoption possible.
What Does Strategic AI Implementation Actually Cost? (And What Does It Save?) →
The Endless Customers framework requires this section, and it is here because it is true: there are situations where addressing AI governance is not the right immediate priority.
If your firm is in the middle of a significant operational change — a merger, a leadership transition, a major platform migration — adding AI policy work to an already-stretched team compounds the friction without improving the outcome. Change management takes real organizational bandwidth. If that bandwidth is already spoken for, prioritize the change that is already in motion.
If your employees are genuinely not using AI tools in their daily work and there is no indication that is changing, the urgency is lower. This is rare in professional services in 2026, but it is not impossible. If it describes your firm, a lightweight AI use statement in your existing acceptable use policy may be sufficient for now.
If your firm is subject to specific regulatory guidance on AI that has not yet been finalized, waiting for that guidance before writing a policy that may need immediate revision is a reasonable position. This applies primarily to firms in highly regulated subsectors where regulators are actively writing new rules — not to most general professional services practices.
What does not qualify as a reason to wait: not having a formal AI strategy yet, not having decided which AI tools to adopt, or not having budget for a full Copilot deployment. Those are reasons to address AI governance with lighter-weight policy work now, not reasons to defer it entirely.
For most professional services firms in the MBTS ICP, the most useful first move is a conversation that covers three things: what AI tools are currently being used across the firm (officially and unofficially), what your Microsoft 365 permissions structure looks like, and whether you have an AI acceptable use policy in place.
Those three questions produce a clear picture of your actual exposure. Most firms that go through that conversation find at least one thing they did not know — either a tool being used that nobody officially adopted, a permissions gap that creates a Copilot risk, or a policy they thought they had but that is either missing or three years out of date.
From there the path is straightforward. Policy work can move ahead of any technology decision — you do not need to have chosen Copilot to write an AI acceptable use policy that covers your current tools and sets the standard for any future adoption. And the permissions and baseline security work that makes Copilot safe to deploy is work worth doing regardless of whether you ever activate Copilot, because it closes the exposure that already exists.
Larry Schwartz and Julie Hodges, a Copilot Expert from Microsoft, are hosting a free webinar covering AI governance, security, and what secure Copilot deployment looks like for professional services firms. There is time for direct Q&A on your specific situation.
Tuesday, June 16, 2026 | 11:00 AM EST | Microsoft Teams (Live + On-Demand Recording)
| Question | Answer |
| What is shadow AI and why does it matter for my firm? | Shadow AI refers to AI tools being used by employees without organizational oversight or approval. In most professional services firms, this means employees using free or consumer-grade AI applications — like free versions of ChatGPT or similar tools — to draft documents, summarize information, or prepare for client meetings. It matters because client data submitted to those tools may be handled under terms your firm has never reviewed, creating potential confidentiality exposure and compliance risk, particularly in regulated industries like legal, financial services, and healthcare. |
| Does blocking ChatGPT on the corporate network solve the shadow AI problem? | Partially. Network-level controls prevent use on corporate devices connected to the corporate network. They do not prevent use on personal devices, personal network connections, or third-party tools that are not on the blocked list. Several of our clients have chosen to block ChatGPT at the network level as a near-term measure while they develop a broader AI policy. That is a reasonable step, but it addresses only part of the exposure. A documented AI acceptable use policy, communicated to employees with clear guidance on what is and is not permitted, is what creates a defensible position. |
| What should an AI acceptable use policy include? | At minimum: which AI tools are sanctioned for use with organizational or client data, what categories of data may and may not be used in AI prompts, how AI-generated outputs are reviewed before client delivery, what employees should do if they believe they have submitted data they should not have, and guidance on personal device and personal account usage for work-related tasks. The policy should be reviewed annually at minimum — AI tools and their capabilities change fast enough that a policy written two years ago may not reflect current realities. |
| Is Microsoft 365 Copilot safe to use without a governance review first? | Copilot operates within your existing Microsoft 365 security and compliance framework, which is a meaningful safeguard compared to consumer AI tools. However, Copilot accesses everything your employees can access. If your Microsoft 365 permissions have gaps — files accessible to more people than intended, sharing settings that were never reviewed, data that should be restricted but is not — Copilot surfaces those gaps faster than most other tools. A permissions and baseline security review before broad activation is not optional for firms with any meaningful compliance exposure. It is what separates a governed deployment from an ungoverned one. |
| How quickly can an AI policy be put in place? | A basic AI acceptable use policy can be drafted and communicated to staff in a matter of weeks — it does not require a lengthy project or a formal AI strategy already in place. The more involved work is the baseline security and permissions review that makes governed AI tools like Copilot safe to deploy broadly. That typically takes two to three weeks for a professional services firm in our ICP. Neither of these requires you to have already decided which AI tools to adopt or commit to a specific platform. |
| What is the difference between an AI policy and AI governance? | An AI policy is the document: the written rules about what employees can and cannot do with AI tools, what data can be used, and how outputs are managed. AI governance is the broader practice of making sure those rules are enforced, updated, and aligned with your firm’s compliance obligations and technology environment. Governance includes the technical controls (permissions, data access configurations), the policy documentation, the employee communication, and the review cadence. You need both. The policy alone, without the technical foundation, is incomplete. The technical controls alone, without employee awareness and documented rules, are also incomplete. |