AI Governance Insight

Your Employees Are Already Using AI Without You. Here's What That's Costing You.

June 24, 2026  ·  12 min read  ·  By Ashwameth Ravilla

New to shadow AI governance? Start with our primer: AI Tools Your Team Is Already Using — And Why That's a Governance Problem.

A majority of office workers use AI tools without IT approval — including features inside software you already pay for. Here's what's actually at stake, and the minimum-viable governance plan any small business can build this week.

The Tuesday Afternoon Problem

It's a Tuesday afternoon. A well-meaning employee at a Maryland marketing firm pastes a client proposal into an AI writing tool to clean up the language. No malice. No awareness of risk. They experience it as editing — the same way they'd use spell-check. The prompt contains the client's name, their budget, their campaign strategy for the next quarter.

Three weeks later, a new enterprise prospect sends over a vendor questionnaire before signing a contract. Question 14: Does your organization have a documented AI acceptable use policy?

The deal stalls.

Here's what makes this scenario worth paying attention to: it's not hypothetical. It's a pattern. A pattern that plays out in professional services firms, healthcare-adjacent businesses, legal offices, and government contractors across Maryland every week — usually without anyone realizing it happened until a client asks, a contract clause surfaces, or an incident occurs.

You already have an unmonitored data pipeline running through your employees. This post names it precisely, shows you what's at stake, and gives you the minimum you need to close it — without an IT department, an enterprise security budget, or a six-month consulting engagement.

What Is Shadow AI — And Why Your Current Policies Don't Cover It?

This is where most shadow AI content gets it wrong: it focuses on rogue tools. Employees secretly downloading sketchy apps. Consumer AI accessed from a personal device.

That's not where most of your exposure lives.

Your real primary risk surface is the software you're already paying for — and the AI features being quietly added to it every quarter:

• Microsoft 365 Copilot accessing your SharePoint files, your emails, your Teams conversations
• Slack AI summarizing channels that contain client communications, deal terms, personnel matters
• Notion AI trained on your internal workspace — strategy docs, meeting notes, client deliverables
• Salesforce Einstein generating outputs from your CRM records — contact data, pipeline details, account history

Every one of these is shadow AI if your organization hasn't reviewed, approved, or set governance rules for it. Most haven't. And the vendors rolling out these features aren't waiting for you to catch up — new AI capabilities ship on their update schedule, not yours.

This is the approved-tool paradox: the tools your team trusts most are the ones creating the exposure you can't see.

What Employees Think Is Happening vs. What Actually Is

The data doesn't create the problem — it reveals how widespread this perception gap already is.

A majority of office workers — 71% in a 2025 survey of enterprise employees — use AI tools without IT approval (Reco AI, October 2025). 38% have shared confidential data with an AI tool without authorization — a figure drawn from a survey of 7,000 employees (CybSafe + NCA, late 2024). GenAI traffic surged 890% in 2024, and most organizations are still in the early stages of forming any governance approach (Palo Alto Networks, 2025). Among those using AI, 58% are doing it daily (KPMG, 2025).

And the dominant platform? OpenAI accounts for 53% of all shadow AI usage across enterprises — more than the next nine platforms combined (Reco AI, October 2025). Which means your exposure isn't spread evenly across a dozen obscure tools. It's concentrated in the most capable, most widely distributed consumer AI product ever built — used daily, at scale, by people who think they're just getting work done faster.

That's not a training failure. It's a design reality. Consumer AI tools are built to feel like productivity features, not data pipelines. The friction is intentionally low. The visibility into what happens to the data is intentionally minimal.

Why Small Businesses Are the Most Exposed

Three factors make small businesses disproportionately vulnerable:

No one is watching the tooling.

In a large enterprise, IT teams monitor application usage, flag new SaaS adoptions, and review AI features before they're enabled at scale. In a 10-person firm, the new Copilot feature ships silently into everyone's Microsoft 365 account on a Tuesday — and by Thursday it's being used on client work, with no review, no approval, and no awareness that anything changed.

No contracts cover the AI layer.

Most small businesses have never executed a data processing agreement (DPA) for an AI tool. They accepted the vendor's standard terms at sign-up and moved on. Those standard terms — for consumer-tier AI — typically give the vendor broad rights to use input data. If your employees are using consumer-tier tools on client data, your client's data is operating under terms your client never agreed to.

Client contracts are already being violated.

This is the piece that surprises founders most. Standard professional services agreements, vendor contracts, and SaaS customer agreements routinely include data handling obligations: data must not be shared with unauthorized third parties; personal data must be processed only under compliant terms; confidential information must not be disclosed. An employee pasting a client's strategy document into an unapproved AI tool may have already breached all three — before anyone knew shadow AI was a category worth thinking about.

For Maryland businesses specifically, MODPA enforcement is now live as of April 2026. If employees are feeding personal data about Maryland consumers into unapproved AI tools — even incidentally, even while doing something else entirely — that's a potential processor obligation that your current vendor stack doesn't address. The law doesn't distinguish between intentional data collection and unintentional AI ingestion. The data moved. That's what matters. (For the full MODPA picture, see our MODPA enforcement deep-dive.)

What's Actually at Risk: The Chain of Liability

Walk through that chain slowly, because each step is where a different kind of damage originates:

Employee action: A team member uses an AI tool on client data. Routine. Intended to be helpful. Entirely invisible to leadership.

AI tool ingestion: The tool processes the prompt under its own terms of service — not yours, not your client's. If it's a consumer-tier tool, those terms may permit using your input to improve the model. The data is now outside your control.

Vendor model behavior: What the vendor does with that data is governed by a contract your employee clicked through at sign-up, not a DPA you negotiated. You have no audit rights, no deletion mechanism, no visibility.

Contract breach: Your client agreement says their data stays confidential and isn't shared with unauthorized third parties. It just was. Whether anyone ever finds out is a separate question. The breach occurred the moment the prompt was submitted.

Client notification obligation: Depending on your contract language and the nature of the data, you may have a legal obligation to notify your client. Most small businesses discover this clause for the first time during an incident, not before.

Regulatory exposure: If personal data was involved — employee data, client PII, health-adjacent information — MODPA, GDPR (for any EU-connected clients), or HIPAA (for health-related businesses) may now apply. Regulators don't require evidence of harm. They require evidence of compliance.

On the financial side: vendor-reported averages suggest AI-related data exposure incidents cost hundreds of thousands of dollars in direct costs before legal fees, client notification, and contract penalties. One industry report puts the average at $670,000 for companies with high shadow AI exposure (Reco AI, October 2025). For context, IBM's 2024 Cost of a Data Breach Report puts the average breach at $4.45 million — suggesting the Reco figure represents smaller-scale incidents. For a 10-person firm, either figure is existential. The more relevant question isn't the dollar amount — it's what your current client contracts say about unauthorized data processing.

One more signal worth tracking: enterprise clients with EU operations are already embedding AI governance questions into vendor security questionnaires. The EU AI Act's high-risk provisions activate August 2026. Maryland SMBs selling to enterprise clients or government contractors are already feeling this pressure — even where direct EU Act obligations don't yet apply. Your procurement contact at a large client may ask before any regulator does.

The Minimum-Viable AI Governance Plan (Built for a 10-Person Team)

Here's what each step looks like in practice:

Step 1 — Take inventory (30 minutes / Founder or Ops)

Ask every person on your team — in a single Slack message or team meeting — to list every AI tool they use, including features inside existing tools. You will be surprised. What this looks like in practice: a shared spreadsheet with three columns: Tool name, What it's used for, Approved by whom. If the third column is blank for most rows, that's your baseline.

Step 2 — Write a one-page acceptable use policy (1–2 hours / Founder, with light legal review)

This does not need to be a 20-page policy document. One page. What's permitted, what requires approval, what's off-limits. What this looks like in practice: "Employees may use [list of approved tools] for [approved use types]. Any new AI tool or new AI feature in an existing tool requires a 15-minute review with [name] before first use. The following data types may never be entered into any AI tool without explicit approval: client PII, financial data, health information, proprietary strategy, personnel matters."

That's it. A policy that exists and has been communicated is infinitely more defensible than a perfect policy that's still in draft.

Step 3 — Set a data classification rule (15 minutes / Founder)

Define what data is off-limits for AI tools by default — no exceptions, no judgment calls required. What this looks like in practice: a one-line rule employees can memorize: "If it has a client's name on it, a dollar figure attached to it, or a person's health information in it — it doesn't go into AI without explicit approval." Simple enough to internalize. Clear enough to enforce.

Step 4 — Create a vendor review checkpoint (15 minutes per tool / Ops or IT contact)

Any new AI tool — or new AI feature in an existing tool — gets a 15-minute review before use. What this looks like in practice: four questions on a shared doc — (1) Who makes this tool and where is their privacy policy? (2) What data does it retain and for how long? (3) Is there an enterprise or zero-data-retention tier? (4) Does our use case involve any data types from our classification rule? If you can't answer all four in 15 minutes from the vendor's own documentation, that's your answer.

Step 5 — Define an incident trigger (30 minutes / Founder)

Decide in advance what counts as a shadow AI incident and who gets notified within 24 hours. What this looks like in practice: "If an employee uses an unapproved AI tool on data covered by our classification rule, or if a client asks whether their data has been used in AI tools, that triggers an immediate notification to [Founder name] and a review within 24 hours. If personal data was involved, we notify our legal contact within 48 hours."

For businesses that need to point clients or partners to a formal framework, these five steps map directly to NIST AI-600-1 — the Generative AI Risk Management Profile released July 2024. Steps 1–2 are Govern. Step 3 is Map. Step 4 is Measure. Step 5 is Manage. You don't need to read the full document. But if a client's procurement team ever asks whether you follow a recognized AI governance framework, you now have a defensible answer.

2026–2027: This Is Becoming a Requirement, Not a Best Practice

The direction is already clear. Connecticut's June 2025 privacy amendments require disclosure when personal data is used to train large language models — effective July 1, 2026. Maryland is watching. The patchwork of state-level AI disclosure requirements is tightening, and the threshold for "we didn't know we needed a policy" is shrinking fast.

The businesses that move now — that inventory their tools, write the one-page policy, and set the data classification rule — won't just be protected. They'll be the ones that close deals faster, answer procurement questionnaires confidently, and earn the trust that comes from being able to say: we have a governance program and here's what it covers.

That is not a compliance outcome. It is a competitive one.

Data Gap Disclosures

Sources

• Reco AI — 2025 State of Shadow AI Report (October 2025) — reco.ai
• CybSafe + National Cybersecurity Alliance — via Cloud Security Alliance (late 2024) — cloudsecurityalliance.org
• Palo Alto Networks — What Is Shadow AI? (2025) — paloaltonetworks.com
• KPMG — Shadow AI Is Already Here (2025) — kpmg.com
• Komprise 2025 IT Survey — via CIO.com (2025) — cio.com
• IBM — Cost of a Data Breach Report 2024 — ibm.com
• NIST — AI Risk Management Framework & Generative AI Profile (AI-600-1) (July 2024) — nist.gov
• Shumaker, Loop & Kendrick — Patchwork of Data Privacy Laws (February 2026) — shumaker.com
• Osano — Maryland Online Data Privacy Act (MODPA) (January 2026) — osano.com

By Ashwameth Ravilla | DARS LLC | All citations verified June 2026