Dark AI & Dark Agents
Your people are already using AI at work. The real question is whether anyone knows which tools, with what data, and who said it was OK.
What We Mean
"Dark AI" (often called shadow AI) is any AI tool used for company work that the company hasn't approved or doesn't know about. It is rarely malicious. It's usually a good employee trying to get through their inbox faster.
Dark AI
AI tools people use without approval. They answer questions and write drafts.
- Personal ChatGPT, Gemini or Claude accounts used for work tasks
- Customer emails, pricing or drawings pasted into a free tier
- AI note-takers that auto-join calls and store the recording somewhere else
- Browser extensions that read every page you open
- AI features quietly switched on inside software you already pay for
Dark Agents
AI agents that take actions, not just answer questions, set up without anyone keeping track of them.
- An agent connected to someone's email, drive or calendar with their login
- Automations that send, file, delete or update records on a schedule
- Agents that read incoming documents, which can carry hidden instructions
- Workflows that keep running after the person who built them leaves
- No inventory of which agents exist, what they can touch, or who owns them
A dark tool might leak information. A dark agent can also do things with your accounts. That's why agents need rules before they spread.
Why It Happens
It's almost never a discipline problem. People find a tool that saves them an hour a day. There's no approved alternative and no written rule, and asking feels like admitting you're behind. So they keep quiet and keep using it.
Leadership often stalls for an understandable reason: it's uncomfortable to sign off rules for something you haven't used yourself. So no decision gets made, and the unofficial use keeps growing.
Banning AI doesn't stop it. It just pushes it further out of sight.
The Real Risks
Data Leaving the Building
Free and personal AI accounts may keep your inputs, and some use them to train their models. Once it's pasted, you can't take it back.
A quality engineer pastes a customer's drawing and spec sheet into a personal chatbot to summarize it.
Agents With the Keys
An agent logged in as an employee can send, share and delete just like that employee can. Hidden instructions in a document or email can steer it.
An inbox agent "helpfully" forwards a supplier's pricing thread to the wrong address.
Contracts & Compliance
NDAs, customer quality agreements, ITAR/EAR, HIPAA and privacy laws all say where data may go. Most of them predate AI, but they still apply.
A customer audit asks, "Which third parties process our data?" Nobody can answer.
Unchecked Output
AI drafts read well even when they're wrong. Without a "human checks it" rule, made-up numbers end up in reports and specs.
A summarized work instruction quietly drops a torque value.
Knowledge Walks Out
The best workflows often live in someone's personal account. When that person leaves, the prompts, agents and history leave with them.
A reporting automation stops the week its builder changes jobs.
Wasted Spend
Five departments, five overlapping subscriptions, nobody comparing notes. Unofficial use means you can't measure the value either.
The company pays for Copilot while half the team expenses ChatGPT Plus.
Dark AI Exposure Check
Eight questions about your organization. Answer honestly. Nobody's watching.
What a Sensible AI Use Policy Covers
A one-page policy people actually read does more than a 30-page one nobody opens. It should answer these questions:
- Approved toolsWhich AI tools are OK for work, on company accounts, with model training on your data turned off.
- Data classesGreen, yellow and red information, and what may go into which tool. Customer data and drawings are not "green."
- Agent rulesWhat an agent may connect to, which actions need a human to approve them (send, delete, pay), and a list of every agent you run.
- Human in the loopWho checks AI output before it reaches a customer, a spec, or a decision.
- An amnesty to startAsk everyone what they're already using, with no penalty. You can't govern what you can't see.
- A path to "yes"A quick way to request a new tool, so the official route is easier than the hidden one.
- ChampionsOne or two people per team who help colleagues and bring good ideas forward.
- A review dateAI tools change monthly. Review the approved list every quarter.
Bringing AI Use Into the Open
Anonymous Pre-Survey
Before the workshop, staff tell us which tools they use and for what, without their names attached. You get the real picture, often for the first time.
Hands-On Privacy Session
Attendees see exactly what happens to data in free and paid tiers, and practice getting the same results with approved tools.
Leave With a Draft Policy
The Deployment Workshop ends with an approved tool shortlist, data classes and agent rules that leadership can sign off with confidence.
Find out what's already running in your business
Talk through a workshop with a dark AI survey and policy session built in.