Most Nigerian companies have a reasonable picture of their obvious data risks. They know where the customer database lives, who has access to the accounting system, and which laptops carry client files. What almost none of them track is the steady stream of company information flowing out through a browser tab: staff pasting documents into personal AI accounts to get their work done faster.
This is shadow AI. It is not malicious, it is not rare, and it is probably happening in your organisation today. This article explains how it works, why it is hard to see, why banning it tends to fail, and what actually reduces the risk. It is part of our series on why Nigerian companies need private AI.
What shadow AI looks like in practice
Shadow AI rarely looks like a security incident. It looks like a productive employee. The following is an illustrative scenario, not a real case, but anyone who has managed an office will recognise it.
Illustrative scenario: A relationship officer at a mid-sized financial services firm in Lagos has a quarterly review due. She exports a client's transaction history to a spreadsheet, pastes it into her personal ChatGPT account and asks for a summary of spending patterns and three talking points. The output is good, the meeting goes well, and her manager praises the preparation. Nobody asks how it was done. The client's name, account activity and the firm's commentary now sit in a chat history on a personal account the firm has never seen, retained under consumer terms the firm never agreed to.
Multiply that by every department. HR drafts disciplinary letters with the employee's details included. Legal asks for a clause comparison across two client contracts. Finance uploads a management accounts pack to get commentary written. Sales pastes a full proposal, with pricing, to improve the tone. Each act is sensible on its own. Together, they amount to an unmanaged data flow.
Where the data actually goes
It helps to be precise about the routes, because they differ in risk.
| Shadow AI behaviour | What leaves the company | Company visibility | Sanctioned alternative |
|---|---|---|---|
| Pasting text into a personal chatbot account | Whatever was pasted, plus the conversation history | None | Private AI assistant on the office LAN |
| Uploading whole files (PDF, Excel, Word) | The entire document, often far more than the task needed | None | Private AI with a controlled document workspace |
| AI browser extensions that read web pages and email | Potentially everything visible in the browser | Low | Managed browser policy plus private AI |
| AI meeting note-takers on personal accounts | Full audio and transcripts of internal or client meetings | Low | Approved tool under a company agreement, or local transcription |
| AI apps on personal phones over mobile data | Photos of documents, voice notes, pasted text | None, bypasses the office network entirely | Private AI reachable over the company VPN |
Three points are worth stressing. First, the risk is not that a cloud provider is reading your files out of curiosity. Reputable providers have strong security. The risk is that the data sits outside any agreement your company controls, in accounts your company cannot audit, close or delete when the employee leaves. Second, consumer and business products are treated differently. Business and enterprise plans generally come with contractual commitments on data use and retention; personal accounts generally come with settings the individual controls. Third, once data is in someone's personal chat history, offboarding does not remove it.
Why nobody is tracking it
Shadow AI slips past the normal controls for structural reasons:
- No procurement trail · Free tiers and personal subscriptions never touch the finance department, so there is no vendor record and no contract review.
- Encrypted traffic · To a firewall, a chatbot session looks like any other HTTPS connection. Without specific filtering, IT sees a domain at most, not what was sent.
- Mobile data · In Nigeria especially, many staff work from phones on their own data bundles, completely outside the office network.
- It looks like good work · Managers see better output, faster. There is no incentive to ask how, and staff have no incentive to volunteer it.
- No policy to break · Many companies have no written AI policy at all, so staff are not technically breaking a rule.
Why bans fail
The instinctive response is to block AI domains on the office network and circulate a memo. It is understandable, and it mostly does not work. The productivity benefit is real, so staff route around the block, usually by switching to their phones. You lose what little visibility you had, and the data flow continues on devices you manage even less.
The honest lesson from companies that have tackled this well is that shadow IT of any kind shrinks when the official option is better than the unofficial one. You do not stop people using AI. You give them an approved AI that is good enough, easier to reach and clearly safe to use with company data.
What works instead
1. Measure before you act
Run a short anonymous survey: which AI tools do you use, for which tasks, and what kinds of data do you put in? Combine it with a look at DNS or firewall logs for the main AI service domains. Frame it as an exercise to provide better tools. You will learn which teams depend on AI most, which become your pilot group.
2. Write a one-page policy
Staff need a clear rule they can remember. A simple three-tier version works: public information may go to approved cloud tools; internal information goes to approved tools only; client, personal, financial and legal data goes only to the company's private AI. Our piece on the NDPA 2023 and generative AI explains why personal data deserves the strictest tier.
3. Provide a sanctioned private AI
This is the step that makes the policy stick. A private AI server on the office network, running open models through a ChatGPT-style web interface, gives staff the convenience they were getting from personal accounts, with company logins, company retention rules and data that never leaves the building. The practical setup is covered in How to Give Every Employee a Private AI Assistant, and the ability to chat with company files privately is often what convinces the heaviest shadow AI users to switch.
4. Keep approved cloud for what it does best
Be candid with staff: frontier cloud models are stronger on the hardest tasks. A hybrid policy that keeps an approved cloud tool, under a business agreement, for non-sensitive work removes the temptation to sneak back to a personal account.
5. Make offboarding include AI
Add a line to the exit checklist: confirm company data has been removed from any personal AI accounts. It will not be perfect, but it sets the expectation.
What a private AI server involves
The hardware is less exotic than people expect. A single GPU server, such as one from our AI Series, sits in a server room or secure cabinet with a UPS and connects to the office network. A pilot-sized machine handles roughly 5–15 light concurrent users of 7–14B models; a department-scale machine handles roughly 15–50 staff on quantised 14–32B models. For the detail, see Sizing an Office AI Server for 10, 50 and 200 Staff and our explainer on AI inference servers for business.
The bottom line
Shadow AI is a symptom of something positive: your staff have found a tool that makes them better at their jobs. Treat it that way. Measure it, write a simple rule, and give people a sanctioned tool that is good enough that they stop reaching for their personal accounts. The data leak nobody is tracking becomes a capability the company owns.
Sephora Systems builds and installs private AI servers for Nigerian companies, from Abuja with nationwide delivery. Book a private AI consultation to talk through your current AI usage and what a sanctioned alternative would look like. You can also ask Kitan, our site assistant, or WhatsApp us on +234 707 096 6669.