Walk through almost any Nigerian office today and you will find someone with ChatGPT, Claude or Gemini open in a browser tab. They are drafting letters, cleaning up proposals, summarising board packs, analysing sales spreadsheets and asking for help with contract clauses. Most of them are using personal accounts. Many of them are pasting in client names, salary figures, bank statements and confidential agreements, because that is what makes the answers useful. Nobody approved it, nobody is logging it, and in most companies nobody in management has asked.
This is not a scare story. Cloud AI is excellent, and your staff are using it because it makes them faster. The question is not whether your company should use AI. It already does. The question is whether the company controls where its data goes when it does. This guide is the hub for our private AI series: what private AI is, what it is not, how it is built, what it costs, and where to read more on each piece.
The problem: AI adoption arrived before AI policy
In most organisations, generative AI did not arrive through procurement. It arrived through individual staff signing up with a personal email, sometimes paying for a subscription on a personal card, and quietly getting their work done faster. That pattern has a name, shadow AI, and we cover it in detail in Shadow AI in Nigerian Workplaces: The Data Leak Nobody Is Tracking.
The honest summary is this. When an employee pastes a document into a consumer AI account, that content is processed on servers abroad, under terms of service the company never signed, with retention and training settings the company cannot see. For a lot of content that is fine. For client data, personal data, contracts and financials, it is a governance gap, and for some industries it is a regulatory one. Our article on the NDPA 2023 and generative AI walks through what the Nigeria Data Protection Act means for this, carefully and without legal theatrics.
What private AI actually is
Private AI means giving your staff a ChatGPT-style assistant that runs entirely on hardware your company owns, inside your building, on your own network. In practice it has four parts:
- An AI server · A GPU-equipped machine, such as one of our AI Series builds, installed in your server room or a secure cabinet.
- Open models · Open-weight language models such as Llama, Qwen, Mistral or Gemma, downloaded once and run locally. No data is sent to the model's creator.
- An inference engine · Software such as Ollama for simplicity or vLLM for higher concurrency, which loads the model onto the GPU and answers requests.
- A private web interface · A browser-based chat front end such as Open WebUI, served over the office LAN, with company logins, chat history and optional document collections.
Staff open a web address on the office network, log in, and use it the way they already use ChatGPT. The difference is that every prompt, every uploaded file and every answer stays in the building. For the step-by-step setup, see How to Give Every Employee a Private AI Assistant (Without Sending Data Abroad).
Chatting with your own documents
The feature that usually wins over management is retrieval-augmented generation, or RAG. You load the company's policies, templates, past proposals, product manuals or case files into a private knowledge base, and the assistant answers questions by citing them. "What is our leave policy for contract staff?" gets an answer from your actual HR handbook, not a generic guess. We explain how this works, and where it goes wrong, in Chat With Your Company Files Privately: Local RAG for Nigerian Businesses.
What private AI is not: the candid part
We build these systems, so it would be easy to oversell them. We will not. Here is where open models running on a single on-site server fall short:
- Hardest reasoning tasks · The leading frontier cloud models remain ahead on complex multi-step reasoning, advanced mathematics and difficult coding. A local 14B or 32B model will not match them there.
- Very long documents · Local models can handle long inputs, but very long contexts consume GPU memory quickly and slow everything down for other users.
- Bundled extras · Cloud products bundle web browsing, image generation and voice in one subscription. A private stack can add some of these, but each one is an extra component to set up and maintain.
For drafting, summarising, rewriting, translation, extraction, classification, and question-answering over your own documents, the gap is much smaller, and for many day-to-day office tasks staff will not notice it. That is why the policy we usually recommend is hybrid: sensitive data goes to private AI, approved cloud tools under a proper business agreement handle the rest.
| Type of work | Where it should go | Why |
|---|---|---|
| Client files, contracts, HR records, financials, board papers | Private AI | Data stays on company hardware under company control |
| Questions about internal policies and documents | Private AI with RAG | The knowledge base lives on your server, not a vendor's |
| Public research, marketing copy without client data | Approved cloud AI | Frontier capability, no sensitive data exposed |
| Hardest reasoning or coding on non-sensitive material | Approved cloud AI | Frontier models still lead on these tasks |
| Anything in doubt | Private AI | The safe default costs nothing extra once the server exists |
Why control matters more than fear
The case for private AI rests on three practical points, not on alarm about cloud providers.
- Control · You decide which models run, who can use them, what is logged, how long chats are retained and when everything is deleted. Nothing changes because a vendor updated its terms.
- Data residency · Prompts and documents are processed in Nigeria, on your premises. That makes conversations with clients, auditors and regulators far simpler, which matters most in sectors we cover in Private AI for Banks, Law Firms and Hospitals in Nigeria.
- Cost predictability · Per-seat AI subscriptions are billed in dollars and scale with headcount and the exchange rate. An owned server is a one-time naira purchase plus electricity. We do the full comparison in ChatGPT Team Subscriptions vs an Owned AI Server: The 3-Year Naira Maths.
For the wider strategic argument about owning compute, our earlier piece on on-premise AI compute versus cloud remains relevant, and data privacy and on-premise AI covers the privacy side for teams training their own models.
What it takes: hardware and rough sizing
The server is the centre of the system. What determines its size is how many staff will use it at once and how large a model you want to run. The table below shows our AI Series tiers with rough guidance for private AI. Treat the user ranges as a starting point; real capacity depends on how heavily people use it, document sizes and the model chosen.
| AI Series tier | Core specification | Price (inc. VAT) | Rough fit for private AI |
|---|---|---|---|
| AI Research | Core i7-14700K, 64GB DDR5, RTX 4070 Ti Super (16GB VRAM) | ₦8,800,000 | Pilot or small team: roughly 5–15 light concurrent users of 7–14B models |
| AI Professional | Core i9-14900K, 128GB DDR5, RTX 4090 (24GB VRAM) | ₦18,600,000 | Department scale: roughly 15–50 staff on 14–32B models, quantised |
| AI Lab | Threadripper or Xeon W, 256GB ECC, 2× RTX 4090 or RTX A6000, 1600W redundant PSU, server chassis | From ₦25,000,000 | Company-wide, 70B-class models; consultation only |
A pilot on the AI Research tier at ₦8,800,000 is the lowest-risk way to prove the idea with one department before committing to an AI Professional at ₦18,600,000 or an AI Lab from ₦25,000,000. For a detailed walk-through by headcount, read Sizing an Office AI Server: Users, Models and Hardware for 10, 50 and 200 Staff. For the underlying hardware principles, our guide to AI inference servers for business and Ollama and LM Studio hardware requirements go deeper.
Power is part of the design
In Nigeria, an AI server without proper power planning is a liability. Every installation we do includes a correctly sized UPS for clean switchover and safe shutdown, and we plan around your inverter, solar or generator so the assistant survives a normal day of grid interruptions. A GPU server drawing several hundred watts under load needs its backup sized for that, not for a desktop PC.
Not every company needs a server room
Branch offices, field teams, project sites and NGOs often need private AI that can move. We build compact AI servers into rugged transport cases with their own UPS, which we describe in The Portable AI Server: A Company Brain That Fits in a Flight Case.
How to start
The companies that get this right do not start with hardware. They start with a short audit of how staff already use AI, a one-page policy on what data goes where, and a pilot with one team. Then they size the server to the real usage they observed. We lay this out week by week in Rolling Out Private AI in Your Company: A 30-Day Plan With Sephora.
Your staff have already decided AI is useful. The decision left for management is whether the company's data keeps flowing to accounts it does not control, or into a system it owns.
Sephora Systems designs, builds, installs and supports private AI servers for Nigerian companies, from Abuja with nationwide delivery. Book a private AI consultation and we will help you size the right system and policy for your team. You can also ask Kitan, our site assistant, or WhatsApp us on +234 707 096 6669.