Most companies that decide they need private AI then stall for months. Not because the technology is hard, but because nobody owns the project, nobody knows where to start, and every meeting reopens the question of whether to do it at all. A plan with dates fixes that. This one gets a company from "our staff are already using ChatGPT" to a working, measured private AI pilot in 30 days.
If you are still deciding whether private AI is right for you, begin with our hub article, Your Staff Are Already Using ChatGPT: Why Nigerian Companies Need Private AI. This article assumes you have decided to try it.
The 30 Days at a Glance
| Week | Goal | Sephora does | You do | Output |
|---|---|---|---|---|
| Week 1 (days 1–7) | Audit and scope | Discovery call, sizing questions, site and power check | Survey current AI use, choose pilot team, name a project owner | Pilot scope and tier decision |
| Week 2 (days 8–14) | Build and prepare | Build and burn-in test the server, prepare models and interface | Gather pilot documents, decide access groups, draft policy | Tested server, document set, policy draft |
| Week 3 (days 15–21) | Install and launch | On-site install, UPS and shutdown automation, accounts, collections | Approve policy, brief managers | Live pilot with trained users |
| Week 4 (days 22–30) | Use, measure, decide | Support, tuning, usage review | Use it daily, collect feedback | Results and a scale decision |
Week 1: Audit Current AI Use and Pick the Pilot
Find out what is already happening
Run a short, anonymous survey: which AI tools do staff use, how often, for what tasks, and have they ever pasted in client, staff or financial information? Make it anonymous and say clearly that nobody will be punished for honest answers, or you will get the answers people think you want. The results usually surprise leadership, and they show you which tasks the private AI must handle on day one. Our article on shadow AI in Nigerian workplaces explains what to look for.
Choose the pilot team
- Size · ten to fifteen people, enough to generate real usage, small enough to support closely.
- Workload · one or two departments with document-heavy work: HR, legal, finance, operations or customer service.
- Mix · include at least one sceptic. Their objections are your best design input.
- Owner · one named project owner with authority to make decisions within a day.
Decide the tier
For a pilot, sizing is rarely the hard part. Using the ranges in our office AI server sizing guide, a pilot of roughly 5–15 light concurrent users fits AI Research, and if you already know the pilot will become a 15–50 person department deployment, starting on AI Professional avoids a later migration. Sephora's AI Series tiers, inc. VAT:
| Tier | Price (inc. VAT) | Typical role in a rollout |
|---|---|---|
| AI Research | ₦8,800,000 | Pilot or small team, roughly 5–15 light concurrent users; 7–14B models |
| AI Professional | ₦18,600,000 | Department scale, roughly 15–50 staff; 14–32B models quantised |
| AI Lab | From ₦25,000,000 | Company-wide, 70B-class models; around 200 staff means AI Lab or multiple servers, by consultation |
If budget is the question, our 3-year subscription versus owned server comparison will help your finance team.
Week 2: Build the Server, Prepare the Documents
While we build and burn-in test the server in Abuja, loading the chosen open models (typically from the Llama, Qwen, Mistral or Gemma families) and the private chat interface, your side has three jobs.
- Gather the pilot documents · current policies, manuals, templates and reference material for the pilot departments. Remove superseded versions now.
- Decide access groups · which collections each group sees. Write it down as a simple table; it becomes the configuration.
- Draft the acceptable-use policy · see below.
The local RAG guide explains why document preparation matters more than any hardware decision.
What the acceptable-use policy should say
- What the private AI is for · anything involving client, staff, patient or financial information.
- What may go to approved cloud tools · public information and generic, non-sensitive work, if you allow it.
- What never goes to unapproved tools · named categories, written plainly.
- Checking outputs · staff remain responsible for anything they send or sign.
- Retention · how long chats are kept and who can review them.
If personal data is involved, have your data protection lead review the draft; our piece on NDPA 2023 and generative AI gives useful background, though it is not legal advice.
Week 3: Install, Protect the Power, Launch
We install on-site anywhere in Nigeria. The install covers the server's placement and ventilation, the office network connection, user accounts and groups, document collections, and the power protection plan: an online UPS sized to the server, inverter or generator backup, and automated clean shutdown so an extended outage never corrupts the document index. Skipping power planning is the most common way a promising pilot ends in a bad first week. See AI workstation power and cooling in Nigeria for the detail.
Training happens on launch day: a short hands-on session for the pilot team covering how to ask good questions, how to check cited sources, what the policy says and where to report problems. Managers get a separate briefing on what the pilot is measuring.
Week 4: Use It, Measure It, Decide
Agree the measures before launch so nobody moves the goalposts afterwards:
- Adoption · how many pilot users used it each week.
- Use cases · what they asked, by category, from the usage logs.
- Time saved · self-reported in a short week-four survey. Imperfect, but directionally useful.
- Behaviour change · whether pilot users report moving sensitive work off public tools.
- Quality · examples of good and poor answers, which drive tuning.
Be honest with the results. Open models trail the best frontier cloud models on the hardest tasks, and some users will notice. If the pilot shows a handful of power users genuinely need frontier models for non-sensitive work, a hybrid policy is the right answer, not a failure.
Why Pilots Stall, and How to Avoid It
- No owner · a committee cannot make the dozen small decisions a pilot needs each week. One person with authority can.
- Messy documents · if the pilot collections contain three versions of every policy, users meet wrong answers on day one and stop trusting the tool.
- No reason to switch · if staff can keep pasting files into public tools with no guidance, many will. The policy and the private tool have to arrive together.
- Silence after launch · a two-line weekly note to pilot users with a useful tip and a fix for last week's complaint keeps adoption from fading in week three.
After Day 30: Scaling
A successful pilot usually leads to one of three moves: adding more departments to the same server, moving up a tier, or, for larger organisations, an AI Lab or multi-server design scoped by consultation. For teams that work across sites or travel, the portable AI server adds a flight-case option. Because the chat interface, accounts and document collections carry across, scaling is an infrastructure change, not a restart.
Ready to start your 30 days? Book a private-AI consultation with Sephora Systems and we will begin with week one: a discovery call, sizing and a site check. We design, build, install and support private AI servers from Abuja and deliver nationwide. You can also ask Kitan, our site assistant, or message us on WhatsApp at +234 707 096 6669.