Not a chatbot bolted onto your files. It is a private workbench connected to your knowledge, workflows, records, media and approvals — so it can answer questions, write drafts, watch for problems, and handle routine follow-up. Always showing where its answers came from, and always with one of your people in charge.
Your organisation’s knowledge — contracts, operating systems, regulator letters, board papers, site reports, policies, photos and project records — is connected and remembered. Ask a question in plain language and you get an answer in seconds, with the exact source it came from. Ask for a report and it drafts one, pulling the evidence straight from your own records.
It also keeps watch. Deadlines, targets, risks, a supplier going quiet — it flags them to the right person before they become a problem. And before anything leaves your business, one of your people reviews and approves it.
The clever parts underneath are the same for every client. What changes is your sources, your rules, and what you ask it to do.
Every answer points to the exact source it came from, so you can open it and check for yourself. If it can’t find a solid answer in your approved knowledge, it says so — it doesn’t guess or make things up.
Everything runs inside your own country, on systems you control. Your data and records are never sent abroad, and they are never used to train anyone else’s AI. Fully in line with Nigeria’s data-protection laws (NDPR and NDPA 2023), and the equivalent elsewhere.
Every question, draft, approval and message is saved in a permanent record that can’t be quietly changed. If an auditor, a regulator, or your board ever asks what happened, you can show them exactly — down to the minute.
The AI does the work; one of your named staff approves before anything goes out. The AI is the tool, never the one who decides. Every outgoing item carries a real person’s name.
Ask anything about your business in plain language — “what’s the status of the Ijebu project?” — and get the answer in seconds, with the source attached.
Finds answersBoard papers, donor reports, audit responses, briefings — drafted from your own records, every figure traceable, ready for someone to read and sign.
Drafts for reviewKeeps an eye on deadlines, targets and risks around the clock, and tells the right person — through the channel they actually check — before something slips.
Watches and alertsReads site photos, drone footage and CCTV like a trained inspector: safety gear not worn, asset corrosion or damage, construction progress, and perimeter incidents — flagged for your engineer to confirm.
Reads photos and videoSpots the partners, suppliers or counterparties you should be talking to, and drafts the first message — ready to send from your own email.
Finds and contactsThe safety net underneath everything: nothing leaves without a person’s approval, and every action is recorded.
Governs everythingYou don’t need to know any of this — but for the engineers and technical evaluators on your team, here is how the workbench actually finds and grounds its answers. It is the difference between a serious retrieval system and a chatbot pointed at a folder.
Every question runs two searches at once — keyword (BM25) for the exact terms and semantic vector search for meaning — then merges them with reciprocal rank fusion. It finds the right passage whether your team recalls the exact wording or only the gist.
Before indexing, each passage is enriched with the context around it — which document, which section, which matter — so a lone figure in a table or a single clause in a 200-page contract is still found correctly on its own.
Answers and drafts are written only from the passages actually retrieved from your records, every statement tied to its source. Nothing is invented from the model’s general knowledge.
If the evidence isn’t in your approved knowledge, the workbench says so plainly instead of producing a confident-sounding fabrication. Refusal is a feature, not a failure.
Your documents are chunked, embedded and stored as a search index inside your own environment. The vector index and the text behind it never leave your control or cross a border.
Scanned PDFs, tables, regulator letters, board packs, site photos and CCTV — handled together and re-indexed as new material is added, so answers stay current without a manual rebuild.
An AI on its own can only talk. To be useful it has to reach your real systems — your document store, your calendar, your email, your finance system, a government registry — to look things up and to do things. A connector is the secure adapter that lets it. Think of it as a labelled, lockable socket: the workbench can only use the sockets you give it, only in the ways you allow, and every use is written to the record.
We build these connectors for your stack. One might let the workbench search your contract archive; another might let it file a regulatory return, post an update to your project tool, pull a counterparty from a public registry, or book a meeting and send the calendar invite — which is exactly what the assistant on our home page does when you ask it to set up a call.
Because they follow the MCP standard, each connector is a clean, swappable part: new ones can be added later without rebuilding the workbench, and you decide what every connector may touch. Nothing acts without the permissions and the human sign-off from the four commitments above.
We sit with your team, find the one thing that would help most, and write a clear plan with a fixed price. $25–75K.
We connect the sources, systems and workflows that show how your organisation actually works.
We build and set up your AI, with the safety checks and access rules your business needs. $150–500K.
We train your team and hand it over. You own it.
Updates, fixes and new features — or you run it yourself. $50–250K / year.
Cleo, the design agent on the home page, sketches a private AI workbench for your business in a few minutes and emails the plan to you. It costs nothing, and commits you to nothing.
Talk to Cleo →