Artificial intelligence
Useful AI: small, specialised, controlled
An assistant that answers fifteen precise questions about your documents well is worth more than a general-purpose agent that invents answers about everything.
The problems we address
What brings you here
The information exists but nobody can find it
Procedures, contracts, minutes: the answer is somewhere in eight hundred documents.
Support answers the same question ten times a day
Skilled time consumed by first-line requests that are already perfectly documented.
Periodic reports are written by hand
Extraction, formatting, commentary: several person-days a month for a repetitive deliverable.
There is a fear that AI will invent things or leak data
A legitimate fear. Without cited sources, access limits and a log, an agent cannot be deployed in a company.
Deliverables
What you receive
- Document-search assistants that cite their sources
- First-line customer-support agents with handover to a human
- Lead-qualification agents
- HR assistants for common internal questions
- Information extraction from documents (invoices, contracts, forms)
- Automatic document classification and indexing
- Assisted generation of periodic reports
- Internal copilots connected to the ERP, the CRM or the IoT platform
- Assisted analysis of datasets
- Automation of repetitive tasks with human validation
Benefits
What it concretely changes
Answers you can trace
Every answer cites the documents used. The user can check rather than trust blindly.
The scope is closed
The agent can only call the tools that have been declared, with typed parameters and limited rights.
A human stays in control
Any action with a real effect — sending, changing, committing — goes through explicit approval.
Cost stays under control
Token, loop and frequency limits per user, with spending tracked.
Features
What we can do in this area
- A knowledge base built from your documents, respecting access rights
- Semantic search filtered by department, project or confidentiality level
- Business tools exposed to the agent as typed functions
- Human approval workflow on sensitive actions
- Full log of questions, sources and tool calls
- Regular evaluation against a reference set of questions
- Explicit fallback: the agent says it does not know instead of inventing
- Web, mobile and messaging interfaces, or integration into your own tools
Our approach
How we go about it
A single use case
We start with one measurable case, with a reference set of questions defined together with you.
Preparing the sources
Cleaning, splitting and indexing the documents, with access rights carried into the index.
Evaluation before deployment
Rate of correct answers, rate of abstention, absence of leakage between scopes. The figures are shared with you.
Supervised rollout
Released to a pilot group, logging enabled, feedback collected.
Follow-up
Unanswered questions feed the knowledge base. Quality improves with use.
Architecture
How it is built
An enterprise agent is first a question of architecture and permissions, not of model. The layers below show where the control points sit.
Interface
Web, mobile, messaging, or integration into an existing tool.
Backend API
Authentication, per-user quotas, input validation, logging.
Orchestration
Splitting the request into steps, budget of tokens and calls, failure handling.
Model
A pinned version to guarantee stable behaviour, replaceable without rewriting the application.
Retrieval
A vector index filtered by permissions, with excerpts returned together with their source.
Tools
An allowlist of typed functions; everything else is refused by default.
Guardrails
Human approval, output filtering, access control and audit log.
Diagram
The full chain, layer by layer
Each layer has its own security and resilience constraints. The diagram scrolls horizontally on a small screen.
Technologies
Relevant technologies
- Python
- TypeScript
- PostgreSQL
- pgvector
- Redis
- Node.js
- FastAPI
- Docker
- OpenAPI
- Sentry
- Elasticsearch
Industries covered
The industries where we apply it
Frequently asked questions
Frequently asked questions
Artificial intelligence and business agents
Is our data used to train a model?
How do you avoid invented answers?
Can the agent act on our systems by itself?
Can the solution be hosted on our side?
Africa Tech Services
A project around “Artificial intelligence”?
A first thirty-minute conversation, with no commitment. We will tell you plainly whether we are the right partner — and if not, we will point you elsewhere.