3 · build · 4-to-6-week pilot
Custom AI agent development, delivered ready to run
An agent, here, is not a chatbot. It is a program that reads your messages and documents, decides on an action and carries it out in your tools, under written rules. Building one takes a few weeks; making it sustainable is the real craft.
Proven starting points
Six types of agent already running at NES Studio
Each one was built, put into service and corrected on NES Studio's own operations before being offered.
Email triage and replies
Sorting incoming messages, extracting the useful information, drafting a reply, sending only after human approval.
Document reading
Invoices, contracts, delivery notes, service books: extraction into structured data, with consistency checks and a history per file.
Internal knowledge base
Your documentation, procedures and decisions, searchable by teams and by agents, and kept up to date with approval.
Content production
Monitoring, drafting, proofreading, visuals, publication: an editorial pipeline where a human approves before release.
Clean-up and governance
Drive, email, shared folders: a plan proposed by the agent, approved, then carried out. Nothing is deleted without consent.
Operations monitoring
For an IT department: daily checks on services, backups, certificates and costs, alerting only on anomalies, with a diagnosis attached.
Process
From scoping to go-live
- wk 1
Scoping
A written scope of what the agent may do and what it will never do. Each action ranked by reversibility. Success indicators and a running budget.
- wk 2–4
Parallel pilot
The agent works on your real data, but proposes without acting; the human keeps doing the work. The gap between the two is measured before anything is entrusted to the agent.
- wk 4–5
Go-live
Gradual activation, one action at a time. Human stop points on anything irreversible, a readable log, daily checks, tracking of costs and failures.
- wk 6
Handover
Documentation, shutdown procedure, training for the people who will work alongside the agent. The code and configuration sit in your repository.
- then
Operation, if you wish
In-house with your trained teams, or delegated to NES Studio with a monthly report.
Technical choices
The model serves the task
The AI model is chosen task by task: a fast, low-cost model to sort and check, a more powerful one to draft or reason. Connections to your tools use standard interfaces (including the MCP protocol), so you are not locked in to one vendor.
Hosting on your systems or in France, depending on how sensitive the data is. Only the necessary fragments pass through the model.
What is never delivered: an agent that sends or deletes without approval, an agent without a log, a system whose code you do not own.
Frequently asked questions
Before you start a project
How long until a first agent is in service?
Four to six weeks for a pilot on a well-defined process, including two to three weeks running in parallel with the human.
Who owns the code?
You do. Code, configuration and documentation are delivered into your repository.
Do we need a technical team in-house?
Not to get started. What you do need is a process owner who approves the agent's actions during the pilot.
What happens if the agent makes a mistake?
Anything irreversible waits for human approval, and every action is logged. A mistake is visible, gets corrected, and produces a check so that it does not happen again.
A process that is costing you time?
Describe it in three lines: you will receive an initial assessment and an estimate.