Artificial intelligence
Artificial intelligence can now search, analyse, write, sort, explain, assist or automate part of the work. But it is only useful when it answers a real need.
We help you identify the right uses, test them on your own data, and integrate them into your information system with a level of control suited to your business.
Our principles
We start by watching what people are actually trying to do: save time, find information, cut errors, make better use of data or automate a repetitive task. That is the result we try to measure. The choice of technology comes afterwards.
An excellent model fed incomplete, badly organised or outdated information will rarely give a good result. So before comparing models we look at the documents and data available, their quality, their access rights and how they can be used.
AI can be wrong. So you have to decide what it may do on its own, what has to be checked, and which actions it must never take without approval. The more consequential the decision, the more explicit the control has to be.
Use cases
There is no single use for artificial intelligence. Here are a few situations where it can bring something concrete. Clearly not all of them suit every company: scoping exists precisely to identify which ones deserve exploring.
Ask a question in plain language across contracts, procedures, technical files, documentation or past records — and get straight to the passages that let you check the answer.
Draft minutes, summaries, tender responses, product sheets or other documents from your own templates and existing information. The user keeps final approval.
Read a request arriving by email or form, identify its subject, extract the useful information and route it to the right process.
Help people quickly find a procedure, a business rule or a piece of information that already exists somewhere in the company’s documentation.
Summarise large volumes of information, spot anomalies or trends, and prepare an analysis that a manager can then verify.
Help technical teams understand a codebase, prepare tests, write documentation, track down faults or speed up the modernisation of an existing application.
Turn scattered, heterogeneous documents into an organised, searchable body of information that both teams and AI tools can actually use.
Help teams understand what these tools can do, where they stop, how to frame a request — and above all the situations where checking remains necessary.
Add AI capabilities to an ERP, a CRM, a document system, an intranet portal or a business application without rebuilding the whole system.
How we proceed
An AI project still holds a lot of uncertainty at the start. Rather than building a complete solution straight away, we move forward gradually. At each step we answer one precise question before committing to the next.
We work with the people who really know the process or task involved. The aim is to understand what takes time, what causes errors and what people would like to improve.
We check that the necessary information exists, that it is reliable enough and that it can be used under the intended conditions. We then compare several technical approaches.
We build a first version on a deliberately limited scope. It uses real data and situations that reflect day-to-day work. The prototype must make results measurable, not just produce an impressive demonstration.
Once the value is demonstrated, it has to become a real service: available to the right people, secure, monitored and plugged into existing tools.
An AI system has to keep being watched after go-live. We monitor usage and errors, update data or models when needed, and gradually transfer control of the system to your teams.
Reference architecture
AI models move very fast. An application should therefore not depend entirely on one model, one supplier or one technology. As far as possible we separate the system’s functions so a single part can be replaced without rebuilding everything.
The technology behind each of these steps can change without forcing the whole project to start over.
How we can work together
You do not have to hand us every stage of the work. eletec can run a one-day scoping session, carry out a study or a prototype, support the move into production, or monitor a system that already exists. The durations given are orders of magnitude.
A day with the people concerned, to surface the possible uses, confront them with real constraints and decide which deserve to be studied.
No commitment to go further.
Check that a use case is genuinely feasible before committing to build it.
Test the system on a limited scope, with situations that reflect how it will actually be used.
Turn a validated prototype into a service that stands up to the company’s real conditions.
Watch how an existing system is used and how well it performs, update its sources and models, and support the teams.
Understand the uses of AI that actually help, learn to handle the tools, check their answers and know their limits.
Programme adapted to the business and to the participants’ level.
Security, control & compliance
The quality of an AI system is not just the quality of its answers. You also need to know what data it uses, who can access it, what it costs, what it is allowed to do and how to react when it gets things wrong. Those questions are dealt with from the design stage.
Decide which information may be sent to an outside service, which has to stay inside your infrastructure, and how access is controlled.
Identify the rules that apply to the project, given the data used, the people concerned and the intended use. Data-protection requirements and the European AI framework are built into scoping whenever the project is in scope.
Show sources where they exist, keep the information needed for follow-up, and let users flag or correct an answer.
Measure actual consumption and use a model suited to the task rather than always reaching for the most powerful one. Provide quotas, caching or usage limits where those help.
State explicitly which actions require approval before they run, especially when they carry financial, contractual or operational consequences.
Depending on the project’s constraints, use European hosting, private infrastructure or models running locally. Document the dependencies so that changing course stays possible.
Build reference cases to check answer quality regularly and detect any drop after the system changes.
Avoid tying a whole application to a single model or supplier. Decide from the design stage what must be replaceable and, where the business demands it, what has to keep working if the AI service becomes unavailable.
Frequent questions
Not necessarily.
The quality, relevance and organisation of the information usually matter more than its volume.
Some projects can start with a fairly limited set of documents or data, provided they represent the need properly.
It depends enormously on the use.
The number of requests, their size, the model used, the hosting and the volume of data processed can swing the cost considerably.
So we estimate running costs before go-live, and set up monitoring where it is needed.
Yes, when the need and the infrastructure allow it.
Models running locally can be particularly attractive for confidentiality, data control or certain intensive uses. Cloud services may be preferable in other situations.
We compare options on the actual use case, not on a generic ranking of models.
Yes — some tasks can be partly or largely automated.
The real question is which ones, with what level of reliability, and with what consequences for how the work is organised.
We favour the uses where automation frees up time, cuts repetitive work or helps people do their job better.
Nobody can guarantee that a generative model will never get something wrong.
The risk can, however, be reduced considerably: narrow its scope, give it identified sources, show those sources, test the answers and require human approval where the consequences matter.
So the right question is not only “can AI be wrong?” but also “what happens when it is?”
That is precisely why we favour modular architectures.
When the different functions are properly separated, changing model remains a technical change rather than a complete rebuild of the application.
First step
You do not need to arrive with a defined AI project. Tell us what you are trying to improve or what wastes your time today. A scoping workshop can look at what AI might contribute — and can equally conclude that another solution would be simpler or more appropriate. Either way, you leave with a clearer direction.
