Surikai engineering, homeLet’s talkContact us
FREN

APPLIED AI · HR AND BUSINESS

Applying AI where it delivers business value.

We design AI solutions connected to your documents, data and processes: natural-language search, assistants or agents. HR is our area of specialisation, and the same approach applies to other business functions. The instructions given to the model, data preparation and the evaluation of results take the requirements of each use into account, including non-discrimination, confidentiality of personal data, GDPR and the AI Act.

Discuss your project The Hub project
Illustration of AI applied to business data

OUR APPROACH

AI applied to specific uses

The clearest gains often come from time-consuming tasks, for example:

  • Finding profiles in a CV database
  • Answering a question from internal policies and agreements
  • Preparing the review of applications
  • Matching job offers with profiles

We start from these uses and the documents actually available, then define with the project team the criteria that will be used to judge the result. Answers link back to their sources, and teams keep the evaluation and the decision.

The AI model is chosen from different providers, according to the use, the confidentiality required and the budget. We recommend European models, such as those from Mistral AI, and computing embeddings on a private server when volume and infrastructure allow. Changing the model also changes the assistant’s behaviour with identical instructions, so tests are run again after each change.

Explore the 100% European architecture

PROMPT ENGINEERING

Instructions designed for HR requirements

The quality, speed and cost of an assistant depend largely on the instructions given to the model. We write, test and maintain them like a specification, with rules adapted to HR uses.

  • Versioned, tested instructions

    Instructions are written, tested and kept like a specification, version after version.

    Details
    • The assistant’s role and the answer format
    • The requirement to cite sources
    • What to do when the information is not in the documents
    • Comparison of each version with the previous one on the same scenarios before it goes live

  • Non-discrimination clauses

    For recruitment or career management, the instructions ask the model to keep to the skills and experience relevant to the need.

    Details
    • Exclusion of the discrimination criteria defined by employment law, such as age, sex, origin, family situation, place of residence or health
    • Pseudonymisation of the texts sent, which limits exposure of identities and of information unrelated to the need
    • Dedicated tests to check that these clauses are applied

  • Answers evaluated on scenarios

    The assistant is evaluated on representative scenarios, using fictitious corpora so that no real data is exposed.

    Details
    • Similar experience and rephrased requests
    • Geographical criteria
    • Cases with no answer and visibility exclusions
    • Review of cited sources with the same care as the quality of the text

  • Controlled response time and cost

    While tuning a search assistant on a test corpus, revising the instructions and configuration brought response time down from about 45 to 11 seconds per question, with the same model and the same architecture.

    Details
    • Fewer model calls and a smaller volume of tokens
    • Tracking of cost per request
    • Usage quotas by user group to keep the budget predictable

Some uses of AI in recruitment or career management fall under the systems the AI Act classifies as high-risk. Documented instructions, testing, traceable answers and human oversight help prepare for its requirements, depending on how the system is qualified and the role of each party.

RAG AND CONFIDENTIALITY

Search that controls the data sent to the model

RAG lets the model answer from documents selected in your corpus. Before any AI processing, texts are pseudonymised: names, contact details and personal links are removed, while the experience and information useful for search are kept. Originals stay in the platform with their own access rights, and pseudonymised texts remain personal data subject to GDPR.

Document selection combines structured criteria, applied together with access rights and visibility exclusions, and a meaning-based search guided by HyDE. The answer relies only on the real documents retrieved, and only the selected excerpts are sent to the model.

At indexing, for each document
Original document

Kept in the platform, with its access rights.

Pseudonymisation

Names, contact details and personal links removed.

Structured data and embeddings

Exact criteria such as the mobility area, and an index for meaning-based search.

For each question
Natural-language need

The user’s request, with its criteria.

Deterministic filters

Structured criteria, access rights and visibility exclusions.

Vector search

Guided by HyDE, within the authorised corpus.

Boundary before the LLM
LLM · sourced answer

Only the selected pseudonymised excerpts are sent; the answer links back to its sources.

Search · Structured criteria and rights are applied deterministically; the LLM receives only the selected pseudonymised excerpts.

Explore GDPR consulting

BEST PRACTICES

Useful questions

Which AI model to choose?

We recommend European models, such as those from Mistral AI, used for example in our Hub project. The architecture remains open to other providers: the model is chosen according to the use, the confidentiality required and the budget, and we plan from the design stage for it to be replaceable.

Should people’s identity be sent to the model?

In most HR uses, it is not necessary. We recommend sending the model pseudonymised texts and producing answers without identity, with access to the named record remaining in the application, after a rights check. Pseudonymisation reduces direct identification; the texts remain personal data under GDPR.

What role should human decision-making keep?

We design the assistant to prepare the work: finding records, examining the elements that match the need and flagging what remains to be checked. Evaluation, exchanges with the candidate and the decision remain with the recruiter, which supports the human oversight expected under the AI Act.

What other uses does this approach apply to?

It applies to a knowledge base, HR policies and resources, the review of applications or the matching of job offers and profiles, as well as document collections in other business functions. For each project, we define the uses, access rights and quality criteria.

Where should a project start?

A first, limited use case, with a representative corpus and a few test scenarios, makes it possible to measure relevance, response time and cost before extending it. Fictitious data can be used for the first trials.

Let’s discuss your project

An opportunity to add value with AI?

Describe the intended use and the documents involved. A first conversation is often enough to see whether AI brings a real gain and under what conditions.

Discuss your project