Responsible AI
How we design, build and operate AI systems so that they are useful, safe and under your control. These are the practices we apply in our AI projects, from the first workshop to production.
Our principles
People decide where it matters
Irreversible actions, such as payments, deletions or messages to customers, require human approval. AI prepares the work; a person with context reviews it and takes responsibility for the decision.
Least privilege by design
Each agent or assistant gets only the tools, data and permissions its process needs. Every external input is treated as untrusted to reduce the risk of prompt injection.
Evaluated before launch and after every change
We build a set of real cases with the expected answer, including the difficult ones, and run it whenever prompts, models or tools change. Regressions are caught before users see them.
Traceable and verifiable
Assistants cite the document each answer comes from and say "I don't know" when the answer is not in the sources. Agents log which data they consulted and which tools they used.
Your data, your permissions
AI systems respect the permissions of the original systems: each person sees only what they can already see today. Permissions are enforced in the search, not only in the interface.
Governance proportional to risk
A writing assistant does not need the same controls as a system that decides on credit or hiring. We classify uses by risk and use frameworks such as the NIST AI RMF or the EU AI Act as a reference.
How it shows up in a project
-
Before building
We define the process, the owner and the metric; classify the use by risk; identify which data the system will use and who can see it; and agree which actions always require approval.
-
While building
We design tools and permissions explicitly, build the evaluation set with real cases, test with malicious or unexpected inputs and measure cost and latency per task from the pilot.
-
In production
Logs, cost and latency alerts, an owner and a way to turn the system off. A sample of decisions is reviewed regularly, and unanswered questions become a list of documentation to fix.
What we will not do
-
Promise full autonomy where a mistake is irreversible
Autonomy is earned with evaluation results, process by process. Until then, the agent assists and a person approves.
-
Hide the way to a person
In customer service, the option to talk to a person must be clear and quick, and the conversation must arrive with its context.
-
Let the AI make up policies
Answers about prices, coverage or conditions come from the official documentation, and commitments go through defined rules.
Information security
Responsible AI rests on the same security controls we apply to every project: restricted access to client information, encrypted channels, multi-factor authentication and an incident response process.