What the committee asks before approving an AI project
Compliance, security, Legal, procurement and technology. Each question with the short answer and, on its own page, the detail with sources.
What data leaves my perimeter when I use private AI?
None, when the model runs on your hardware inside a segmented network with no internet egress. Sensitive and confidential data is processed inside the perimeter, without exception. Internal, non-sensitive data may use an enterprise cloud with zero retention only if compliance approves. Public data stays free.
Do you use our customers’ data to train the model?
No. No end-customer data goes into training. The model learns voice, format and craft from approved, masked answers, product documents and examples written by your experts. Sensitive data is consulted at answer time, with the user’s permissions and a record of every access.
How are data subject rights and the right to object to automated decisions handled?
No agent decides alone: it gathers, organizes and hands off, and a person reviews before any effect on a customer or an employee. That meets article 27 of Mexico’s data protection law by design. The audit log shows which data each answer consulted, to handle access, rectification, cancellation or objection requests (ARCO rights).
What happens if there is a data breach in the system?
The incident response protocol in the evidence file kicks in: what counts as an incident, severity levels, response times, who shuts the system down, communication to compliance and to the authority, and a post-incident review. The law requires notifying the data subject immediately. We don’t promise it won’t happen; we promise what happens if it does.
How is the system isolated from the internet and the public cloud?
The production server is installed on a segmented network with no internet egress, inside your perimeter. The model, the permissioned knowledge base and the audit log live there. It connects to your systems only to read, never to modify. The demonstration runs on a computer with no internet, so the committee sees it working before signing.
Do you run controlled attack testing (red teaming), and who leads it?
Yes, before the pilot and at every major version. Your security team leads it, not us: it is part of independent validation. The report shows the attack success rate by category, the fix applied and the retest. Findings are fixed before the pilot.
Is there an AI law in Mexico that companies must comply with?
No. Mexico has no general AI law: the Senate draft has no committee report and first requires a constitutional amendment. What does apply: the 2025 data protection law, with an authority that already fines, and each sector’s third-party contracting rules.
Who owns the trained model and what it produces?
You do. The model weights, its identity adapters, the tests and the audit log belong to the client and are exported at no cost in a standard format. The base model keeps its open license. Work generated solely by a machine has no copyright; your brand is protected by trademark registration and every piece goes through your approval queue.
What licenses do the base models have, and what does Legal review?
Llama 3.3 is used under Meta’s community license: a copy of the license, “Built with Llama” attribution, a technical name that includes “Llama” and the acceptable use policy. Qwen3 and Mistral 3 are under Apache 2.0; earlier versions have research licenses, so each version is reviewed. The commercial seal, Dabar, is separate.
What happens if the AI vendor disappears or defaults?
With the hardware on your site, you already have everything in your company: hardware, model weights, knowledge base, audit log and manuals. The exit plan sets format, deadlines, transition and certified deletion, and it is tested every year. When the hardware is ours, optional escrow of models and code with a third party is available.
Which model do you use, and why not a frontier model from OpenAI or Anthropic?
A current-generation open model (Llama, Qwen or Mistral), chosen for each project, installed on your hardware and measured against your own questions. It is not a frontier model and we say so. The reason is the perimeter: sensitive data cannot leave. Commercial models are used in our studio to design and code, never on your sensitive data.
How do we know the model works before it goes to production?
With a held-out question test: your experts write questions the model never saw, the threshold is set in advance and results are reported by segment. Your team grades it, not ours. A version is released only if it matches or beats the previous one, after a week with internal administrators and the written sign-off of the process owner.
What about hallucinations? Can the model make up a policy or a clause?
It can be wrong; no model is infallible and we don’t promise otherwise. That is why it answers from the permissioned knowledge base: it searches your documents and cites the source. Every week we measure the share of answers supported by the documents they cite, with automated grading and a sample reviewed by people. Unsupported answers go to a person.
How do you prevent bias and automated decisions about people?
The first project never includes an automated decision about people: credit scoring, underwriting, medical decisions or hiring are deferred by rule. The agent gathers, organizes and hands off; a person decides. Test results are reported by segment to detect differences, and the impact assessment documents risks and mitigations before anything is built.
What happens when the model ages or our operation changes?
Drift is monitored daily: changes in the kind of questions and in answer quality against the baseline. The knowledge base is updated within two business days. Retraining is quarterly and every new version passes the same test. Without operation, answers age within six months; that is why continuous operation exists.
How do you manage model changes and versions?
No version reaches production without three conditions: an equal or better test result, a week with internal administrators and the written sign-off of the process owner. Every version is signed, archived and can be restored. Each one carries its model card and its row in the registry. Any change of base model is announced in advance.
What guarantee does the Assessment offer if the project doesn’t pay off?
The Assessment comes with a guarantee: fixed price, 50% at the start and 50% at the final report. If the final report does not identify a viable use case (one with a positive twelve-month business case and no block from Legal), the final 50% is not charged. Implementation is paid by demonstrable milestones. No one moves forward without seeing their own model work.
Why DEBIR and not a large integrator certified under ISO 42001?
We are not certified under ISO 42001 and we don’t say we are. We deliver the documents an auditor of that standard would ask for: policy, impact assessment, registry, model cards and incident protocol, mapped to the NIST AI RMF and to SR 26-2. The difference: perimeter, identity and operation in one dedicated team, from day one to month twelve.
Discovery sessions, no commitment
With technology, compliance and two process owners. Includes an offline demonstration prepared for your sector, without using your company’s data.
Inventory and obligations
Data in four levels and a first review of what your sector requires.
Three candidate use cases
And the Assessment proposal. The Discovery is credited against the Assessment.
Founding clients program
Three seats, with an explicit discount in exchange for a publishable case and a reference.
Terms
Discovery and evidence pilot at no cost, Assessment at a discount, in exchange for a publishable case, a reference and feedback. Continuous operation is always paid at list price.