Do you know how much your organization spends on AI and what it's using it for?
PromptShield is a context gateway: it logs every request, evaluates it against your rules before it reaches Claude, Gemini or ChatGPT, and blocks the ones that break them.
Request a demoExecutive summary
Your people already use artificial intelligence.
With company accounts, with personal accounts, with several providers at once. No one knows how much it's being used, what information left, or whether it was used for the business. And what people learned in each conversation stays scattered across everyone's computers.
-
Rules that enforce themselves
If a request falls outside the line of business or includes information that must not leave, it is blocked, citing the policy behind the decision.
-
Sensitive information doesn't leave
A blocked request never reaches the cloud. If you have a local model, it answers in-house.
-
How much, who and for what
Every request is logged, including blocked ones: which area uses AI, with which model, for what, and what was stopped.
-
The log stays in-house
What your people ask and learn stops being scattered across personal accounts: it stays in your infrastructure.
How it works
Every request goes through the gateway before it reaches any AI model.
-
Logs the request
Every request is logged as it enters the gateway: who, when and from which application. If it is then blocked, you decide whether its content is kept.
-
Evaluates the context
It checks the request against your restriction registry: rules for formatted data, such as tax IDs, national IDs or account numbers, and a local model that understands context. It isn't a word filter.
-
Blocks with a reason
If it breaks a rule, it doesn't leave. If you have a local model, it answers in-house; if not, the person receives a rejection with the reason and the policy behind it. Both are logged.
-
Sends to the cloud
If it complies, it goes to the model your organization uses, such as Claude, Gemini or ChatGPT, and its answer is logged.
It passes and is logged
An analyst asks to compare two versions of an internal policy. It breaks no rule: it goes to the cloud, and the request and answer are logged.
Customer information
A quality engineer writes: "Help me with the 8D for Automotive Customer X's complaint about lot 24-1187." That customer is protected by contract: the request doesn't leave, the plant's local model answers it, and both are logged. Without a local model, the engineer would receive: "This request mentions a customer with a contractual restriction (AI Use Policy, section 4)."
Outside the business
Someone asks for help planning a vacation. PromptShield replies: "This request is outside the authorized use of AI (AI Use Policy, section 2)," and the rejection is logged.
If something fails, the information doesn't leave.
- A single door. Direct access to AI services is closed; everything goes through the gateway.
- Fail closed. If the evaluation doesn't respond, the request doesn't leave.
- You hold the keys. Provider credentials live only in PromptShield; no one else knows them.
- Everything is logged, in-house. The log keeps every request with its answer or rejection. For blocked requests, you decide whether to keep the content or only who, when and which rule applied. Nothing in the log leaves your infrastructure.
- The log is yours. Because it can contain sensitive information, you decide who can access it and how long it is kept.
What's next
-
Stage 2 · Organizational knowledge In development
Logged requests and answers are processed to identify your organization's know-how and propose it for a knowledge base; nothing goes in without approval from a person on your team. It includes a usage-against-budget dashboard and automatic model selection by task.
-
Stage 3 · Data replacement In development
Identifiable data is replaced before it leaves and restored in the answer, so requests with customer data can use the cloud without exposing it.
Regulatory framework
Less information in third-party hands. More evidence for audit.
PromptShield reduces the information that reaches third parties and produces auditable evidence. It doesn't replace your legal team's analysis. The frameworks below are Mexican.
Finance
- Banking secrecy
- Article 142 of Mexico's Credit Institutions Law. A request with customer names, accounts, loans or amounts doesn't leave: it is blocked or handled by the local model.
- Personal data protection
- Minimization, control over transfers, and a record of every request.
- Circular Única de Bancos (CNBV)
- Third-party and cloud services: a single control point for all AI use, with an exportable log.
- Operational risk (R-28)
- Blocked attempts are logged as evidence for the incident register, along with what was actually sent to external providers.
Manufacturing and other industries
- Confidentiality agreements
- Rules per customer: if a contract prohibits using external AI, anything that mentions that customer is blocked or stays in-house.
- Customer audits
- Concrete answers for security questionnaires that already ask about AI use.
- ISO 27001 and industry certifications
- Evidence of control over how AI is used.
- Personal data protection
- Employee data, such as names, social security numbers and salaries, doesn't leave when using AI.
In any industry: logging requests and their answers is processing of your employees' data. Disclose it in your internal privacy notice and define its retention.
Rollout plan
We start small, measure and grow.
-
Assessment
Who uses AI today, with which providers and for what. Usage policy and a first restriction registry by area.
-
Pilot
One or two areas. We measure usage, blocks, and how well it detects sensitive information in your real documents.
-
Production
Direct access to AI services is closed and every area goes through the gateway.
-
Continuous improvement
Regular reports to management and adjustments to the restriction registry.
No phase moves forward without your team's approval. No system catches 100% of cases; that's why we measure with your documents before going live, and the most sensitive areas can work with the local model only.