You own the accounts
Systems are built and run in accounts registered to your business: your cloud, your automation platform, your AI provider account.
AI systems are only useful if they can reach your information and act in your tools. That is exactly why they need careful limits. This page explains how those limits are set, in plain terms.
Straight up: Georgy Consulting is a one-person practice and does not hold certifications such as SOC 2 or ISO 27001. Instead, the work is designed to fit inside your security and compliance framework, uses services you approve, and leaves you in control of every account and credential.
Where possible, every part of the system lives in accounts your business owns. The one thing that usually leaves is the text a model needs to read, and only the part the task requires.
Systems are built and run in accounts registered to your business: your cloud, your automation platform, your AI provider account.
Each system gets only the permissions its task needs: read-only where reading is enough, limited to specific folders, mailboxes or records.
Only the information a task needs is sent to an AI model. Where practical, personal or sensitive details are removed or masked first.
Actions with real consequences, like sending to customers, moving money or changing important records, can require a person's approval.
Inputs, outputs and actions recorded at a level agreed with you, while keeping logs free of unnecessary sensitive data.
Systems are tested against real examples, including awkward and adversarial ones, before they're used for real work.
Most AI systems send text or documents to a model provider for processing. Which provider, under which terms, and in which region are decisions you should make knowingly.
Business data goes through providers' business or API offerings, whose terms on data use and retention differ from free consumer apps.
Before anything is built, you'll see how the chosen provider says it handles data (retention, training use, where processing happens) so your team can approve it. Providers change their terms, so this is checked for each project.
Where needed, models can be accessed through your existing cloud provider, in a specific region, or, for some tasks, run on infrastructure you control.
Not every step needs the most powerful model. Matching the model to the task can reduce both cost and how much data leaves your systems.
No design removes these risks completely. The aim is to reduce them to a level you're comfortable with, and to make problems visible when they happen.
| Risk | What it looks like | How designs respond |
|---|---|---|
| Wrong or invented answers | Confident output that is incorrect | Answers grounded in your documents with citations, outputs checked against your data, human review where accuracy matters |
| Manipulated inputs | Text in an email or document trying to steer the system (“prompt injection”) | Tight limits on what the system may do, instructions kept separate from content, approval for sensitive actions |
| Data leaking between users | An assistant revealing something a person couldn't otherwise see | Your existing permissions respected rather than pooling everything together |
| Behavior changing over time | Model updates or new kinds of input changing results | A saved test set re-run after changes, and monitoring of real use |
If you work under rules such as GDPR, HIPAA, financial services regulation or professional confidentiality duties, those requirements shape the design from the first conversation: which data may be used, which providers are acceptable, what must be logged and what needs human review.
Georgy Consulting is not a law firm and doesn't give legal advice. For regulated data, your legal or compliance advisers should confirm the approach meets your obligations, and the design documentation is written to make that review straightforward.
Questions about a specific project? Email info@georgyconsulting.ai.
Send it over. You'll get straight answers, including where the answer is “that's handled by your provider” or “not yet”.