Security: where your data lives and who can see what
Your records stay in your systems. Each person connects with their own account, so the permissions you already have still apply. Anything hard to undo waits for a team member.
In a Deep Blue Coding setup, company data is not moved into a new database. The AI agent reaches each system through a connector at the moment a question needs it, with the sign-in of the person asking, and gets back only what that person is allowed to see. Connectors for internal systems run on the company's own infrastructure, and actions that are hard to undo wait for a named team member to approve them.
How a question travels
Who can see what
The same question gives a different answer depending on who asks. A finance lead sees every account and amount. A sales rep sees the accounts they own. A support agent sees the tickets and not the invoice amounts.
That is not a second permission system to maintain. It is the one you already have: each person signs in to each connector with their own account, and the source system decides what to return. A few older systems only offer one shared account. For those we agree with you what the connector may show, and to which group.
- Sign-in
- Through your identity provider, with single sign-on.
- Connectors
- Your administrators choose which are available, and to which groups.
- Actions
- Each connector exposes a fixed list of actions. Read-only is the default.
- Leavers
- Removing someone from your identity provider removes their access.
Connectors on your infrastructure
A connector we build for an internal system runs on your infrastructure, next to that system. The agent reaches it through one secured address that your IT team controls, and each person signs in to it with their own account. Your IT team reviews its code and the access it needs before it is deployed, and the source code stays in your repository.
More on custom MCP connectorsApprovals and records
When a process reaches a decision, such as approving a purchase or sending a request to a client, it stops. The right team member gets a message in Teams or Slack with the facts and two buttons. The process continues only after they answer.
Each decision is recorded: who was asked, what they were shown, what they answered and when. The record is stored by the approval service, which can run on your infrastructure or be hosted by us, and you can export it at any time.
More on process automation with approvalsThe AI model and your data
The agents run on a commercial AI model, Claude by default. The records a question reads become part of that conversation, which is stored in your company's own workspace with the model's vendor. Before anything is connected we settle these points with your security team, in writing:
- Training
- Whether inputs and outputs can be used to train models, and the plan under which they are not.
- Retention
- How long conversations are kept, and who can delete them.
- Certifications
- Which audits and certifications the vendor holds, and where to download the reports.
- Admin controls
- Single sign-on, audit logs and role-based access: which ones your plan includes.
When we build software for you
The same rules apply to a web product, a mobile app or a dedicated team as to an automation.
- Confidentiality
- We sign an NDA before you share anything, if you want one.
- Code ownership
- Full code ownership stays with you, with clean documentation, so you can walk away at any time with everything you need.
- After launch
- Bug fixes and security updates are part of the support and maintenance plan.
What we can access
During a build we work with the access you grant, on test data wherever possible. That access is listed in writing and removed at handover.
If you take a monthly upkeep plan, we keep only the access needed to check and update connectors, and you can see and revoke it at any time.
Questions security teams ask us
Is our data used to train the model?
That depends on the model and the plan, so we check it with you before anything is connected. The business plans of the main AI vendors exclude training on customer data by default, and we confirm it in writing against the vendor's current terms for the plan you choose.
Where does our data live?
In your own systems, where it is today. A question reads the records it needs through a connector, and those records become part of the conversation of the person who asked, which is stored in your company's workspace with the model's vendor. Nothing is copied into a database owned by Deep Blue Coding.
Can an employee see data they should not?
No more than they can today. Each person connects with their own account wherever the source system supports it, so a connector returns only what that account can already see. Where a system only offers a shared account, we say so and limit that connector to what the whole group may see. Your administrators also decide which connectors each group can use.
Can the agent change or send something on its own?
Only where you have decided it can. Connectors can be read-only. For anything hard to undo, the process stops and a team member approves it in Teams or Slack.
What access does Deep Blue Coding have?
Only what you grant for the build, on test data wherever possible. It is listed in writing and removed at handover unless you take a monthly upkeep plan, in which case it is limited to what upkeep needs.
Will you answer our security questionnaire?
Yes. Send it before the assessment and we answer it in writing. We also join the review with your security team before anything is deployed.
Send us your security questionnaire.
We answer it in writing before the assessment starts, and we join the review with your security team before anything is deployed.
Email social@deepbluecoding.com
We reply within one working day.
Deep Blue Coding