One foundation for IT, not another integration project.
Connectors, shared definitions and role-scoped agents.
What costs technology teams today
What IT has to settle before letting an agent touch a system. Pick yours above and Xophia shows how it would be solved in your architecture.
Every new integration starts from scratch
Another connector, another mapping, another service account. What was built for one project isn't reused in the next.
We don't know what an agent can touch
An AI pilot with broad access to your systems is a risk nobody signs off: without scoped permissions and a record, it doesn't pass the security review.
Security asks for evidence we don't have
Who approved what, with which data and when. If that lives in scattered emails and logs, every audit is a reconstruction.
The business buys tools we don't control
Every team adds its own SaaS and IT ends up maintaining integrations, licenses and access that nobody designed to work together.
One run, end to end
This is how a module runs: the connector reads with a scoped role, the agent proposes, the policy holds the write and the person responsible approves. Everything is recorded.
Illustrative example with fictional data. Roles, approval policies and identities are defined with your team for each deployment.
01 · Read
The connector reads with a scoped role
The reconciliation module's agent queries bank movements and open invoices in the ERP with read-only permissions.
- Role
- finance.read
- Read
- 412 movements · 388 invoices
02 · Hold
The policy holds the write
The agent proposes the matches. Because writing to the ERP is a high-impact action, the policy pauses the run until someone with the right role approves.
- Proposal
- 409 matches · 3 differences
- Policy
- erp_write · approval required
03 · Record
The write runs with its owner
Once approved, the write runs under the operator's identity when the system supports it. Calls, approval and outcome stay in the record.
- Approved
- Finance · 10:42
- Write
- 409 lines · operator identity
Read them as technical documents
Each reference module shows its architecture, the systems it touches and an illustrative trace of tool calls. They all run on the same foundation.
- Finance reconciliationSeparate processes for bank, cards, SII and Buk within one controlled workflow against the ERP. Writes wait for approval and use delegated identity or a least-privilege service account.
- Sales supervisionReads the CRM and a semantic layer of sales metrics, and never writes to the CRM without approval. The trace shows which tools it calls and with what permissions.
- KPI copilotSQL compiled from the semantic layer, validated against the schema and row-level security, and run with a read-only role in your warehouse.
Security and compliance, no embellishment
What's available and what's in progress, stated separately. The full detail is in Security.
Identity and least privilege
SSO via SAML 2.0 or OIDC, resource-level role-based access and role-scoped tools for every agent. Identities and service accounts are documented per deployment.
Approval before any write
Writes to your systems and high-impact actions require human approval. Approvers and limits are defined per deployment.
Records and actual status
Calls, approvals and outcomes are recorded for review, with TLS 1.2+ in transit and AES-256 at rest. SOC 2 Type II readiness is in progress: we don't claim that report or an ISO 27001 certification.
Reading before the technical review
See the whole blog- What an agent needs to workAn agent doesn't work in a vacuum. Connecting it to operations means defining how it queries, how it acts and which controls it needs; the specific pieces depend on the task.
- MCP: superpowers inside your companyMCP standardizes how an assistant discovers and uses capabilities. It includes optional HTTP authorization; business permissions, approvals and logging still need their own design.
- Security for agents: identity, least privilege and prompt injectionAn agent with access to your systems is a new kind of user, one that reads everything it's given and sometimes obeys instructions it shouldn't. Protecting it starts with fifty-year-old principles.
Questions from a technical review
How does it connect to our systems?
Through MCP-compatible tools and APIs such as REST, GraphQL and SQL, plus events and webhooks, with ERP, CRM, banks, SII, Buk and BI. The exact adapter, permissions and write scope are confirmed during implementation.
Where does it run?
In Xophia's managed cloud, with logical isolation per organization, or inside your company's VPC, depending on scope. Region, retention, data flows and backups are confirmed and documented before deployment.
What can an agent do without approval?
Only use the tools and data its role allows, such as reading. Writes to your systems and high-impact actions wait for a person's approval, and every call is recorded.
Can you answer our security questionnaire?
Yes. Send us the questionnaire or request a review with engineering. There we show which controls are available, what's in progress and what evidence can be reviewed, including the current scope of penetration testing.
What would it take for your team to approve it?
Request a technical review: architecture, connectors, permissions and deployment, with the people who build the platform. If you have a security questionnaire, bring it.