📅 Book a 30-min Demo📞 Call/text (571) 293-0242

Appendix EConnector and Broker Policy Catalogue

3 min readFrom Sovereign AI Engineering by Miguel AmigotLast updated
Share

A starting catalogue of the systems a sovereign platform should reach through Model Context Protocol connectors, by sector, with the scopes a broker policy should distinguish and the audit fields every call produces. It is a template to be edited into the institution's own policy table, not a complete list.

Audit fields every call produces

Whatever the connector, the broker writes one line per call with these fields. They are the minimum a regulator, an auditor or an incident review will ask for.

FieldContent
timestampWhen the call was made
userThe person on whose behalf the call was made
roleThe role the policy was evaluated against
agentThe agent that made the call
surfaceWeb, side panel, desktop, mobile, terminal, scheduled
toolConnector and tool name
scopeThe scope granted: self, caseload, department, organization
parametersThe call's inputs, with secret fields redacted
decisionAllowed or denied, and the policy rule that decided
result_sizeRows or bytes returned
classificationThe data classification of what was returned
model_targetWhich model later received the result, from the router

Higher education

SystemExample productsRead scopesWrite scopesClassification
Student information systemBanner, PeopleSoft Campus, Jenzabar, ColleagueOwn record (student); caseload (advisor); program (chair); all (registrar)Holds, notes (registrar roles only)Restricted (FERPA)
Learning management systemCanvas, Blackboard, D2L, MoodleOwn courses (student); taught courses (instructor); all (admin)Announcements, grades (instructor)Restricted
Recruitment and advancement CRMSlate, SalesforceOwn territory (recruiter); all (director)Notes, tasksConfidential
Advising platformEAB Navigate and equivalentsCaseloadNotes, alertsRestricted
Ticketing and service deskServiceNow, TeamDynamix, Jira Service ManagementOwn tickets (requester); queue (agent)Create, comment, resolveInternal
Identity directoryActive Directory, Okta, institutional IdPOwn profile; lookup by roleNoneInternal
Document storesGoogle Drive, SharePoint, BoxOwn and shared-with-meCreate in own spaceVaries by item
Library and research databasesDiscovery layers, publisher APIsCatalogue searchNonePublic

Enterprise

SystemExample productsRead scopesWrite scopesClassification
CRMSalesforce, HubSpot, DynamicsOwn accounts; team; allNotes, tasks, opportunity fieldsConfidential
ERP and financeSAP, Oracle, NetSuite, Workday FinancialsOwn cost centre; department; allDraft journal entries, approvals (finance roles)Restricted
HR systemWorkday, SuccessFactors, BambooHROwn record; direct reports; all (HR)Time-off requestsRestricted
Support deskZendesk, Freshdesk, IntercomOwn tickets; queueCreate, reply, resolveInternal
Project and work trackingJira, Asana, MondayOwn projects; teamCreate, updateInternal
CollaborationSlack, Teams, emailOwn channels and mailPost, send (approval recommended)Internal
Data warehouseSnowflake, BigQuery, RedshiftGoverned views by roleNoneVaries by view
Contracts and documentsDocuSign, SharePoint, BoxOwn and sharedCreate draftsConfidential

Healthcare

SystemExample interfacesRead scopesWrite scopesClassification
Electronic health recordFHIR APIs; vendor-sanctioned interfacesOwn patients (clinician); panel (care team); none (non-clinical)Draft notes only, clinician review requiredRestricted (HIPAA); local models only
SchedulingPractice management systemsOwn schedule; departmentRequestsRestricted
Clinical documentsDocument managementOwn patientsNoneRestricted
BillingRevenue cycle systemsDepartmentNoneRestricted

Healthcare connectors should route exclusively to models on the restricted list, which in practice means local inference on institution hardware.

SystemExample productsRead scopesWrite scopesClassification
Matter and case managementClio, iManage, NetDocumentsOwn matters; practice groupNotes, draftsPrivileged; local models recommended
Document managementiManage, NetDocuments, SharePointOwn and matter-sharedCreate draftsPrivileged
Billing and timeClio, Elite, AderantOwn entries; partner viewTime entriesConfidential
Market and reference dataVendor feedsLicensed scopeNoneLicensed
Research databasesLegal and financial research platformsLicensed scopeNoneLicensed
KYC, AML and complianceScreening and case toolsOwn queue; compliance teamCase notes, escalations (approval required)Restricted

Writing the policy table from the catalogue

Each row above becomes several rows in the broker's policy table: one per role and tool, with the scope the role gets and whether the tool may write. Three rules keep the table honest.

  • Start empty. Deny by default; add the first agent's needs; add the next agent's. The table should read as a history of deliberate grants.
  • Scope before allow. Never grant a role "all" when "own" or "caseload" is what the agent needs. The source system's own permissions are the backstop, not the control.
  • Writes require an approval mode. Any tool that changes a system of record should be paired with the ask-first approval mode in the harness, at least until the agent has an evaluation history.

Run the stack this book describes

ibl.ai is the agent operating system you run yourself: you own all the code and the data, in your cloud, on-premise or air-gapped, with any model, one identity across web, browser, desktop, mobile and terminal, and no per-seat pricing.