Enterprise workflow automation
Agentic AI Time Tracking Assistant for Enterprise Workflows
Overview
Designed and implemented a local-first Agentic AI assistant that connects authenticated calendar activity with live PSA work structures and permits only explicitly approved draft writes.
Context / Problem
Preparing accurate time records required reconciling calendar activity with live project, assignment, and milestone structures. A useful assistant had to handle ambiguity without bypassing enterprise authentication, overwriting existing data, or turning a planning aid into an autonomous submission mechanism.
My Role
I designed and implemented the local-first workflow, its browser and PSA integrations, typed planning model, mapping memory, safety validators, confirmation gate, audit trail, and post-write verification.
What I Did
- Read calendar activity through an already authenticated browser session using Playwright and normalized events into a reviewable form.
- Discovered and validated live Salesforce and Certinia work structures, then combined them with local mapping memory and ambiguity handling.
- Generated a typed draft plan, required explicit confirmation, wrote only scoped draft records, and verified the resulting state.
Architecture
The workflow starts with a normal authenticated calendar session, normalizes events, maps them against validated local memory and the live PSA catalog, and produces a draft plan. Only an explicit human confirmation opens the narrow draft-write path; the assistant then reads the saved state back for verification.
- Calendar sessionImplemented
- Event normalizationImplemented
- Mapping memoryImplemented
- Live PSA validationImplemented
- Draft planImplemented
- Human confirmationHuman approval
- Scoped draft writeImplemented
- Post-write verificationImplemented
No submit · No approve · No generic write tool
Key Decisions / Trade-offs
- Keep execution local-first
- Local orchestration preserves direct user control and reuses ordinary authenticated sessions without introducing an unattended runtime.
- Use a narrow write contract
- Typed draft-only operations make destructive or unsupported actions unavailable instead of relying on a broad generic write tool.
- Validate memory against live data
- Stored mappings accelerate recurring work but never override the current assignment and milestone catalog.
Implementation Scope
The local assistant, integrations, planning flow, validation, audit logging, guarded draft writes, and read-back verification were implemented. The design remains compatible with a future managed Agentic runtime, but that runtime was not part of the implementation.
Safety / Security / Governance
The assistant uses normal enterprise SSO and MFA through an authenticated browser session. Existing entries are preserved, submitted or approved records are rejected, every write needs explicit confirmation, and no submit, approve, or generic Salesforce write capability exists.
Scope Boundary
Implemented scope is local-first and limited to reviewable draft creation. The assistant does not submit or approve timecards, bypass SSO or MFA, or expose a generic write tool. AWS AgentCore Runtime and Strands are target-compatible design options only and are not claimed as implemented.