AgentFlow turns fragmented, manual processes into governed, auditable automations — keeping people in control where it matters and applying AI where it accelerates, with all data (and optionally the AI models) under your own roof.
AgentFlow replaces email-and-spreadsheet processes with standardized, auditable automations that keep humans in control and apply AI under governance. On-premises delivery gives your team full custody of identity, data, and — optionally — the AI models themselves.
Capabilities are described at solution level; partial or preview features are flagged with honest maturity.
| Capability area | What [Customer] gets | Maturity |
|---|---|---|
| Visual workflow designer | Low-code drag-and-drop canvas, rich library of process building blocks, inline validation, version history | GA |
| Orchestration engine | Reliable, crash-safe, horizontally scalable execution; concurrent branches | GA |
| AI agents | Per-step model selection, structured output, token budgets, sensitive-data redaction | GA |
| Knowledge / retrieval | Grounding on your own documents, version-stable scope, per-field source citations | GA |
| Open AI interface | Agents consume external AI tools; a secure interface exposes workflows to approved external AI assistants | GA |
| Human tasks & forms | Dynamic forms, org-chart-aware assignment, SLA, approve/reject routing, review mode | GA / recent |
| Document generation | Templates → PDF + Word, dynamic expressions, Nghị Định 30 format | GA |
| e-Signature | Multi-signer, multi-document envelopes, publish-time coverage validation | GA |
| Integrations | Application connectors (HubSpot CRM); read-only internal data lookup | GA |
| Delivery | Email + webhook notifications, durable delivery queue | GA |
| Platform & governance | Single sign-on, role-based access, multi-tenancy, append-only audit, notifications, bilingual UI (en/vi) | GA |
| Analytics / process-mining | Process-intelligence dashboard (Overview / Map / Variants / Conformance / Insights) | Preview |
A workflow is a graph of these blocks connected by transitions. Every block is validated before a process can go live, and each run is frozen into an immutable record so republishing can never alter a run already in flight.
Beyond the blocks: typed workflow variables, a shared expression engine with publish-time validation, unified triggers (recurring schedule or connected-app events, with duplicate-fire protection), per-step error handling, version history with restore, organizing groups, and per-workflow start-access control (restricted by default).
AgentFlow is organized into clear logical layers. The experience layer calls the orchestration layer; orchestration drives the intelligence, business, and integration services; and every layer rests on shared platform services (identity, audit, notifications, data & storage, observability).
Reliable by design. Business logic is cleanly separated from storage and transport; every state change is transactional and audited; and each run is frozen into an immutable record so an in-flight process can never be altered by a later change — a governance property most automation tools lack. The on-premises engineering (high-availability data platform, storage, and inference options) is detailed in the companion deployment plan.
The same workflows run against either external AI services or self-hosted models — configurable per deployment. Which branch you choose is a data-governance call, not an engineering one.
Choose from data policy, not engineering. If governed egress is permitted under zero-retention terms, Branch A is markedly cheaper and unlocks the strongest models plus the CRM connector. If a true air-gap is mandated, Branch B is the compliant path — plan for AI infrastructure and a quality bake-off. A mixed approach (simple steps on a self-hosted model, complex steps external) is also possible.
Delivered as a self-operated, highly-available on-premises platform, pre-provisioned to peak load. Figures are planning estimates, not quotes — full engineering and sizing are in the companion deployment plan.
| Line | External-AI option | Self-hosted-AI option |
|---|---|---|
| One-time investment (single site) | ~$480k–835k | + AI infrastructure (hardware-dependent) |
| Annual operating | ~$725k–1.6M / yr | Same base + AI infrastructure running cost |
Annual operating cost is dominated by the platform operations team and support subscriptions. A dual-site disaster-recovery target roughly doubles the data, storage, and network lines.
Honest readiness note. The platform is architecturally close to this scale; a scoped set of hardening steps and a load test precede full go-live. These are folded into the phased delivery plan in the proposal — none are speculative.