Technical Solution Proposal
Technical Solution Proposal 

AI-augmented workflow automation, delivered
on-premises

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.
 AI as a first-class workflow step
 On-prem data sovereignty
 Human-in-the-loop governance
Engineered for 50,000 concurrent users
In one line. Process owners design automations on a visual canvas, publish them under validation, and run them on a reliable, crash-safe engine — where each run can pause for human approval, invoke an AI agent grounded in your own documents, generate and e-sign official documents, read internal data, connect to third-party systems, and compose sub-processes — all inside one governance fabric.
01 / BUSINESS OUTCOMES

What it delivers

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.
Standardize & automate cross-team processes
Visual workflows authored by business analysts; reusable sub-processes.
Faster, rule-driven, org-aware approvals
Human approval steps with dynamic manager-hierarchy assignment, SLA deadlines, action-button routing.
Consistent, compliant documents
Templates rendered to PDF/Word with Nghị Định 30 fidelity and built-in e-signature.
Tamper-evident audit & governance
Immutable, append-only audit trail; per-run immutable process records.
Trustworthy, controllable AI
Grounded retrieval with per-field source citations, token budgets, and sensitive-data redaction; models run under your control.
Data sovereignty
Full on-premises deployment; single-tenant hard isolation; optional air-gapped AI inference.
02 / CAPABILITIES AT A GLANCE

Solution capabilities

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
03 / PROCESS BUILDING BLOCKS

The building blocks of a process

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.
👤
Human Approval / Task
Human
Assignable task with optional form, org-aware assignment, prefill, approve/reject, SLA, review & signing modes.
🤖
AI Agent
AI
AI as an auditable, budgeted, governed step — per-step model, tools, knowledge grounding, structured output.
◇
Decision / Branch
Flow
First-match-wins rule-based branching over ordered conditions, with an optional default path.
📄
Document Generation
Docs
Renders forms and templates to PDF or Word with dynamic expressions and Nghị Định 30 fidelity.
✍️
e-Signature
Docs
Multi-signer, multi-document signing envelopes with template placement and coverage validation.
🔌
System Integration
Integrate
Third-party connector (HubSpot today) — live read plus durable writes, connection pinned per run.
🔎
Data Lookup
Integrate
Read-only, parameterized query against an internal database, capturing results into workflow variables.
🗄️
Data Persistence
Integrate
Native structured-data capture — create-or-update an internal record with mapped, validated fields.
🔁
Batch / Loop
Flow
Fans an inline sub-process over a collection as isolated child runs — parallel or sequential, bounded.
⧉
Sub-process
Flow
Synchronously invokes another published workflow as a child run — reusable, composable process libraries.
⇉
Synchronization
Flow
Explicit join barrier — holds until every incoming branch completes, then fires once. No silent stuck runs.
{x}
Variables & Mapping
Flow
Assigns values mid-run (reference, template, or literal) — stateful workflows without external storage.
✉️
Email
Deliver
Templated outbound email with attachments, durably queued for reliable delivery.
🔗
Webhook
Deliver
Outbound HTTP call (method, headers, auth, templated body) — push events to any endpoint.
▶
Start / End
Flow
A single, deterministic entry point and a terminal step that declares the workflow's output contract.
i
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).
04 / LOGICAL ARCHITECTURE

A layered solution architecture

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).
Experience Layer
Visual Workflow Designer
Low-code canvas for business authors
Task Inbox & Forms
Where people receive and complete work
Admin Console
Users, roles, connections, settings
Dashboards
Operational & process visibility
▼
Orchestration Layer
Workflow Engine
Reliable · crash-safe · horizontally scalable
Process Building Blocks
Human · AI · Decision · Document · e-Sign · Integration · Data · Sub-process · Notify · Batch · Sync · Variables
Triggers & Scheduling
Recurring calendars + connected-app events
▼
Intelligence Layer
AI Agent Service
Governed AI invocation per step
Knowledge Base / Retrieval
Grounding on your own documents
AI Governance
Budgets · redaction · source citations
Open AI Interface
Secure access for external AI assistants
▼
Business Services
Forms & Human Tasks
Dynamic, org-aware, SLA-driven
Document Generation
PDF / Word, Nghị Định 30
e-Signature
Multi-party, multi-document
▼
Integration Layer
Application Connectors
CRM / HubSpot and more
Data Source Connectors
Internal databases · read-only
Delivery Channels
Email · webhook
▼
Platform Services (shared)
Identity & Access
SSO · role-based access · multi-tenant
Governance & Audit
Immutable trail · sensitive-data redaction
Notifications
In-app · email · real-time
Data & Storage Platform
Highly available, on-premises
Observability
Metrics · logs · tracing
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.
06 / INVESTMENT & DECISIONS

Indicative investment

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.
Headline commercials (indicative)
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.
Two launch-gating decisions
⚠️ Disaster recovery
A single site has a hard recovery ceiling. A funded second site delivers near-real-time protection. This is a data-residency and budget call.
⚠️ AI inference posture
Controlled egress (Claude + HubSpot available) vs. air-gapped self-hosted (maximum data control). Decide before hardware is ordered.
i
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.
Why AgentFlow
🔒 On-prem data control
Identity, data, and optionally the AI models stay under your roof — sovereignty by architecture, not by promise.
🤖 Governed AI
Every AI step is grounded, budgeted, redacted, and traceable to its source — AI you can put in a regulated process.
🧾 Immutable audit
An append-only audit trail and per-run frozen records make every execution deterministic and defensible.
👤 Human-in-the-loop
People stay in control where it matters, with exactly the approval, review, and signing controls each step needs.
📄 Document fidelity
Native Vietnamese government (Nghị Định 30) formatting and legally-binding e-signature inside the process.
🔓 Open & composable
Reusable sub-processes, connectors, and a secure interface for approved external AI assistants.
05 / AI STRATEGY

AI inference — a governance decision

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.

BRANCH A Controlled egress

  • AI agents call external AI services through a governed, allow-listed gateway.
  • Claude and HubSpot both work — the strongest models and the CRM connector are available immediately.
  • The trade-off: prompts, CRM records, and rendered documents leave the premises — this must be a signed governance decision with zero-retention terms.
Minimal added cost · usage billed per token

 BRANCH B Air-gapped / self-hosted

  • AI runs on self-hosted models inside your data center — nothing leaves the building.
  • Claude cannot be self-hosted; those steps move to a self-hostable model. The HubSpot connector is unavailable air-gapped.
  • Highest data control; requires dedicated AI infrastructure and a model-quality validation pass.
Added hardware investment · no external usage bill
!
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.

Hotline: (+084) 028 7309 9009

Email: info@innomation.vn

Website: https://www.innomation.vn/

INNOMATION TECHNOLOGY CORPORATION

Product

Artboard 26

Privacy Policy

http://thenounproject.comThe Noun ProjectIcon TemplateRemindersStrokesTry to keep strokes at 4pxMinimum stroke weight is 2pxFor thicker strokes use even numbers: 6px, 8px etc.Remember to expand strokes before saving as an SVG SizeCannot be wider or taller than 100px (artboard size)Scale your icon to fill as much of the artboard as possibleUngroupIf your design has more than one shape, make sure to ungroupSave asSave as .SVG and make sure “Use Artboards” is checked100px.SVG

Contact Us

© 2024 Innomation. Copyright Registered.