Human-Centred Agentic Intelligence (HAI): A Three-Layer Framework for Designing Human-Agent-System Journeys

Pothiraj, Anandakumar Muniasamy

doi:10.5281/zenodo.20389202

01 Abstract

What this paper proposes.

This whitepaper introduces the Human-Centred Agentic Intelligence (HAI) framework, a dual-mode, three-layer design methodology for mapping human–agent–system journeys. As organisations deploy AI agents across both discrete product tasks and extended customer lifecycles, existing design tools fail to capture agent reasoning, system infrastructure, and failure handling alongside the human experience.

HAI addresses this gap through a Unified Three-Layer Journey Matrix comprising a Human Layer (user/customer experience), an Agent Layer (reasoning, autonomy modes, and failure protocols), and a System Layer (data, infrastructure, protocols, and governance). The framework operates in two modes: a User Journey Matrix for step-based product and task design, and a Customer Journey Matrix for stage-based lifecycle design — sharing identical Agent and System layers.

Complementary structures include a Goal Failure Response Protocol with seven failure types and five failure states, a Multi-Agent Role Taxonomy with layered accountability, a three-tier agent memory architecture, protocol integration guidance for MCP and A2A standards, EU AI Act compliance mapping, and a three-layer measurement framework.

This edition adds a Sovereign AI alignment layer: a Sovereignty Boundary dimension in the System Layer and a structured sovereignty taxonomy, treating data, compute, model, and regulatory sovereignty as design-time specifications rather than runtime constraints.

Keywords Agentic AIHuman-Centred AIJourney MappingMulti-Agent SystemsEU AI ActSovereign AIDigital SovereigntyMCP ProtocolA2A ProtocolUX Design Framework

02 The Two Journeys

One product. One relationship. Two design problems wrongly conflated.

Modern AI deployments span two fundamentally different journey types — and most organisations conflate them, producing matrices that are either too tactical (missing lifecycle strategy) or too abstract (missing operational precision). Neither journey can be designed well without the other.

Bounded · Sequential · Step-specific

The User Journey

The experience a person has performing a specific task inside a product or service. Onboarding a user, completing a checkout, resolving a support ticket. Failure has immediate, localised consequences.

Owned by product designers, UX teams, agentic engineers.

Spanning · Lifecycle · Stage-based

The Customer Journey

The experience a person has across their full relationship with an organisation — awareness through advocacy. Trust evolves, intent shifts, memory accumulates. Failure has strategic, relationship-level consequences.

Owned by CX strategists, marketing, customer success.

03 Unified Three-Layer Matrix

Three layers of design, asked at every phase of every journey.

Traditional journey mapping captures the human dimension but ignores the agent and system. Agentic AI frameworks address the agent but ignore the human. HAI addresses all three — for both journey types — through a single matrix structure.

Layer 01

Human Layer

What the person sees, does, feels, and needs — labelled "User Layer" in step-based mode, "Customer Layer" in stage-based mode.

Layer 02

Agent Layer

What the AI thinks, why it acts, how autonomous it should be, and what it does when things go wrong. Identical across both modes.

Layer 03

System Layer

What data is required, how fresh it must be, which protocols govern connections, and what compliance, regulatory, and data-sovereignty constraints apply.

Q1What is the human trying to accomplish, and how do they feel?

Q2What triggers the agent to act, and what is it trying to achieve?

Q3How autonomous is the agent at this phase, and why?

Q4What data must be available and what infrastructure is required?

Q5Which agents are involved, and how do they interact under normal operation?

Q6What happens if the agent cannot achieve its goal, and when does control escalate?

04 Dual-Mode Architecture

Two matrix templates. Same Agent and System layers. Distinct Human vocabularies.

The framework provides two matrix templates that share an identical Agent Layer and System Layer, but carry Human Layers tailored to their purpose — step-based product design, or stage-based lifecycle design.

Dimension

Mode A — User Journey

Mode B — Customer Journey

Unit of analysis

A discrete task step

A lifecycle stage

Time horizon

Minutes to hours

Days, weeks, months

Primary audience

Product, UX, agentic engineering

CX, marketing, customer success

Failure impact

Localised — affects task completion

Strategic — affects the relationship

Memory emphasis

Short-term working memory primary

All three tiers; procedural critical

Cross-mode design rule

The Agent Mode established in the Customer Journey Matrix for a stage sets the ceiling for agent autonomy within that stage's User Journey steps. Awareness as Assistive means no individual step within Awareness may run Autonomous.

05 Agent Mode Taxonomy

Three modes of autonomy, calibrated to phase safety and demonstrated trust.

Mode 01

Assistive

Human initiates. Agent informs only. Human decides every action.

When → high-risk · awareness · new users

Mode 02

Advisory

Agent suggests. Human approves or modifies before anything happens.

When → consideration · configuration · decision support

Mode 03

Autonomous

Agent acts within guardrails. Human can override or undo post-action.

When → proven competence · retention · routine steps

06 Failure & Handover

Every phase has a documented failure path, not just a happy path.

Agentic systems regularly encounter ambiguous inputs, data unavailability, conflicting constraints, low-confidence outputs, and timeouts. Without a documented failure protocol, agents fail silently, retry indefinitely, or escalate without context. HAI defines seven failure types, five failure states, and a nine-element handover context package for transferring control to a human.

State 01

Active

Normal operation. Normal interaction.

State 02

Uncertain

Clarification request or advisory disclosure surfaced.

State 03

Degraded

Retry with fallback. Mode degrades. Reduced capability.

State 04

Escalating

Handover context package being prepared. Human informed.

State 05

Resolved

Original agent moves to monitoring role. Continuity maintained.

07 Protocol Standards

Grounded in the protocols agents actually use.

As of 2026, the agent protocol landscape has consolidated under the Linux Foundation. HAI's System Layer captures which protocol governs each phase's interactions — analogous to how TCP/IP, HTTP, and TLS layer to form the internet.

Agent ↔ Tools

MCPModel Context Protocol — tool, data, and API connectivity. Donated to AAIF (Linux Foundation) Dec 2025.

De facto standard

Agent ↔ Agent

A2AAgent-to-Agent Protocol — task delegation, multi-agent collaboration. v1.0 March 2026.

Production-ready

Agent ↔ User

AG-UIAgent-User Interaction Protocol — frontend streaming and interaction (CopilotKit).

Emerging

Observability

OTel GenAIOpenTelemetry Semantic Conventions — logging, tracing, monitoring. EU AI Act audit infrastructure.

Becoming de facto

Governance

EU AI ActArticles 9, 13, 14 — risk management, transparency, human oversight. Full enforcement Aug 2026.

Binding regulation

08 Sovereign AI Alignment

Who decides what, on whose infrastructure, under whose law.

Sovereign AI is the capability to specify, govern, and operate AI systems under your own law, on your own infrastructure, with your own accountability structures — independent of any single foreign provider, model, or cloud. HAI treats sovereignty as a design-time specification problem, not a runtime constraint. Runtime controls can be retrofitted; sovereignty intent must be specified when the matrix is designed.

Type 01

Data Sovereignty

Where data is stored and processed, the jurisdiction whose law governs it, and cross-border transfer obligations.

System Layer · Data Layer Needs · Sovereignty Boundary

Type 02

Compute Sovereignty

The physical and jurisdictional location of inference, training, and fine-tuning compute.

System Layer · Physical/Infra · Sovereignty Boundary

Type 03

Model Sovereignty

Provenance, licensing, and ownership of the model — domestic-trained, open-weight, foreign foundation, or sovereign fine-tune.

Agent Layer · Agent Type · Sovereignty Boundary

Type 04

Regulatory Sovereignty

Which legal regime governs the deployment — AI-specific law, sectoral rules, and cross-border data transfer obligations.

System Layer · Regulatory & Sovereignty Classification

Type 05

Operational Sovereignty

Domestic workforce, supply chain, and operational control of the agent system across its lifecycle.

System Layer · Compliance & Governance

Type 06

Security & Trust Sovereignty

Independent verifiability of identity, authentication, audit, and observability — without reliance on a single foreign provider.

System Layer · Protocol Layer · Compliance & Governance

The Sovereignty Boundary dimension

A new System Layer dimension specified per step or stage, documenting five elements: data residency jurisdiction, compute residency requirement, model provenance constraint, cross-border data flow permission, and accountable jurisdiction. Where a phase has no sovereignty constraints, it is marked "Non-sovereign" with a brief justification — an explicit audit artefact, not a blank cell.

WHY HAI SUITS SOVEREIGN AI → Three-layer separation lets sovereignty be specified per layer, not per system — a foreign foundation model with in-jurisdiction inference and native-language UX. · The framework is technology-agnostic, so a matrix can be retargeted to a sovereign stack by changing only System Layer cells. · Existing HAI artefacts — failure protocol, role taxonomy, memory architecture — already produce the layered-accountability and right-of-erasure documentation sovereign oversight bodies require.

09 Use Cases

Worked examples in both modes.

Case 01

SaaS Onboarding

Initial configuration step with Advisory-mode agent recommending sensible defaults from team-size and industry signals. Failure-handover triggered after two recommendation rejections.

Mode → User Journey · Step-based

Case 02

E-Commerce Lifecycle

Awareness through Retention. Agent mode escalates from Assistive to Autonomous as trust accumulates. A2A pipelines coordinate recommendation, checkout, fraud detection, and renewal agents across stages.

Mode → Customer Journey · Stage-based

Case 03

Enterprise Support

Triage, Resolution, and Escalation steps. Orchestrated pipeline of Triage, Classification, Routing, Resolution, Verification, and Reply Draft agents — with structured Human Liaison handover when confidence drops below 65%.

Mode → User Journey · Multi-agent

10 Read & Download

The full paper. The templates. Supporting materials.

↓ Whitepaper · Sovereign AI Extension PDF · 3-Layer Edition · Final v2 → ◎ Zenodo Record · DOI 10.5281/zenodo.20389202 → ⊞ Journey Matrix Template XLSX · incl. Sovereignty Boundary → ⌖ ORCID Record 0009-0007-3713-7704 →

11 How to Cite

BibTeX, for your reference manager.

@techreport{pothiraj2026hai,
  author      = {Pothiraj, Anandakumar Muniasamy},
  title       = {{Human-Centred Agentic Intelligence (HAI):
                  A Three-Layer Framework for Designing
                  Human-Agent-System Journeys}},
  institution = {Independent Researcher},
  address     = {London, United Kingdom},
  year        = {2026},
  month       = may,
  type        = {Whitepaper},
  version     = {v2 --- Sovereign AI Extension},
  doi         = {10.5281/zenodo.20389202},
  note        = {ORCID: 0009-0007-3713-7704},
  url         = {https://doi.org/10.5281/zenodo.20389202}
}

Plain text Pothiraj, A. M. (2026). Human-Centred Agentic Intelligence (HAI): A Three-Layer Framework for Designing Human-Agent-System Journeys (Version v2, Sovereign AI Extension) [Whitepaper]. Zenodo. https://doi.org/10.5281/zenodo.20389202

12 Author

Anandakumar
Muniasamy Pothiraj

Independent Researcher · London, UK

orcid.org/0009-0007-3713-7704

Andy is an independent researcher developing methodology at the intersection of human-centred design and agentic AI systems. The HAI framework is the first structured methodology purpose-built for designing human–agent–system journeys, treating the agent as a first-class journey participant rather than as a tool layered onto an existing experience.

The framework's design-time artefacts — the Unified Three-Layer Matrix, Agent Mode Taxonomy, Goal Failure Protocol, Multi-Agent Role Taxonomy, and the Sovereignty Boundary dimension — are intended as a bridge between product, design, engineering, CX, and compliance teams under emerging governance regimes including the EU AI Act and NIST AI RMF.