Case Study: Designing an Agentic Coaching System for Iranians Living Abroad
Article hnarimani@gmail.com July 13, 2026

Case Study: Designing an Agentic Coaching System for Iranians Living Abroad

The real problem for Iranians living abroad is not a lack of information. It is turning fragmented information into an executable decision in an unfamiliar system.They face questions about residency, work, tax,...

The real problem for Iranians living abroad is not a lack of information. It is turning fragmented information into an executable decision in an unfamiliar system.

They face questions about residency, work, tax, housing, language, healthcare, community, and career direction. The answers are scattered across official portals, Telegram groups, videos, forums, and personal anecdotes. A human coach can help, but capacity does not scale. A general chatbot can answer, but it does not own context, risk, or follow-through.

This is not a chatbot problem. It is an agentic decision-system problem.

Coaching Is Not a Conversation

Immigrant coaching is often designed as a simple exchange. A person asks. The system responds. That model breaks under real conditions.

The user rarely needs a standalone answer. They need a decision path shaped by their country, legal status, available time, financial constraints, language ability, and immediate goal.

What Most Teams Get Wrong

Most teams frame the product as “answering immigration questions.” That creates a searchable pile of advice, not an operational support system.

A better framing is simple: reduce uncertainty, identify the next action, and escalate high-risk cases to a qualified human.

A useful agent does not make the decision for the user. It reduces the cost of making a bad one.

Why This Matters

Iranians abroad are not one user segment. A graduate student, skilled worker, founder, family member, and recent arrival have different constraints and different failure modes.

A single generic flow reduces trust. Full one-to-one human support raises operating cost. An agentic system exists to manage that tension without pretending automation is magic.

System Objectives

  • Turn vague questions into structured decision problems
  • Retain useful context without repeatedly interrogating the user
  • Ground sensitive answers in traceable, current sources
  • Produce short action plans rather than long advice lists
  • Escalate legal, financial, and health-related risk to humans
  • Capture feedback that improves workflows and knowledge quality

From Chatbot to Decision Architecture

A sound System Architecture does not treat one model as an all-purpose expert. Each responsibility needs a clear boundary.

For this case, the system uses five layers. The separation makes failures easier to observe, audit, and repair.

Layer One: Context Intake

The intake agent gathers only the context that changes a decision. Country, residency status, deadline, primary goal, language level, and key constraint are usually enough to begin.

It should not force users through a long form. Every question must earn its place by changing the next step.

Layer Two: Problem Classification

Each request should become an operational category: employment, residence, housing, tax, cultural adjustment, business setup, or mental-health support.

Classification enables routing. Without it, a large language model must guess everything at once.

Layer Three: Verified Knowledge Retrieval

The system should never present legal, visa, or tax claims as certain from model memory. This layer should use Retrieval-Augmented Generation, or RAG, against a curated knowledge base.

Every document needs a source, validity date, geographic scope, owner, and risk level. In sensitive domains, an unsupported answer is a product defect.

Layer Four: Action Planning

The final response should become a bounded plan. Three relevant actions are usually more valuable than ten generic recommendations.

The planner must distinguish between an immediate action, further research, and human escalation. That is where a conversational tool becomes an operating system.

Layer Five: Quality Control and Escalation

A supervisory agent checks for unsupported claims, excessive certainty, conflict with the user profile, and high-risk content.

When risk crosses a threshold, the system should state its limit clearly. It should then route the person to the appropriate professional pathway.

The PATH Framework

PATH is a practical framework for designing coaching workflows where personalization, safety, and scale must coexist.

ComponentPurposeOperational Output
ProfileCapture the real user situationA decision profile with country, goal, constraints, and life stage
AssessEvaluate the request and riskIntent classification, urgency level, and escalation need
TraceFind trustworthy evidenceOfficial sources, internal knowledge, and confidence boundaries
Hand-offMove from insight to executionAn action plan, reminder, follow-up, or specialist referral

PATH is not a conversation framework. It is an operational control model.

A Practical Scenario

Consider an Iranian professional in Germany with a new job offer. They are uncertain about residency renewal and health insurance requirements. A general chatbot may return links and broad explanations.

An agentic system first identifies the residence permit type, expiry date, employment contract type, and whether the job change triggers a reporting requirement. It then retrieves only country-specific, permit-specific material.

A useful output might be:

  1. Confirm your current permit type and expiry date using the residence documentation or official portal
  2. Check the employer-change condition against the official rules for that permit category
  3. If the condition remains unclear, escalate with a prepared document checklist to an immigration professional

The distinction matters. The system did not merely generate an answer. It created a lower-risk path.

Chatbot Versus Agentic System

DimensionGeneral ChatbotAgentic Coaching System
Context memoryLimited or temporaryStructured, governed user profile
EvidenceOften unclearRetrieved, dated, and reviewable sources
Next stepUsually broad adviceA specific action, follow-up, or escalation
Risk handlingPrompt-dependentRules, thresholds, and human-in-the-loop controls
Product learningUnstructured conversation dataEvents, feedback, failure analysis, and workflow refinement

Implementation Guidance

Do not start with a complex multi-agent stack. Start with one high-demand, low-risk workflow, such as job-search preparation for one country.

Set Product Boundaries First

  • Constrain the initial geography and user segment
  • Identify the three highest-frequency problems from real interviews and usage data
  • Define escalation policies for legal, financial, and medical topics from day one
  • Define answer-quality metrics before selecting a model

Build Governed Knowledge

Internal knowledge needs versioning. Each source should have an owner, review date, primary reference, and defined applicability.

If the team cannot explain where an answer came from, it should not publish that answer in a high-stakes context.

Design the Orchestration

Orchestration determines which agent acts, when it acts, and what data it can access. It matters more than the model brand.

A basic flow can include intent detection, risk screening, knowledge retrieval, action planning, and final review. Simpler flows are easier to observe and improve.

Measure Operations, Not Activity

Do not optimize for conversation count alone. That metric often rewards longer, less useful responses.

  • Suggested-action completion rate
  • Correct human-escalation rate
  • Rate of answers backed by valid sources
  • Repeat requests for the same unresolved issue
  • Time from question to first useful action
  • Human reviewer correction rate

Operational Reality

The difficult part is not the language model. The difficult part is knowledge maintenance, accountability boundaries, and human operations design.

Immigration rules change. Official sources can be difficult to interpret. Users do not arrive with neatly categorized problems. The system must be designed for that mess rather than hiding it.

Human-in-the-Loop

Human-in-the-loop does not mean routing every conversation to a person. It means placing human judgment at the highest-value and highest-risk bottlenecks.

Humans can review legally consequential outputs, approve new sources, and turn recurring failures into product rules.

Common Failure Modes

Building One Agent for Everything

An agent responsible for work, residency, tax, mental health, and networking will produce shallow and difficult-to-govern outputs. Separate responsibilities.

Treating Community Anecdotes as Evidence

Community experience can reveal user pain. It is not a legal authority. Use it as a signal, not as the decision source.

Personalizing Without Clear Consent

Residency, income, health, and family details are sensitive. Data minimization and explicit consent belong in the architecture, not in a forgotten legal footer.

Measuring Satisfaction Instead of Outcomes

A user may appreciate a polite response and still make no progress. The primary measure should be reduced uncertainty and forward motion.

Trade-offs and Constraints

More personalization can create more value, but it also increases privacy risk and data-management cost. More agent autonomy can improve speed, but it can weaken control.

Supporting more countries expands the market, but makes knowledge quality harder to maintain. Early product scope should be intentionally narrow.

DecisionBenefitCost
Support multiple countriesLarger addressable marketFragmented knowledge and harder quality assurance
Deep user memoryMore precise guidancePrivacy risk and data complexity
Greater automationLower operating costHigher impact from preventable errors
Broad human reviewHigher trust and safetyMore cost and operational capacity required

Key Takeaways

  • Immigrant coaching should be designed as a decision system, not a question-answering tool
  • Profile, risk assessment, traceable evidence, and next action are the core components
  • RAG without knowledge governance is simply a faster way to produce uncertain answers
  • Human escalation should be a designed capability, not an emergency escape hatch
  • Narrow initial scope is more valuable than broad, unreliable coverage
  • The most useful metric is user progress, not conversation length

FAQ

What is an agentic system for immigrant coaching?

An agentic system combines specialized agents, governed knowledge, risk rules, and operational workflows to move a user from an ambiguous question to an action or an appropriate referral.

Can an AI agent replace an immigration lawyer or advisor?

No. An agent can organize information, create an action path, and identify when expert support is needed. Legal decisions and regulated advice should remain with qualified professionals.

How many agents are needed for a first version?

Three roles are enough for an initial version: context intake, knowledge retrieval, and quality control. Early complexity rarely improves quality.

What data matters most for personalization?

Collect data that changes the next decision: country, residency status, goal, deadline, and key constraints. More data does not automatically create more value.

How should answer quality be evaluated?

Evaluate source traceability, profile fit, risk classification, clarity of the next action, and human correction rate. Immediate user satisfaction alone is not sufficient.


A good system does not make users dependent on answers. It moves them toward decisions they can execute and verify.

Sources [1] EMPATHIA: Multi-Faceted Human-AI Collaboration for ... https://arxiv.org/html/2508.07671v1 [2] Leveraging Ai Agents In Design-Based Teaching: Insights From Utilizing Synthetic Users To Design Tool Support For Immigrants https://aisel.aisnet.org/treos_ecis2026/104/ [3] A Case Study to Foster the Individual-level DMC Through ... https://oiccpress.com/jee/article/download/18704/19970 [4] United States Citizenship and Immigration Services – AI Use Cases https://www.dhs.gov/ai/use-case-inventory/uscis [5] [PDF] AI-Driven Conversational Models for Supporting Migrant Career ... https://www.ijsreat.com/archiver/archives/ai_driven_conversational_models_for_supporting_migrant_career_guidance_and_labour_market_integration_a_scoping_review.pdf?v=1753779431 [6] how organizations are using AI agents https://www.youtube.com/watch?v=g13UP2B29Ek [7] Agentic AI-Led Migration Assistance-Fast-Reliable-Empowering https://www.datamatics.com/resources/case-studies/agentic-ai-led-migration-assistance-fast-reliable-empowering [8] AI Agents for International Student Services & Study Abroad https://www.autohive.com/use-case/international-student-services-and-study-abroad [9] [PDF] Artificial intelligence and automation in the migration governance of ... https://files.eric.ed.gov/fulltext/EJ1416666.pdf [10] AI-Coach with Artificial Emotional Intelligence (AEI) - MIT Solve https://solve.mit.edu/solutions/89558 [11] Mobile Applications Aiming to Facilitate Immigrants’ Societal Integration and Overall Level of Integration, Health and Mental Health: Does Artificial Intelligence Enhance Outcomes? https://docs.iza.org/dp13933.pdf [12] AI Agent Case Studies | Success Stories | BuildAiAgents https://buildaiagents.app/case-studies [13] AI Agent Useful Case Study – Three Real World Examples https://arize.com/ai-agent-useful-case-study/ [14] AI Agent Useful Case Study: A Day in the Life of an ... https://www.druidai.com/blog/ai-agent-useful-case-study-day-in-the-life-of-ai-agent [15] 10 AI Agent Use Cases with Real Results https://www.intellectyx.com/ai-agent-useful-case-study-10-real-world-examples-applications/

Ready to apply this in your own product? Book a Strategy Call and get a clear roadmap for your next sprint.

Comments (0)

Be the first to leave a comment.
Login / Sign up