How We Build Custom AI Agents for SEA Businesses: Process, Stack, and What to Expect

Most AI agency content describes what AI agents can do in theory. This article describes what we actually build, how we build it, what the process looks like from the client's side, and what you walk away with at the end of an engagement.

If you are evaluating AI agent development for your B2B business in Thailand or Southeast Asia, this is the most honest account we can give of what the work involves.

Table of Contents

• [Our Design Philosophy: Narrow and Deep, Not Broad and Shallow](#our-design-philosophy-narrow-and-deep-not-broad-and-shallow)
• [The Tech Stack We Build On](#the-tech-stack-we-build-on)
• [Our Build Process: Phase by Phase](#our-build-process-phase-by-phase)
• [What Clients Receive at the End of an Engagement](#what-clients-receive-at-the-end-of-an-engagement)
• [The Agents We Build Most Often](#the-agents-we-build-most-often)
• [What Makes an AI Agent Project Succeed or Fail](#what-makes-an-ai-agent-project-succeed-or-fail)
• [Pricing and Timeline](#pricing-and-timeline)
• [Frequently Asked Questions](#frequently-asked-questions)

Our Design Philosophy: Narrow and Deep, Not Broad and Shallow

The most common mistake in AI agent development is building a general-purpose agent that can "do anything." In practice, these agents do nothing well because the goal is too vague, the tools are too many, and the success criteria are undefined.

We build the opposite: narrow agents with a single well-defined job.

A lead qualification agent qualifies leads — it does not also handle support tickets or generate reports. A support triage agent handles first-response classification and routing — it does not also research competitors. This narrowness is not a limitation; it is what makes the agent reliable enough to run in production.

When a client needs multiple agent capabilities, we build a system of narrow agents — each handling its specific function, orchestrated by a workflow layer that routes tasks to the right agent.

The Tech Stack We Build On

Layer Tool Why

Reasoning engine Claude (claude-sonnet-4-6) Best instruction-following, long context, strong Thai language performance

Workflow orchestration n8n Self-hosted option for data residency, cost-efficient at volume, native LINE support

Database / memory Supabase (PostgreSQL) Persistent storage for agent memory, conversation history, and task logs

Lead enrichment Apollo.io Company and contact data for Thai and SEA businesses

CRM integration HubSpot or Pipedrive Where agent outputs are logged and sales team receives results

Communication LINE Messaging API, Gmail, SendGrid Thai-first communication channels

Document processing Claude Vision + n8n PDF, image, and scanned document reading

Monitoring Custom n8n error workflows + Slack alerts Production reliability

We do not build on proprietary agent frameworks that lock clients into our infrastructure. Every agent we build runs on open, well-documented tools — clients can audit, modify, or hand off to another team.

Our Build Process: Phase by Phase

Phase 1: Discovery and Scoping (Week 1)

We start by understanding the workflow the agent will replace or augment.

What we map:

• The current manual process, step by step
• Where the process breaks, slows, or creates errors
• The inputs the agent will receive (what data, in what format, from what source)
• The outputs the agent must produce (what deliverable, in what format, to what destination)
• The success criteria: how do we know the agent did the job correctly?

What we produce:

• A workflow diagram showing the current state vs. the agent-powered future state
• A tool inventory: which APIs, databases, and systems the agent needs access to
• A risk assessment: where the agent is most likely to make mistakes and how we handle those
• A signed-off scope document before any build begins

This phase prevents the most common engagement failure: building the wrong thing.

Phase 2: Prototype (Week 2)

We build a functional prototype using synthetic data — no real customer information is processed during this phase.

The prototype covers:

• The agent's core reasoning loop
• Integration with the primary data source (CRM, email, LINE, or document input)
• The primary output (CRM update, draft document, notification, or database record)

At the end of Week 2, we demo the prototype to the client with a structured walkthrough of 10–15 test scenarios, including edge cases.

Client involvement: 1-hour demo review session. Feedback captured and incorporated before moving to production build.

Phase 3: Production Build (Weeks 3–4)

The prototype is hardened into a production system:

• Error handling: every step has explicit failure handling — if an API call fails, the agent retries, then falls back, then alerts a human
• Logging: every agent action is logged with a timestamp, input, output, and success/failure status
• Human review checkpoints: for consequential outputs (sending a message, updating a financial record), a human approval step is built in by default
• Rate limiting: API calls are managed to avoid hitting platform limits
• PDPA compliance: consent handling, data minimization, and opt-out logic are built in for any agent that processes personal data

Phase 4: Testing (Week 4 — parallel with production build)

Testing covers three dimensions:

Happy path testing: 50+ scenarios where inputs are clean and expected. Measures accuracy of classification, quality of outputs, and correct tool usage.

Edge case testing: inputs that are ambiguous, incomplete, in unexpected language, or outside the agent's defined scope. Measures graceful degradation — does the agent fail safely (route to human) or unsafely (produce a confident wrong answer)?

Load testing: what happens at 10x normal volume? Does the workflow queue correctly? Do rate limits trigger? Does the logging system keep up?

Phase 5: Deployment and Handover (Week 5)

Live deployment to the client's environment:

• All API credentials are in the client's accounts — no dependency on our infrastructure
• n8n workflows are deployed to the client's n8n instance (cloud or self-hosted)
• Database tables are created in the client's Supabase project
• A 60-minute team training session walks through: how the agent works, how to monitor it, how to update the knowledge base, and how to contact us if something breaks

Handover package:

• Complete workflow documentation with annotated diagrams
• Prompt documentation (the exact instructions given to Claude for each step)
• Runbook: common issues and how to resolve them
• Source access to all n8n workflows

What Clients Receive at the End of an Engagement

Clients own everything:

• The n8n workflows (exported JSON, importable to any n8n instance)
• The prompts used to instruct Claude at each step
• The database schema and all historical data
• Complete documentation of how the system works

We do not create lock-in. If a client wants to move the agent in-house or to another vendor, they have everything they need to do so.

The Agents We Build Most Often

Lead qualification agents — the most common first engagement. Research + score + notify + log. Typically 2–3 week build.

Customer support triage agents — email and LINE first-response handling. Typically 3–4 week build including knowledge base structuring.

Document processing agents — invoice, contract, or form extraction and routing. Typically 2–3 week build.

Proposal generation agents — discovery notes in, first-draft proposal out. Typically 3 week build including template design.

Business intelligence agents — weekly automated reports from multiple data sources. Typically 2 week build.

What Makes an AI Agent Project Succeed or Fail

Success factors:

• Clear, narrow goal defined before build begins
• A human who owns the agent and is accountable for its outputs
• A knowledge base or data source that is accurate and maintained
• Willingness to run with human-in-the-loop first and automate progressively

Failure factors:

• Undefined or changing scope during the build
• Expecting the agent to handle everything from day one without a monitoring period
• No defined success criteria — "make it useful" is not a goal
• Insufficient attention to the knowledge base or input data quality

The technology is not the constraint in most failed AI agent projects. The constraint is process clarity and organizational readiness.

Pricing and Timeline

Agent Type Timeline Investment Range

Single focused agent (lead qual, doc processing) 2–3 weeks $2,000–$3,500

Customer support agent with LINE integration 3–4 weeks $3,000–$5,000

Multi-agent system (2–3 specialized agents) 5–7 weeks $5,000–$9,000

Monthly maintenance retainer Ongoing $300–$600/month

All engagements include the handover package described above. Maintenance retainers include monitoring, monthly optimization reviews, and knowledge base updates.

[See full pricing and packages →](/pricing/)

Frequently Asked Questions

Do you work with businesses outside Thailand? Yes. We work with B2B companies across Southeast Asia — Singapore, Malaysia, Vietnam, Indonesia, and the Philippines. English is our working language for cross-border engagements; Thai for domestic Thai clients.

Can you audit our existing AI implementation before recommending a new build? Yes. We offer a 2-hour AI workflow audit as a standalone engagement. We review what you have, identify gaps and inefficiencies, and recommend a prioritized roadmap. This is often the right starting point for businesses that have already experimented with AI tools.

Do you use OpenAI or Anthropic? We build on Claude (Anthropic) as our primary reasoning model. For clients with existing OpenAI integrations we want to preserve, we can build on GPT-4 Turbo — the architecture is the same; only the model API changes.

What ongoing support do we get after handover? Handover includes 30 days of email support for questions about the delivered system. After that, clients can engage a monthly maintenance retainer for monitoring, optimization, and expansion work.

Related reading: [What Is an AI Agent?](/blog/what-is-an-ai-agent/) | [AI Agent vs Chatbot for B2B](/blog/ai-agent-vs-chatbot-b2b/) | [Custom AI Agent Services](/services/custom-ai-agents/)