AI agent development

In short: FYAdvisers builds AI agents that work out the steps of a task and act across your tools, with narrow permissions, a person approving actions that matter and a log of every run. Where the steps never change, a fixed workflow is cheaper and easier to test.

AI agent development means building software that uses a language model to work out the steps of a task and carry them out across your tools: reading a request, looking up records, updating a system and preparing a reply. FYAdvisers is a Singapore technology consultancy that builds agents for small and mid-sized businesses. Every agent we build has a narrow job, only the permissions that job needs, an approval step before actions that matter, and a log of everything it did. We build on n8n, Make or custom code, depending on the task and who will maintain it.

We will also tell you when you do not need an agent at all.

What an agent does that a workflow does not

A normal workflow automation follows fixed steps: when a form arrives, add a row and send an alert. An agent is given a goal and a set of tools, and decides which tools to use and in what order. That helps when requests vary, such as a customer message that might be a booking change, a complaint or a new order, each needing different lookups.

The flexibility has a cost. Agents are slower, cost more per task to run and are harder to test than fixed workflows, because the same input can lead to different steps.

An example from our work

For an individual PropNex property agent, we built a WhatsApp deal-conversation workspace. It groups chats into deals, extracts the people, details and outstanding items, and prepares reply drafts and English to Chinese translations. It also has Google Calendar tools. The agent reviews every draft and decides what to send.

Typical uses for Singapore B2B companies

• Inbox and enquiry handling. Read incoming messages, identify the request, look up the customer and prepare a reply or a task for the right person.
• Operations follow-up. Check which jobs, quotations or orders are waiting, chase missing information and prepare reminders for approval.
• Document-led tasks. Read a tender, contract or form, extract what matters and update the right records. See AI document processing.
• Sales research. Research an account, summarise recent events and draft outreach for approval. See AI lead generation.
• Internal questions. Answer staff questions from your own documents and records, with links to the source.

Approvals, guardrails and logging

These are part of the design from the first day, not an add-on.

• Narrow permissions. The agent gets read access by default and write access only where the task needs it, using its own service account rather than a person’s login.
• Approval steps. Sending to customers, changing financial records, deleting data or committing your business waits for a person. The approval screen shows what the agent wants to do and why.
• Guardrails. Rules check inputs and outputs: required fields present, amounts within limits, no personal data sent where it should not go. Limits on steps, retries and spending stop runaway loops.
• Logging. Every run records the request, the steps taken, the tools called, what was approved and by whom. That is how errors are found and the instructions improved.
• A way to stop it. Each agent can be paused without breaking the rest of your systems.

How a project runs

1. Discovery. We map the task as your team does it today, including the exceptions, and decide which parts need judgement and which follow rules.
2. Prototype with your real data. We run the agent on past cases in a test environment and compare its steps and outputs with what your team actually did.
3. Build. We connect the tools, set permissions and approvals, and build the review screen into the place your team already works, often through a system integration or a small custom application.
4. Human review. The agent starts with every action needing approval. Automatic actions are added only for low-risk steps you choose.
5. Monitoring. We review logs, failures and rejected actions, and adjust tools, rules and instructions.

n8n, Make or custom code

• n8n has a large set of app connections and AI agent features. It can run on n8n’s cloud or on your own server. n8n’s pricing page listed cloud plans from 20 euros a month (Starter) and 50 euros a month (Pro), billed annually, when we checked in October 2026, and says a self-hosted version is available on GitHub.
• Make is a visual automation platform with AI features. Its pricing page listed a free plan with 1,000 credits a month in October 2026; check the page for paid plans.
• Custom code suits agents that need their own interface, fine-grained permissions, high volume or deep integration with your own systems. It costs more to build and needs a developer to maintain.

Off-the-shelf agent products are worth considering first. Microsoft Copilot Studio suits teams already on Microsoft 365, Salesforce Agentforce suits Salesforce users, and Zapier, n8n and Make offer agent features that a capable operations person can set up.

What it costs

We outline our own fees after a conversation. The main cost drivers are:

• Number of tools and systems. Each connection needs permissions, error handling and testing.
• Risk of the actions. Agents that change records or contact customers need more guardrails, review screens and testing.
• Model usage. Agents call a model several times per task, and providers bill per token, so running cost grows with the number of steps and volume.
• Platform fees. Automation platform plans, hosting and any paid data sources.
• Ongoing care. Apps change their APIs, models are updated and new edge cases appear.

Data and governance

PDPC’s Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems, issued on 1 March 2024, explain how the PDPA’s consent, notification and accountability obligations apply when AI systems use personal data. IMDA and the AI Verify Foundation released the Model AI Governance Framework for Generative AI on 30 May 2024, covering dimensions such as accountability, data, trusted development and deployment, incident reporting, testing and assurance, and security. Agents that act on your systems touch most of these, so we document what data each agent can reach, where it is processed and who answers for its actions.

When not to use an agent

• The steps are always the same. Use a fixed workflow. It is cheaper and easier to test.
• The decision is predictable. Use rules.
• A mistake would be costly and hard to undo, and a person would need to check every step anyway. A drafting tool may be enough.
• The task happens a few times a month. Setup and upkeep may cost more than the time saved.
• Your data is scattered or out of date. Fix the records first, or the agent will act on wrong information.

Who this suits

• Teams handling varied requests across several tools.
• Businesses that have already automated the predictable steps and are left with the ones needing judgement.
• Owners who want a person approving important actions, with a record of what happened.

What to prepare

• The task, described step by step as your team does it, including the awkward cases.
• The tools involved and who has admin access.
• Which actions are safe to automate and which must be approved.
• A set of real past examples, with personal details removed.

Get in touch and we will tell you whether an agent, a workflow or neither is the right fit.