Custom WhatsApp AI Agent: 6 Things You Need First

What does a custom WhatsApp AI agent actually require?

• One completed outcome: Define the job in operational terms, such as confirming a delivery status from the approved record and escalating exceptions, rather than “answer customer questions”.
• A source of truth: Name where prices, stock, order status, policies and customer records live, who keeps each source current, and what wins when records disagree.
• Official channel access: Use the WhatsApp Business Platform with clear ownership of the business account, phone number, message templates, permissions and support relationship.
• Connections to the work: Give the agent only the access it needs to read or update Xero, an order system, a calendar or a shared sheet. The integration is what turns a reply into completed work.
• Boundaries and approvals: Write down what the agent may do alone, what needs a person, and what it must refuse. Refunds, payment changes, unusual discounts and ambiguous identities are common approval points.
• A tested exception path: Design the handover before launch. The staff member should receive the conversation, relevant records, the attempted action and the reason the agent stopped.
• An operating owner: Someone must review failures, keep source data current, approve workflow changes and retest the agent after policies or integrations change.

Which workflow should an SME automate first?

Choose a frequent, repetitive workflow with a clear finish and a manageable downside when something goes wrong. Order-status enquiries, appointment qualification and collecting complete details for a quotation are usually better starting points than open-ended complaints or discretionary refunds.

Trace the work from the first customer message to the final record. List every lookup, copy-and-paste step, decision, approval and follow-up. If the map ends at a reply, a standard platform may be enough; if it crosses tools and business-specific rules, the custom part has a reason to exist. Our platform versus custom agent guide covers that decision in more detail.

Define success as completed work: a supported answer, a correct record and a clean handover where needed. Reply volume and response speed matter, but neither proves the underlying job was done correctly.

What data and integrations need to be ready?

The agent needs a narrow, current set of approved sources—not every folder the company owns. For a delivery workflow, that might be the order record, payment status and dispatch status. Product brochures in old email threads should not compete with the current catalogue.

Each integration needs an access rule. Read-only access is enough for many lookups. Any write action should be limited to the exact fields required, logged, and reversible where practical. Do not give broad access simply because it makes the first demo easier.

Messy data does not rule out automation, but the exception must be visible. When an order number is missing or two customer records conflict, the agent should ask a precise question or hand over rather than guess.

How do PDPA and WhatsApp rules affect the build?

Meta's WhatsApp Business Messaging Policy says businesses may use automated replies within the 24-hour customer-service window only when they also provide a prompt, clear and direct escalation route; outside that window, businesses may send only approved message templates. Build those behaviours into the workflow instead of treating them as footer text.

Singapore's PDPC says bespoke AI service providers acting as data intermediaries have Protection and Retention obligations under the PDPA. Before live data is connected, document the purpose, fields, access, processing locations, retention, deletion, incident support and exit process.

IMDA's Model AI Governance Framework for Agentic AI recommends bounding an agent's powers, setting meaningful human approval points, applying controls throughout its lifecycle, and training end users. For an SME, that becomes a practical permissions table, an action log, an approval queue and staff who know when to intervene.

How should you test a custom WhatsApp AI agent?

• Use real patterns, not live personal data: Build an anonymised test set from ordinary chats and the awkward cases staff already see: missing references, mixed language, voice notes, changed addresses and conflicting records.
• Write the expected outcome first: For each case, state the approved source, correct action, required approval and acceptable customer response before the agent runs.
• Test safe failure: Include unavailable systems, stale data, unsupported requests and a customer asking for a person. A dependable agent should stop clearly, not improvise.
• Review actions as well as words: Check the records read, changes proposed, approvals requested and handover context. A fluent answer can still accompany the wrong operational action.
• Retest after every material change: A new policy, catalogue structure or integration can break cases that worked before. Keep a repeatable test set and run it before expanding scope.

Worked example: a Singapore distributor

Consider a distributor whose customers ask for order and delivery updates on WhatsApp. Today an administrator identifies the customer, searches the order sheet, checks payment in Xero, asks the warehouse about unclear dispatch records, types the reply and notes the follow-up.

The first version of a custom agent should own only that path. It matches the customer and order, reads approved order, payment and dispatch fields, drafts or sends a supported update, and hands unclear cases to the administrator with the relevant records attached. It cannot change payment status, promise stock, approve a refund or invent a delivery date.

The team compares the pilot with the manual baseline: how much staff handling remains, which enquiries finish without rework, where unsupported claims occur, and whether handovers contain enough context. Once the result is dependable, the business can consider adjacent tasks such as delivery rescheduling; it should not add them merely because the agent can chat about them.

This is the kind of end-to-end workflow behind Biks customer operations. In the existing Rosalie case study, daily WhatsApp order admin fell from more than four hours to 20 minutes.

What drives the cost, and how should you compare it?

Cost is driven by workflow scope, the condition of the source data, the number and difficulty of integrations, message volume, exception complexity, testing, reporting and ongoing maintenance. An agent that answers from one approved policy is a smaller build than one that checks payment, updates an order and coordinates delivery.

Compare the full operating cost with the current manual process and with hiring another administrator. Measure staff time spent reading, searching, retyping, following up, supervising and correcting the target workflow; then include the work that remains after automation. A low software fee is not cheap if staff still resolve every exception or maintain duplicate records.

Do not compare proposals until they separate implementation, WhatsApp message charges, ongoing support, integration maintenance and optional changes. Never accept a savings estimate that assumes perfect automation or ignores the human review the workflow genuinely needs.

What should be ready before launch?

If any warning sign remains, narrow the scope rather than launch a vague agent. A small workflow that finishes reliably creates more value than a broad assistant that sends polished replies and leaves the team to repair the work behind them.