What determines WhatsApp AI agent cost in South Africa?

01

Define the first viable outcome before asking for a price

“An AI agent on WhatsApp” could mean answering service questions, collecting enquiries, retrieving authenticated account data or making controlled changes in an ERP. Those are different products.

Use the WhatsApp capability guide to choose one of these boundaries. A narrow end-to-end release is more useful than several half-connected features.

02

Separate implementation from the cost of operating it

Total ownership cost for a chosen period = initial delivery + recurring platform costs + measured usage + human operations + ongoing change. Define the period and avoid counting the same supplier allowance twice.

• Initial delivery: discovery, content preparation, conversation design, integrations, testing and launch readiness.
• Channel and platform: messaging charges, provider subscriptions, inbox seats and account administration.
• AI usage: model calls, context length, retrieval, media processing and retries.
• Application operation: hosting, queues, storage, monitoring and backups appropriate to the service.
• Human operations: exception handling, knowledge reviews, quality sampling and support ownership.
• Ongoing change: platform updates, new business rules, regression tests and approved feature work.

Ask which costs are included, passed through, capped or charged separately. A subscription that includes support may still exclude business-content updates or work on a changed integration.

03

Use current supplier rates, not copied pricing tables

Meta's current platform pricing page describes delivery-based charges that depend on the recipient market and message category, with free service messaging during the customer-service window and specific exceptions and volume tiers. Check the applicable rate card and conditions when preparing the estimate.

The business's South African location does not mean every recipient attracts the same rate. Separate your recipient markets and message types in the estimate. Also identify provider fees that sit outside Meta's charges.

For a rand-denominated budget, ask how foreign-currency supplier bills are converted, when rates are reviewed and who carries exchange-rate changes. This is a budgeting assumption, not a reason to convert an overseas project benchmark into a South African development price.

04

Give the delivery team evidence they can estimate

1. Supply representative customer messages and the intended resolution for each.
2. Identify the maintained source of truth for answers, availability and account status.
3. Confirm which systems have usable APIs, test environments and named owners.
4. Separate read-only features from writes that require approval or recovery.
5. Describe expected normal demand, busy periods, languages and media inputs.
6. Agree how a person receives unresolved work and when that team is available.
7. Write acceptance criteria for the first viable release and exclusions for later phases.

Where the risk is unknown, estimate a bounded investigation first. For example, proving access to one authorised ERP record may remove more uncertainty than polishing an unrestricted demo conversation.

Ask the team to identify confidence in each estimate and the evidence that would change it. Delivery timing should follow the actual dependencies and agreed launch criteria rather than a generic calendar promise.

05

Account for AI-assisted development without assuming a fixed saving

Tools such as Cursor, Claude and Codex may assist the delivery team with implementation, test drafting and repetitive changes. Ask the team to estimate the work using its real toolchain and validated delivery approach. There is no project-independent percentage saving to apply.

Faster code production does not resolve incomplete business rules, unavailable APIs, account approvals or the need to test real customer journeys. Evaluate the completed outcome, not how quickly an initial implementation appeared.

For the agent itself, Anthropic's architecture guidance recommends adding agentic complexity only when it improves the result. Fewer necessary model steps can simplify both evaluation and operation.

06

Compare like-for-like scope and ownership

Give each supplier the same brief and ask for a response against the same acceptance criteria. Clarify ownership of the WhatsApp business account, application code, integration credentials, knowledge material and operational data. Ask what can be exported if you later change supplier.

Compare usage assumptions, overage handling, support coverage, service dependencies and the process for approving additional work. Confirm whether staff training, a working human handover and security verification are included.

A useful quote explains what happens when the agent cannot resolve the request. A low initial figure is not comparable if your team must later supply an inbox, rebuild the integration or create the missing operational process.

07

Measure cost per correctly resolved outcome

Use the pilot to measure a complete operating picture. Count successful resolutions, repeat contacts, incorrectly routed requests, staff follow-up effort and total attributable run cost. Keep unresolved and abandoned conversations visible.

Operating cost per verified resolution = attributable operating cost ÷ verified resolved requests. Track quality alongside that measure so a cheaper response does not hide more errors. Do not claim a unit cost when there are no verified resolutions.

Before building the next feature, ask whether the current service resolves the intended problem, where users still need help and what the business can realistically maintain. Use the AI use-case check to make that starting boundary concrete.

Questions

Frequently asked questions