Meta Business Agent Pricing: Route Conversations Before You Automate

What Changed With Meta Business Agent Pricing?

Meta introduced Business Agent on 3 June 2026 as AI that can answer business-specific questions, recommend products, book appointments, qualify leads, support sales and let a team decide when a person steps in. Meta said businesses could get started for free and that paid subscription offerings would follow in the coming months.

Business Today and TechTimes subsequently reported that token-based billing began on 1 August 2026 for Meta Business Agent use on WhatsApp. Both reported a global rate of US$2 per one million tokens. The official June announcement confirms the move towards paid access, but the specific August rate is supported here by those two media reports rather than the announcement itself.

The rate may look small in isolation. The operating issue is larger: longer, more complex conversations consume more processing and often carry more customer-service risk. A team needs to know whether AI is removing repetitive work or merely delaying the cases that need judgement.

Why Should Conversation Routing Come Before Automation?

A single chat channel can receive very different tasks. A customer may ask for opening hours, check stock, request a recommendation, dispute a payment or raise a complaint. They all arrive as messages, but they do not have the same data needs, consequences or review requirements.

If every message enters one automated lane, three problems can grow together:

• routine questions compete with urgent or sensitive cases;
• the agent may reach beyond approved information or actions;
• usage costs rise without a clear view of which conversations created value.

Meta says businesses can decide when a team member steps in. That makes handoff design a core part of implementation, not an exception added after launch.

How Can You Build a Three-Lane Routing Map?

Start with 20 recent, anonymised customer enquiries. This is an RxAI working method, not a Meta requirement or a statistical benchmark. The aim is to reveal the real mix of work before changing systems.

Classify each enquiry into one of three lanes:

1. Green — AI can answer: The answer is stable, public and low risk. Examples include opening hours, locations and basic service scope.
2. Amber — AI prepares, a person approves: The request needs order data, stock checks, contextual recommendations or a commitment. AI can collect fields and draft a summary, but a person confirms the response.
3. Red — a person takes over: The request involves payment disputes, refunds, complaints, personal information, health or legal issues, or another decision that may materially affect customer rights.

Then write one or more escalation triggers for each lane. Useful triggers include missing information, a second expression of dissatisfaction, a refund request, a delivery commitment or an answer that cannot cite approved business information.

These examples are operational guidance, not legal or industry-specific compliance advice. Adjust them for Australian Privacy Principles, consumer obligations, internal policy and the consequences of an incorrect response.

What Should the 20-Conversation Audit Record?

Use a simple spreadsheet with five columns:

• question type;
• risk lane;
• approved data required;
• human-handoff trigger;
• final handling outcome.

Review which low-risk question type appears most often. Choose one high-repeat, low-consequence category as the first trial. Avoid beginning with a workflow that changes prices, processes refunds, commits delivery dates or exposes personal data.

For the pilot, record whether the response was correct, whether the right source was used, whether the handoff happened at the intended point and whether a person had to repair the result. This gives the team evidence about service quality and workload before it expands the automation boundary.

How Should an Australian SMB Start?

Keep the first release narrow and reversible:

1. Anonymise 20 recent enquiries.
2. Assign each one to green, amber or red.
3. Document approved sources and escalation triggers.
4. Select one green question type for a pilot.
5. Test ambiguous, missing-data and unhappy-customer cases.
6. Monitor handoffs, corrections and reported usage.
7. Expand only after the owner approves the evidence.

Pricing changes can create pressure to optimise quickly. A better sequence is to establish the service boundary first, measure the pilot and then decide whether broader automation is commercially useful.

RxAI helps Australian businesses design controlled AI workflows with clear ownership, evidence and human review. Explore our automation services or contact RxAI to map a customer-service pilot.

Sources

• Meta official announcement - product capabilities, supported Meta messaging channels, free starting access, paid-offering direction and human handoff controls.
• Business Today pricing report - reported 1 August billing start, US$2-per-million-token rate and plan distinctions.
• TechTimes cross-check - independent reporting of the start date and rate.