FAQ
Answers to the questions we hear most while clients integrate Meta Business Agent (MBA). For anything not covered here, check Known limitations & workarounds or reach out to your 360dialog contact.
Getting started
Do I need a credit line to try MBA?
No — not to test it. You can complete eligibility, enable, configure, test, and go live on an allowlist (ALLOWLISTED_ONLY) without any payment method on file. A payment method is only required before you switch the agent to answer everyone (EVERYONE). See Roll out to everyone.
What payment methods are supported?
Today, a direct credit line with Meta, managed via Billing Hub. From 8 September 2026, credit-card payment is also supported as an alternative, rolling out gradually — some accounts may see it before others.
Can I use 360dialog's (or my partner's) credit line for a client's MBA?
No. A credit line belongs to the Business Manager that owns the WABA, and MBA billing follows that same rule — a 360dialog or tech-provider credit line can't be reused for an end client's agent, even if that credit line is already attached to the WABA for other purposes. Each client needs to add or apply for their own credit line.
My client's Business Portfolio doesn't even show an option to apply for a credit line — what can they do?
There's no workaround today. Two options: wait for credit-card support, or have the client check their eligibility directly on Meta's Payment Methods page in Business Manager — 360dialog can't apply for a credit line on a client's behalf.
How is MBA billed?
Usage is postpaid: charged at the start of the following month. There's no prepayment and no mid-month charge. If you're on a credit line and utilization hits 100%, MBA access is suspended until it's resolved — a usage budget is the practical way to avoid hitting that ceiling unexpectedly.
Does MBA usage show up in my normal WhatsApp billing analytics?
Not yet. MBA usage doesn't currently appear in the standard WhatsApp message billing analytics. In the meantime, track usage via Conversation Turns and Budget. Meta has also announced near-real-time cost-monitoring APIs, targeted for around 15 September 2026 — we'll document them once they ship.
Does the pricing change coming 1 October 2026 affect MBA?
No — MBA pricing itself doesn't change. The 1 October update is about service messages sent by anything other than MBA (a human agent, or your own 3rd-party AI/backend) — those start being billed per message from that date, with one exception: the first 1,000 service messages per phone number per month are free. That free allowance applies to non-MBA service messages, not to MBA messages, and it counts per message, not per conversation.
MBA messages have been billed separately since 1 August 2026, at a flat rate of $2.00 per 1 million tokens (roughly 4–5 US cents per message), covering message delivery and AI processing in a single charge. Nothing about that rate or mechanism changes on 1 October.
How do I onboard — API, or something more hands-off?
Three ways, depending on how much control you want:
Via API — build your own onboarding and management experience. Full control, available today. This guide covers that path.
MBA Integrated Onboarding — a hosted, guided flow (similar to 360dialog's Number Integrated Onboarding) that handles terms acceptance and setup for you. Live now.
360dialog Hub — enable and manage MBA directly from the Hub UI, no integration work at all. Best for a single account that doesn't need programmatic control.
Is MBA available for my number?
It depends on eligibility (account type, region, and a few other factors) — run the eligibility check before assuming. See 1. Check eligibility. Availability was regional during early rollout (Brazil first, EU from mid-July 2026); that restriction appears lifted, but if a number in a newer region reads as not-eligible, re-check after a short wait.
Can I use MBA on Instagram?
Not today. Meta's current MBA documentation covers Messenger and WhatsApp only; Instagram is mentioned as a future direction but isn't available yet. Set up and configure the agent for WhatsApp as usual — if Instagram support ships, it's expected to be the same underlying agent, just connected to an additional channel, not a separate setup.
Testing
Can I test MBA before opening it to real customers?
Yes — that's what the allowlist is for. Set ai_audience to ALLOWLISTED_ONLY, add your own numbers, and message the agent from those numbers on real WhatsApp. See 5. Go live.
Does the agent_test endpoint work during allowlist testing?
Use it for quick automated checks (it starts a fresh conversation each time unless you pass back a conversation_id), but for validating a production number before public launch, testing on real WhatsApp under an allowlist is the more representative method. See 4. Test your Agent.
How do I see why the agent replied the way it did?
Pull that conversation's turns — each one shows the internal steps the agent took (which tool it called, what it returned, how long each step took, and whether anything errored or timed out). See Conversation turns.
Configuration
What can I actually teach the agent?
Three things: knowledge (business info, FAQs, files, websites it can reference), behavior (skills — your workflows and tone, written as natural-language directives), and capabilities (connectors and tools that call your own APIs, plus UI Skills for rich WhatsApp components). See 3. Configure your agent.
Can I upload a spreadsheet or a PDF with a table as a knowledge source?
Not reliably today. CSV files, and tables inside PDFs, error out when the agent tries to read them. Put structured or tabular facts into FAQs or business info instead. See Known limitations.
Can the agent send a file or image back to a customer?
Yes. Sending rich messages (images, buttons, carousels, lists, locations, flows) via UI Skills is supported and is the right tool for that.
What's the difference between a regular Skill and a UI Skill?
A regular Skill tells the agent how to think and respond — tone, priorities, workflows, all in plain text. A UI Skill tells it when to reply with a rich WhatsApp component instead — a button, image, list, carousel, location, or Flow — again as a natural-language instruction rather than a JSON payload you build yourself. See UI Skills.
Can I connect the agent to my own systems (order lookup, booking, inventory)?
Yes, via Connectors and tools — you define an external API once (a Connector), then add individual operations on it (Connector Tools) with a clear description so the agent knows when and how to call them. See Connectors and tools.
Can I connect an MCP server as a connector?
Not yet. MCP-based connectors are on Meta's roadmap but not shipped. Integrate via the REST connector + tool model for now. See Known limitations.
Can I cap how much the agent spends?
Yes — set a usage budget, capping token or AI-turn usage within a rolling time window (one day up to thirty days). Once the cap is hit, the agent stops responding until the window resets. See Usage budget.
Can I make the agent send a specific, exact-match message at a certain point (e.g. for conversion tracking)?
You can instruct it to — write a skill telling the agent to reply with an exact phrase when a specific condition is met (e.g. "once the booking is confirmed, reply with exactly: '...'"). In our own testing this behaves as expected, but treat it the way you would any AI-generated output: it's a strong instruction, not a hard guarantee of byte-for-byte output every time. If you're wiring this into an exact-match tracking system (e.g. a conversion tool that matches on literal text), monitor it after launch rather than assuming it'll never drift.
Handoff & control
What happens when the agent hands off to a human?
It stops answering that chat entirely — make sure a human handover path is actually staffed before you launch, or those customers get silence. On hand-back, the agent also loses the earlier conversation context, so don't assume continuity across a handoff. See Configure your agent — Handoff.
Can I take over a conversation the agent is handling, and give it back?
Yes. You take control implicitly the moment you send a message to that conversation — from then on, the consumer's messages arrive on messages instead of standby, and the agent stays quiet. Hand it back with the Thread Control endpoint's release action. There's no endpoint to read who currently holds a conversation — infer it from messaging_handovers webhook events. See Thread control.
Does the agent always reply in the customer's language?
Not always automatically — reinforce language behavior explicitly in your skills where it matters.
Can I A/B test MBA against my existing WhatsApp flow on the same number?
Yes, but it's not a built-in percentage-split toggle — you build the split into your own routing. The pattern: keep your existing flow as the default, and only release thread control to MBA for the subset of conversations you want in the test (for example, a random split at the point a user enters a specific journey). MBA doesn't take over a conversation until thread control is released to it, so whatever owns your routing logic upstream — your flow builder, your backend — decides which users get the MBA-led experience. See Thread control.
Known limitations & open questions
Is the platform stable enough for production traffic?
MBA is still in early access. Endpoints can intermittently return a 4XX or 500 with no obvious cause while Meta debugs ongoing instability — this isn't necessarily a sign your configuration is wrong; retry after a few hours before treating it as a real error. See Known limitations — Platform stability.
Is billing fully finalized?
Not entirely. The payment-method mechanics above are confirmed, but Meta is still shipping billing-related tooling (e.g. the near-real-time cost APIs targeted for mid-September) and MBA usage still isn't in standard billing analytics. Treat commercial terms as provisional and check Known limitations for the latest.
Can I read who currently owns a conversation via the API?
No — there's no GET for thread control today; only release is supported, and ownership has to be inferred from webhook events (messaging_handovers) and the standby vs messages routing. We've asked Meta for a read endpoint.
Will agent_test ever work properly under ALLOWLISTED_ONLY?
Historically it didn't — it required the agent to be fully published (EVERYONE) to return a real reply. Check 4. Test your Agent for the current behavior, since this is one of the areas Meta has been actively changing.
What's still coming that isn't documented yet?
Near-real-time token/cost monitoring APIs — targeted around 15 September 2026.
MCP-based connectors — on Meta's roadmap, no date yet.
A thread-control read endpoint — requested, no date yet.
We'll fold each of these into the relevant page as soon as they ship — this FAQ and Known limitations are the two pages to check back on.
Last updated
Was this helpful?