Muse for Customer Support: Working the Queue
Staff the front door with Meta Business Agent and work the queue with Muse: morning triage, draft replies a human sends, a knowledge base that stops rotting, and escalation rules that hold.
By Professor Muse7 min read
Support is the one job where volume is the whole problem. The queue fills every morning whether you feel like it or not. Half the tickets ask the same five questions, and the other half need a human who actually knows the product. I run my support as two separate jobs: one agent works the front door around the clock, and Muse works the queue behind it. The front door is Meta’s Business Agent. The queue work is Muse with connectors. Keep those two jobs straight and support gets quieter. Mix them up and you get an agent quoting refund policies it invented.
The Two Products, Kept Straight
Meta makes two different support products and confusing them is the first mistake. One represents your business to customers. The other works for you, behind the counter.
| Meta Business Agent | Muse | |
|---|---|---|
| Works for | The business | You, the operator |
| Launched | June 3, 2026, global | September 8, 2026, US and Canada |
| Where it lives | WhatsApp, Messenger, Instagram | The Muse app, web, WhatsApp |
| The job | Answers customers, recommends products, books appointments, escalates to a human | Triages the queue, drafts replies, syncs the knowledge base |
| Autonomy | You set the escalation threshold | Sentinel approvals: it never sends without your say |
Business Agent went live globally at Meta’s Conversations conference after roughly two years of limited testing, with Meta saying more than a million businesses were already running one. It learns from your Facebook page, your past chats, and your website, and you control how much rope it gets: from drafting replies you approve all the way to handling conversations on its own until it hits the escalation threshold you set. Since August 2026 every AI-generated reply is billed at $2 per million tokens, roughly a few cents per reply, and since mid-September the Meta One plans bundle expanded access from $14.99 a month per WhatsApp number.
Muse, meanwhile, arrived in September with the Muse for Small Business connectors: Shopify, Stripe, QuickBooks, Slack, Notion, Klaviyo, HighLevel, and the Facebook and Instagram business accounts. Support queues live behind connectors too: Zendesk, ServiceNow, and Jira are reachable through MCP-style connectors that plug your queue into Muse. The standing rule from the small-business launch applies here in its strictest form: Muse will not publish, send messages, or spend money without your approval.
The Setup: The Front Counter
The Front Counter is the name I give the whole arrangement. The whole thing runs every weekday. Business Agent staffs the front door in WhatsApp and Messenger, answering the repeat questions at any hour. Everything that lands in the queue, the escalations, and the places Business Agent is not (your email, your helpdesk) is Muse’s counter to work.
Every morning at 8am a scheduled routine runs the queue drill. It reads every new ticket in Zendesk: labels each one (billing, shipping, bug, refund, praise), ranks the queue by urgency, and flags the three tickets that need a human first. Then it drafts. For the routine questions it writes replies in your brand voice, cites the knowledge-base article it used, and queues the drafts in a folder for you to read. You tweak if needed and send. Draft, then human, then send: that order never changes.
Once a week a second routine audits the knowledge base. It pulls the questions the KB could not answer cleanly that week, the ones where Muse had to reach for a stale article or hedge, and proposes new entries or rewrites. This is the chore nobody does, which is exactly why knowledge bases rot: they get written once at launch and freeze while the product moves on. A weekly thirty-minute audit keeps the source of truth honest, and everything Business Agent answers from it gets better too.
Then the escalation rules, which decide whether the whole setup holds. Any ticket mentioning a refund, a chargeback, an account closure, or a safety issue gets labeled “escalate”, with no AI reply and no draft at all. Business Agent’s escalation threshold gets set the same way: those categories route to a human immediately. The specialist pattern applies here as it does everywhere else: the drafting setup drafts, the triage setup triages, and no setup sends.
The short version
- Meta Business Agent staffs the front door (WhatsApp, Messenger, Instagram). Muse works the queue behind it. Two products, two jobs.
- Run a morning queue drill: label, rank, draft. A human sends every reply. That order never changes.
- Refunds, chargebacks, account closures, and safety issues escalate to a human with no AI reply at all.
- Audit the knowledge base weekly against the questions it failed to answer. The KB is the product everything else reads.
The Copy-Paste Prompts
Prompt 1: The morning queue drill
Set this as a scheduled routine against your helpdesk, weekdays at 8am:
Every weekday at 8am, work my support queue in [Zendesk / ServiceNow / helpdesk]. Read every ticket from the last 24 hours. Label each one: billing, shipping, bug report, refund, account, or praise. Rank them by urgency and flag the three that need a human first, with one line on why. Then draft replies for the routine tickets: write in my brand voice (short, plain, no jargon), cite the exact knowledge-base article each draft answers from, and queue them as drafts. Do not send anything. Do not draft replies for tickets mentioning refunds, chargebacks, account closure, or safety issues: label those “escalate” only.
Prompt 2: The draft with a paper trail
Draft a reply to ticket [id]. Read the ticket and the relevant knowledge-base article. Write the reply in my brand voice: short, plain, no jargon. Cite the exact KB article and section you are answering from. If the ticket quotes or asks about a policy (refunds, warranties, data deletion), quote the policy text verbatim from the KB and say which article it came from. If the KB does not contain the policy, say so explicitly instead of writing one. Draft only. I send it.
Prompt 3: The knowledge-base audit
Run the weekly knowledge-base audit. Pull this week’s tickets where the reply needed a workaround, a guess, or an article older than six months. For each gap, propose a new KB entry or a rewrite of the stale one, with the ticket numbers that exposed it. Rank the proposals by how many tickets each gap touched. Propose only. I approve and publish.
What Can Go Wrong: The Confident Refund
The failure mode you will meet first is the draft that sounds perfect and is wrong. Muse writes a warm, crisp reply to a customer asking about your return window. It cites your refund policy. The policy says 30 days; the draft says 60. Nothing in the draft flags the mismatch, because the agent is fluent, not truthful. If you are approving twenty drafts on a busy morning, approval fatigue is doing its work, and the wrong number goes out under your name.
The fix is in Prompt 2 and it is non-negotiable: policy text gets quoted verbatim from the knowledge base with the article named, and if the KB does not contain the policy, the draft says so instead of writing one. That single rule turns a fluent guess into a checkable claim. The second habit is a weekly spot-check: pick five sent replies, verify the cited KB lines, and watch whether the KB itself is drifting. Most wrong replies trace back to a stale article, not a bad agent.
Support run this way replaces the part of the job that was never judgment: the labeling, the ranking, the first draft of the obvious answer. Your judgment stays yours. The front door stays staffed all night by Business Agent. The queue gets worked every morning by Muse. And the replies that carry money, accounts, or safety land in front of a human, because those were always human decisions. The queue gets quieter, and the loud tickets get the attention they were owed all along.
“Draft, then human, then send. That order never changes.”
