What HighLevel connected
The connector went live on 29 September 2026, according to Meta's newsroom announcement. Muse gets access to four things inside HighLevel: customer records, calendars, communication threads and backend automations. In practice, the assistant can look up who a contact is, check whether a slot is free, confirm an appointment and write the result back into the CRM without a human copying anything between browser tabs.
HighLevel is not the only tool on the list. Meta's small-business connector expansion covers more than 15 platforms, among them Shopify, Intuit QuickBooks, Stripe, Slack, Canva and Zoom. The pattern repeats in each case: an assistant that once answered questions now executes tasks in software it does not own.
The second integration in three months
Meta and HighLevel already shipped together once this summer, when Embedded Appointment Booking for Facebook Lead Ads rolled out in June 2026. That pairing moved a lead from a Facebook ad into a calendar. The September connector goes further, because it gives Muse write access to the operational layer rather than just the booking step.
Muse itself has changed shape over the same period. It launched as a standalone personal assistant on mobile and web on 8 September 2026. With "Muse for Small Business" arriving on 29 September, it is now positioned as something that sits across several business platforms at once.
What one agent run costs
Neither company published a price for the connector. Token costs are still worth modelling, because any workflow that reads a thread, checks a calendar and writes back to a CRM is a metered API workload. Take a modest assumption: 8,000 input tokens to read the conversation and the relevant records, 500 output tokens for the reply and the CRM update. Every figure below comes from the vendor's own published pricing page, linked on the model name, checked on 1 October 2026 and converted at €0.86 per dollar — the EUR/USD rate we applied on 1 October 2026, so re-check it before you model a contract. The comparison covers only models with published list pricing; the open-weight option sits at the bottom because its cost is hardware, not a licence.
| Model | Input / 1M tokens | Output / 1M tokens | Input (EUR) | Output (EUR) | One run, 8k in / 500 out |
|---|---|---|---|---|---|
| GPT-6.1 Sol | $4.00 | $20.00 | €3.44 | €17.20 | €0.036 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | €1.72 | €8.60 | €0.018 |
| Grok 4.7 | $2.00 | $6.00 | €1.72 | €5.16 | €0.016 |
| DeepSeek-V4.1-Flash (peak) | $0.30 | $1.20 | €0.26 | €1.03 | €0.0026 |
| DeepSeek-V4.1-Flash (off-peak) | $0.15 | $0.60 | €0.13 | €0.52 | €0.0013 |
| GLM-5.3-Flash | $0.15 | $0.50 | €0.13 | €0.43 | €0.0012 |
| Muse Spark 1.3 (self-hosted) — open weights, no licence fee; the cost is your own GPU time and electricity | |||||
Two things stand out. A mid-tier model such as Claude Sonnet 5.5 or Grok 4.7 lands under two cents per run, and the cheap open-weight options sit an order of magnitude below that. Ten thousand automated conversations a month at Sonnet pricing works out to about €180 in tokens. The same volume on GLM-5.3-Flash is closer to €12. Whether that difference matters depends on the seat cost, which neither vendor has disclosed for this integration.
The assumption is also generous. Production agents pull more context than a chat message does: past appointments, notes, tags, the last three emails. If the agent reads 30,000 tokens per run instead of 8,000, multiply those figures by roughly four. Modelling that before rollout is cheaper than discovering it on an invoice.
What a European agency has to check
Client records, calendars and message threads are personal data. Under GDPR, the agency stays the controller. Meta and HighLevel act as processors, and each flow needs a data processing agreement that genuinely covers it. Ask where the data is processed and whether the connector can be scoped to EU regions. Neither company published a data residency commitment alongside this launch, so the answer has to come from the contracts you sign rather than the announcement.
The regulatory situation moved during 2026 as well. The EU AI Office and national authorities now enforce the AI Act actively, and Article 50 transparency duties apply, including machine-readable marking of synthetic content. Annex III high-risk obligations, by contrast, were deferred by the Digital Omnibus on AI. An agency using Muse to draft confirmations and update records does not turn itself into a high-risk provider overnight. It does have to disclose AI-generated content where Article 50 applies, and it should be able to say which model wrote what.
Self-hosting is the alternative. Muse Spark 1.3 has been available as open weights since 2 September 2026, and GLM-5.3-Flash ships under an MIT licence. Running either on your own hardware keeps the records inside your network, at the cost of GPU time. That is the class of model we test on the AI Arena rig, and it is a reasonable option for agencies handling sensitive client data that cannot wait for a vendor's residency answer.
For document-heavy pipelines, Mistral OCR 4.1 is priced at $4 per 1,000 pages, about €3.44, and is a European option worth comparing against the US incumbents. It covers parsing, not workflow execution.
Where the market already moved
The connector fits a wider change in how companies buy AI. Text-only chatbots and prompt pilots are no longer the product. What is being bought in late 2026 is multi-agent systems wired into existing software, with governance attached and a return that can be measured. HighLevel is an agency platform, and agencies resell. Once an assistant can book, confirm and update records on its own, the reseller has something billable per client instead of a demo to show.
Is the Muse connector available in the EU?
Meta's small-business connector expansion covers more than 15 tools including HighLevel, but regional availability follows each vendor's own rollout. Check Meta's and HighLevel's regional documentation, and confirm with your account manager before you build a client-facing workflow around it.
Do I need Muse to build this?
No. The same loop, reading a thread, checking a calendar and writing back to a CRM, can be built with API calls to any of the models in the table above. Muse's advantage is setup time, not capability.
Does the AI Act treat this as high-risk?
Booking and CRM updates are not Annex III use cases. Article 50 transparency duties do apply to synthetic content, and the Annex III high-risk deadlines have been deferred by the Digital Omnibus on AI — that deferral is the timetable to plan against if a use case ever moves into that category.
For agencies, the decision point arrives earlier than the roadmap suggests. Before Muse goes anywhere near a client-facing workflow, three things need an answer on paper: which records the agent may read and write, where those records are processed, and who signs off on the output before a customer sees it. The models above differ by an order of magnitude in token cost, which matters less than the seat price neither vendor has published for this connector — and far less than a residency answer you cannot put in a contract. Agencies that settle those three questions first can resell the workflow with a number attached; those that wire it up and ask afterwards will be explaining a data flow and an invoice at the same time.
