LuminorixWatch it do the work. Then decide.
The 90 second recording of a real AI agent doing a real job goes here once it is cut. It is made on the real app, not staged in a web page. Below it, how the work runs.
Recorded on a test build, 2026-10. In testing with pilot customers, not in the App Store yet.
Built for work that still runs on people, email and portals.
Payments and remits, filed and summarised
Logs into the payer portals and files each remit. A person takes the steps a portal will not allow a bot to do.
Month-end documents, chased for you
Knows which clients still owe documents, checks what came back against your checklist, and drafts each reminder for you to send.
Example. No customer yet; we would build the first checklist with you.
The reports and follow-ups that eat the week
The status report your prime contractor asks for, the vendor report your suppliers expect, bills matched and held when something is off.
Distributors: running at a reference customer. Construction: example only, no customer yet.
What we can stand behind today.
Counts from our own systems. No customer is named. For the team to decide House rule 5 allows only one proof line; these four are shown so you can react to them.
Median time from the end of your sentence to the start of the spoken answer, measured on the voice service.
Backup restores we tested and verified.
Remit files the billing agent renamed and filed to a client's shared folder in its first run, 2026-10-07.
Billing company pilot so far. Every customer today is a pilot.
Agents switched on in May. Profitable every month since.
Reference customer. June to August 2026.
A billing company's week with an AI employee.
This is how the billing agent works for one medical billing company. Names, payers and amounts are left out.
- Monday, 7:00 am Pacific. The agent files the previous week's remits and leaves one line in the team's Teams thread: what arrived and what waits on a person. A payments summary is available on request. The weekly run starts Monday 2026-10-12.
- Any day a remit arrives. The agent renames the file to the date, the amount and the portal, and files it to the company's shared folder.
- When a portal asks for a robot check. The agent does not try to get around it. It asks a person, who signs in on a shared screen, and the agent carries on from there.
- Replies to the client. The system runs a quality check and a wording check on a reply before it goes out.
- De-identified billing operations only today. No patient information is processed until a BAA is signed with us and with the model provider.
- One billing company pilot so far, since August.
- Some portals need a person every time.
Why not just use Claude yourself?
Claude is very good. Using it on its own, or through a desktop tool such as Cowork, is a fine way to get one person more done. Running a business function on it is a different job.
| Where it gets hard | Claude or Cowork on your own | Luminorix |
|---|---|---|
| Across the organisation | A good fit for one person's own work. Sharing it across a team is left to you. | Set up and run for a team, so a new person uses what is already working. |
| Your data | It sits in the account that set it up. Where it goes is yours to manage. | A workspace for each customer, a record of actions, and a plain list of what we do not have yet. |
| What it costs | Model usage adds up with every task, and someone has to watch it. | Model usage is our job to manage. You see the work done, not a usage meter. |
| When nobody is at the desk | Usually runs when a person starts it. | It runs on a schedule and on events, and writes to you in Teams, Telegram or email. |
| Who sets it up and runs it | You do. | We do. A named person on our side answers when something looks wrong. |
Built on Claude, operated by us, embedded in your business. It works when nobody is at the desk, lives in your channels, has a name, a role and a memory, has control built in, and is done for you.
A person on your team approves or sends.
AI agents draft the next step. For the steps that matter, you get a card with the reason and you approve or hold. Every action is on the record, and where an action can be reversed you can undo it. A spoken yes never approves; you tap.
- On its own: read, draft, prepare.
- Comes to you as a card: sending, posting, paying, changing a price.
- Our hard stop list: moving money, writing to someone new, deleting.
Send 2 reminders. Hold 1 for you.
You read each draft first. Where an action can be reversed, you can undo it.
Example card on example data. Nothing here is connected to a real system.
See it on your own business.
A short call with our team. We start from your kind of work and show you the first result. Pricing is shared on the call.
- A short call with a founder.
- We ask what takes your week before we show anything.
- A person on your team approves or sends. Every action is on the record.