Bring Your Own AI Expert to Every Meeting
Every team knows the moment. A meeting is going well, someone asks a question nobody can answer on the spot, and it ends with "let me check and get back to you." The decision waits a week for a fact that took two minutes to find.
At Alchemy Loop we built ourselves an answer: Casey, Your Personal AI Meeting Expert, for Microsoft Teams & Mobile. It began as an internal product, an AI expert who joins our Teams calls and travels with us on a phone. Casey knows our subjects and our documents and, when we connect them, the live figures in our systems. Here is how he works, and what it takes to bring an expert like him into your own meetings.
He listens, and speaks when it helps
Once everyone has agreed, Casey listens along with the team. Ask "Casey, what was last quarter's survival rate?" and he answers out loud in a sentence or two, saying where the answer came from. "Tell us more" gets the longer version.
Because he has been following the conversation, he understands "that site" or "the number you just gave", and he doesn't repeat what people have just heard. If the meeting wants it, he can also join in on his own: when a question is put to the room and nobody picks it up, he offers the answer. He is deliberately cautious about it. A rhetorical question, small talk, or a question aimed at a particular person is left alone, because an unwanted interruption costs far more than a missed chance to help. Or he can simply wait quietly until he is called by name.
Research while you wait
The moment people like most: "Casey, research this and let me know when you're ready." He goes off in the background, the meeting carries on, and a few minutes later he says the briefing is ready. Research is the one task where Casey goes beyond what your team has given him: he uses the AI's own knowledge and searches the public internet, so the meeting can draw on outside expertise on the fly. The briefing is kept, so the next call starts from what was learned in this one.
Questions about your own data
Casey has no access to your databases on his own. Paired with our product LLModel Query, he can answer questions about them: "how many orders shipped late last month?" LLModel Query turns the question into a read-only query, runs it inside your network, and writes the answer there. On its own, LLModel Query never sends your rows to an AI. Casey works differently: to answer out loud, in context, he passes that answer and the handful of rows behind it to your AI account, and he remembers his answers for the rest of the call. The AI never sees or connects to your database, and LLModel Query keeps nothing it reads. If even that small result must stay private, run the AI on your own server (below), and nothing from your database reaches an AI company.
Where everything goes
Your Teams meetings are probably already transcribed by Teams. Casey adds one service to hear and speak, and the room agrees before he listens. You should still know exactly what goes where, so here is the whole list. Two outside services are involved: ElevenLabs, which turns speech into text and gives Casey his voice, and the AI model, which runs on your own business account with Anthropic or OpenAI, under your own agreement with them. Business accounts with both providers are not used to train their models by default.
| What | Where it goes |
|---|---|
| The meeting's audio, while Casey is listening | ElevenLabs, live, to turn speech into text. It is not sent to the AI as audio. |
| Casey's replies | Written by the AI; ElevenLabs turns the words into his voice. |
| A question put to Casey | Your AI account, with the matching parts of your documents and the names of the people on the call. With the transcript on, also the conversation so far and his earlier answers. |
| Joining in unasked (if switched on) | Your AI account sees the last few lines of conversation to decide whether to speak. Needs the transcript on. |
| Research | Your AI account and public web searches, with the topic and related material from Casey's own knowledge pack. Nothing from the meeting is sent. |
| Questions about your database | Only through LLModel Query: the answer to that one question and the few rows behind it go to your AI account, so Casey can word his reply and refer back to it later in the call. ElevenLabs receives his spoken reply. |
| Meeting notes | With the transcript on, the transcript and Casey's answers, figures included, are kept on your own systems, and the AI reads them to write a summary at the end. With it off, none of these. |
Everything Casey keeps stays on your own systems, but the meeting is processed by the services you choose, and ElevenLabs keeps its own copy unless you arrange otherwise. On an ElevenLabs Enterprise plan your data isn't used for training, and you can ask for zero retention. On any other plan, switch off training in your account settings. Review your AI account's terms the same way.
The transcript is your choice. Turn it off and the AI sees only the question put to Casey, never the conversation around it. Casey then answers when he is called by name, but he can't follow references to earlier discussion or join in on his own, and no meeting notes are written. ElevenLabs still hears the meeting while Casey is listening, because that is how he hears his name. And anything people say aloud, a figure repeated across the table included, is part of the conversation.
Or keep the AI entirely private
Instead of Anthropic or OpenAI, the AI can be an open model running on your own private GPU server, rented from a service such as RunPod and switched on only during business hours, which keeps it affordable. Questions, transcripts and research then go to a machine only you use, not to an AI company. Open models such as gpt-oss-120b, Qwen3.5 and DeepSeek-V4-Flash are capable of the whole job.
The same expert in your pocket
Not every meeting is on a screen. For a meeting in a room, a site visit or a lab, Casey runs on a phone:
- Tap to ask. Hold the button and ask, or tap once to start and again to stop. Casey hears only the question, not the room.
- Always listen. Read the consent line to the room, and Casey follows the whole meeting, answering when called by name.
- Research. Ask for a briefing out loud, or type a topic, and Casey says when it is ready.
- Who is here. A short list of who is in the meeting; Casey identifies people in the room only from it.
- End meeting. One press, and Casey stops listening.
The phone reaches Casey over Tailscale, a private, encrypted network that admits only the devices your company has approved, so Casey is never open to the public internet.
Guardrails
- Consent first. Casey does not listen until the room has agreed.
- He says where each answer came from: your documents, your live data, or the web.
- Live figures and web searches are kept apart: Casey doesn't search the web while answering from your data, and he labels which is which.
- Your AI account, your agreement. The AI runs under your business account, not ours.
The best meetings end with decisions, not homework. An expert in the room who can look things up while you talk is one of the simplest ways to get there.
Alchemy Loop builds AI into everyday business software, from help on the screen to an expert on the call. Learn more at alchemyloop.ai.