Record the meeting. The machine does the rest.
A room full of people, one laptop, one tap on record. Freehold's engine picks the recording up out of the company's own drive, separates and identifies every voice by its thumbprint, confirms the ones it doesn't know yet over email, and publishes a fully attributed transcript into the company's system of record. Then agentic librarians get to work — updating the business's documentation based on what was said, what was decided, and by whom. Nobody took notes. Nobody uploaded anything. Nobody wore a mic.
Meetings are where a business actually runs — and where its decisions evaporate.
The real operating system of most companies is people in a room talking. That's where priorities get set, work gets assigned, and disagreements get settled. Then everyone stands up, and the record of it is somebody's memory. Even good meeting notes tell you what was decided; they almost never prove who committed to it. Six weeks later, "we agreed to that" and "no we didn't" are both unfalsifiable.
With an attributed record, the answer is a name, a timestamp, and the person's own words — not a reconstruction.
Decisions land in the system of record the day they're spoken, linked to the meeting that produced them.
Curator agents reconcile what the room said against what the docs say — and flag the gaps instead of letting them rot.
A readable, speaker-attributed record — not a 57-minute audio file nobody will ever open again.
From a tap on "record" to self-updating documentation
Six stages. One of them is a human answering a short email. Everything else runs itself.
Anyone taps record. No special hardware, no bot joining a call, nothing to install. The meeting happens the way meetings happen.
The engine watches the company's shared drive. A new recording appears; work begins. Nobody uploads, exports, or files anything.
The hour becomes hundreds of wall-clock-stamped segments, each assigned to a distinct voice in the room — before anyone knows whose voice it is.
Each voice is compared against the enrolled voiceprint store. People the system has heard before are named automatically, with a confidence score.
For voices it doesn't know, an agent emails the meeting organizer — one plain-language question at a time. Each answer enrolls a new thumbprint. Next meeting, no questions.
The attributed transcript lands in the system of record. Librarian and curator agents then update the company's documentation from it — decisions, owners, corrections.
What the engine hears: who held the floor, minute by minute
illustrative rendering — not the client's meeting
Every burst is speech assigned to one voice. The grey lane is the part the system refuses to guess about — voices too far from the microphone, or two people blending on one mic, stay unattributed rather than misattributed.
The whole training process is answering an email.
There's no enrollment ceremony, no "please read this paragraph into the mic." While the thumbprints are still being established, the agent interviews the meeting organizer — briefly, in plain language, quoting the moments it needs help with.
I've identified two of the voices in today's session from previous meetings. One voice I don't recognize spoke for about four minutes, mostly in the second half — for example at the 25-minute mark, walking through how the work should be divided. Who is that?
that's our newest project manager
Got it — enrolled. I'll recognize them automatically from now on. One more: two people seem to share the microphone around minute 40. I won't attribute that stretch to either of them unless you'd like me to.
Why this design
The organizer already knows the answer. They were in the room. A ten-second reply does what an enrollment workflow, a hardware budget, and a training session would otherwise do — and the system gets permanently smarter with each answer. Confirmation is also an accountability gate: a human vouches for every name before it enters the record. The machine proposes; a person who was there confirms.
A record you can act on, because it refuses to guess
Attribution is only useful if it's honest. This engine is built to say "I don't know" — loudly, in the record itself.
Every line carries a trust label
Voice-matched and human-confirmed segments are marked high trust. Stretches where two speakers blend on one microphone are marked do not attribute — permanently. The system will not enroll a blended voiceprint, because a corrupted thumbprint would quietly poison every future meeting. Better an honest gap than a confident error in the company's system of record.
Physics is stated, not hidden
One laptop microphone hears the people near it best. The engine measures that limit and reports it — who was cleanly separated, who wasn't, and why — instead of inventing speakers to fill the silence. When it says four voices were identified, that claim has evidence behind it that an engineer can audit. A system that admits what it can't hear is one you can trust about what it can.
That honesty is what makes the downstream automation safe. In the working session behind this study, assignments that had lived in the documentation as "inferred — treat as provisional" became confirmed records the same afternoon, because the people responsible said so out loud, in voices the system could prove were theirs. And an analysis the AI had drafted days earlier was promoted from "machine-generated" to "corroborated" when the room independently arrived at the same structure on the record. Machine work rises on human evidence — never on its own eloquence.
Run the business by talking.
Attributed meetings are the foundation, not the product. Once the machine reliably knows who decided what, decisions stop being things people have to remember to act on — they become inputs to agentic workflows.
Attributed transcripts in the system of record; librarians update the documentation from what was said and decided.
Agents pick commitments off the record and file them: tasks with owners and dates, follow-up drafts ready for approval, doc corrections with provenance.
Whole business processes triggered by conversation — decide out loud, and the machine in the background actions it, with humans approving anything that matters.
The end state is simple to say: the leadership team sits around a table and works the way people naturally work — talking, arguing, deciding. The estate listens, keeps the honest record, updates what the company knows, and sets the machinery in motion. The meeting is the interface.
Your meetings could be doing this.
Everything on this page runs today, for a real company, on commodity hardware — a laptop on a table and an engine behind it. The client's people, numbers, and words stay theirs; the capability is what's for sale.