Can an AI notetaker be trusted with a listing appointment?
Written for agents, brokers, lenders, contractors, home services.
The short answer
An AI notetaker listens to a call and produces written notes from it. Granola is one of the better ones and the category behaves similarly across products.
Trust it to save you an hour a day. Do not trust it as the only copy of anything somebody promised.
The distinction that matters: a transcript is a record of what was said. Auto-generated summary notes are a machine's account of what it thinks mattered. Two documents, two reliability levels, and they get filed as though they were one thing.
Documented failures from one call on 2026-08-07: a third participant appeared in the meeting record who never speaks anywhere in the transcript, and one side of the conversation dropped out of the capture partway through.
The tool has no verification step and no way of marking what it is unsure about. That is defensible design. It means the checking has to be yours.
A lender on six calls a day cannot take notes on six calls a day. Neither can an agent running back to back listing appointments, or a contractor who walks four jobs before lunch. So the notetaker gets switched on, and for the first month it feels like the best thing anybody bought all year. Notes appear. Action items appear. Nobody reconstructs a Tuesday from memory on a Thursday.
We use these tools. This is not an argument against them. It is an account of four ways we have watched them fail on real calls, each of which produces output that looks exactly as confident as the output that was right.
What is an AI notetaker?
An AI notetaker is software that listens to a call, either by joining it as a participant or by capturing the audio on your machine, and produces written notes and usually a transcript. Granola is one of the widely used ones, and several products in the category behave much the same way.
Most of them produce two separate things from one call. A transcript, which is a line by line record of the words the system heard, usually labeled by speaker. And summary notes, which are a shorter document written by a language model describing what it thinks the call was about and what it thinks was agreed.
Almost everybody reads the second one and files it. Almost nobody reads the first one. That habit is where most of the trouble below comes from.
What these tools are genuinely good at
- Volume. Six borrower calls produce six sets of notes without you doing anything. The alternative is one set of notes and five gaps.
- The obvious action items. When somebody says "I will send you the survey Thursday", it gets caught. Plain commitments in plain words come through.
- Names of things that came up. Which inspector, which subdivision, which loan program. Useful as an index even when the surrounding sentence is unreliable.
- The boring middle. Nobody remembers minute 14 of a 40 minute call. The transcript does.
- Being there while you were not writing. On a job walk or a listing appointment your attention belongs on the person in front of you.
That is a real hour a day for a busy lender or a team leader, and we are not talking anybody out of it. The rest of this piece is about the boundary of what the output can carry.
Failure one: a person who was never there
On a discovery call recorded 2026-08-07, the notetaker put a third participant into the meeting record. No third voice appears anywhere in the transcript. Not one line. That person was somewhere in the calendar orbit of the meeting and the tool wrote them into the attendance record as fact.
Consider what that does downstream. A call record shows three people on a conversation about a listing price, a commission split, or what a borrower disclosed about their income, when two were there. Six months later that record is what somebody reads to work out who knew what, and nothing on its face says it is wrong.
The check takes thirty seconds. Open the transcript and confirm that every name in the participant list speaks somewhere in it. A name that never speaks does not belong there.
Failure two: a stretch of the call that simply is not there
On that same 2026-08-07 call, one side of the conversation dropped out of the capture partway through. Not garbled. Absent. The other participant's own words later in the recording confirm that something significant happened during the gap, which is how the gap got noticed at all.
We went looking in three places: the notetaker, the video platform the call ran on, and the notes tool the record was filed into. The search closed on 2026-08-10 with a plain finding. No capture existed anywhere. The words are gone.
This is the failure that should change how you use the tool, because a transcript with a hole in it does not look like one. It reads continuous. No marker, no gap indicator, no line saying capture was interrupted. If the missing stretch is where the seller agreed to a price reduction, you hold a document that reads complete and omits the only part that mattered.
The same recording also collapsed a long stretch into a single speaker block, so the ordering inside it is unreliable. Who said what, and in what order, is exactly what you go back to a transcript to settle.
Failure three: any word the model has not heard before
Machine notes mangle unusual vocabulary, and they fail first on the words carrying the most meaning. Proper nouns, industry terms and religious language are where it breaks. A separate transcript rendered the product name "Wispr Flow" as "Whisper Flow" throughout, which is a reasonable guess by a system that has heard one of those words a great deal and the other one never.
In this trade that means surnames, street names, subdivisions, title companies, loan programs and trade terms. A borrower's last name is a rare word to a speech model and a load bearing word to you.
The rule we hold: correct the name in anything you write from the transcript, and leave the transcript alone. A transcript is a record of what was heard. Editing it to say what you now believe was meant removes the one property that made it worth keeping.
Failure four, and the expensive one: summary treated as transcript
What is the difference between a transcript and AI meeting notes?
A transcript is a record of what was said. Auto-generated summary notes are a machine's account of what it thinks mattered. They have different reliability, they fail in different ways, and the common mistake is filing them as though they were the same document.
A transcript can be wrong about a word. It is rarely wrong about whether a sentence was uttered, except in the missing-audio case above. Summary notes can be wrong about whether something was said at all, because writing a summary means deciding what a conversation meant, and deciding what it meant involves filling in.
Here is what that produced in one case. Four claims about a client's software stack were carried in a working record, and all four rested only on auto-generated notes rather than on the transcript. Two of the four were later put back in front of that client as things her team keeps. That is a machine's inference being spoken to the person it was an inference about.
Nothing about the note said "inferred". Summary notes carry no confidence marking. A sentence the model heard verbatim and one it constructed from context are typed in the same font.
There is a timing version of the same problem and it is subtler. Notes from 2026-08-07 described an install as pending. A first-hand account put the same install as done. Both statements can be true when they are four days apart, and the note carries no date for the claim inside it, only a date for the meeting.
The caution we wrote down afterwards: do not treat the wording in the notes as a date. "Pending" in a summary means pending as of whatever moment the speaker was describing, which may not be the day of the call. In a transaction with a close date, that decides whether you are chasing something.
How to use one of these without getting caught
- 01Keep the transcript, always. It is the only version that is a record. Export it if your tool allows and put it with the deal. Summaries can be regenerated. A transcript cannot.
- 02Treat the summary as a draft. Read it the way you would read a new assistant's notes on their second day.
- 03Check the participant list against the transcript. Every named person should speak. Thirty seconds, and it catches the invented attendee.
- 04Re-read anything you are about to say to the client's face. Before you repeat a fact about their business, their loan or their property, find the line where they said it. If you cannot find it, ask them instead of asserting it.
- 05Never let a summary be the only copy of a commitment. Anything promised gets written into the place your business keeps promises, the same day, in words.
- 06Watch for the single-speaker block. A long stretch attributed to one person on a two-way call usually means the ordering inside it cannot be relied on.
None of that takes more than five minutes per call, which leaves you far ahead of typing notes yourself.
A worked example: where the notes should end up
PAD is People, Assets, Deals: three folders of plain text files on your own computer, plus one operating file telling an AI assistant how to work inside them. Here is a listing appointment call moving through it, which is the same shape whether you use folders, a CRM or a legal pad.
- The transcript goes into the deal folder, under the call it came from. Unedited, wrong names and all, with the date.
- The summary goes beside it, labeled as a summary. Same folder, different file. The label does real work, because a reader six months from now needs to know which document they are holding.
- Any promise made on the call is written onto the person it was made to. The person, because a promise is owed to a human, and the deal will close and stop being opened.
- Anything the summary asserts that the transcript does not support is checked or dropped. If it is worth keeping, it gets marked unverified with the date it was noticed.
Separating those files makes the reliability question answerable later. One glance tells you whether you are reading what somebody said or what a model decided they meant.
The four boxes
The four box sort applied to AI notetakers, using Granola as the example. What is there and should be, what is missing and should be there, what is there and should go, and what is missing for good reason.
There, and should be
Automatic capture across a full day of calls, a searchable transcript, and a summary written without you.
- Capture that happens without a decision is the whole value. Any system requiring you to remember to start it will be off on the call that mattered.
- The transcript is the durable artifact. Keep it even when the summary is all you read.
- Speaker labeling is useful when it works and worth checking when the stakes are real.
Missing, and should be there
A place for what came out of the call to live, attached to a person or a deal, once the call is over.
- State the gap as a job: every commitment made on a call needs to land where a human will look next week.
- A second gap: the habit of reading the transcript before repeating anything to a client. That is a process rather than a purchase.
- A third: a name list, so whatever receives the notes has correct spellings to check the machine's guesses against.
There, and should not be
The practice of filing the summary and ignoring or deleting the transcript.
- It looks like tidying. It throws away the only document that can settle a dispute about what was said.
- Also remove the habit of quoting a summary back to the client. Two of four claims in one case were inferences, and both were repeated to the person they were about.
- Treating a word like "pending" in a summary as though it carries the meeting date belongs here too.
Missing, and correctly missing
There is no verification step and no confidence marking on any line of the output.
- We think the absence is defensible. A notetaker flagging its own uncertainty on every line would be unreadable, and a page of hedged sentences gets skimmed, which returns you to having no notes.
- A confidence score would also become a number people came to trust, and a wrongly confident number is worse than no number.
- The cost is exact: the reader cannot tell a heard sentence from an inferred one. So the discipline has to live outside the tool, in the five minutes after the call.
Should you use an AI notetaker on a listing appointment?
Yes, with two conditions. Get consent from everyone in the room, since recording rules vary by state and by contract. And treat the summary as a draft that has to be checked against the transcript before any part of it is repeated to the client or written into the file.
The appointment is exactly the situation these tools are built for. Your attention belongs on the seller, and notes taken by hand during a listing presentation are either bad notes or a bad presentation.
What changes with the stakes is the review. A quick internal call gets a glance. An appointment where somebody agreed to a price, a term or a concession gets the transcript read.
What this does not do
Our summary for this trade: keep using it, keep the transcript, read before you repeat, and never let a machine's account of a conversation be the only place a commitment lives.
Common questions
- What is an AI notetaker?
- Software that listens to a call and produces written notes, usually alongside a transcript. Granola is one of the widely used ones. Most products in the category produce two documents from one call: a line by line transcript of what was heard, and a shorter summary written by a language model.
- What is the difference between a transcript and auto-generated meeting notes?
- A transcript records what was said. Summary notes are a machine's account of what it thinks mattered, which means the summary can assert things nobody said. They should be filed as two separate documents so a later reader knows which one they are holding.
- Can an AI notetaker add someone who was not on the call?
- Yes. On a discovery call recorded 2026-08-07 a third participant appeared in the meeting record while no third voice speaks anywhere in the transcript. The check takes thirty seconds: confirm that every name in the participant list speaks somewhere in the transcript.
- Can an AI notetaker miss part of a conversation?
- Yes, and it will not tell you. On the same 2026-08-07 call one side dropped out of the capture partway through, and the transcript reads continuous with no gap marker. A search across the notetaker, the video platform and the notes tool closed on 2026-08-10 having found no copy of the missing audio anywhere.
- Why do AI notes get names and industry terms wrong?
- Speech models guess rare words toward common ones. A separate transcript rendered the product name "Wispr Flow" as "Whisper Flow" throughout. Surnames, subdivisions, title companies and loan programs are all rare words to the model and load bearing words to you. Correct them in what you write from the transcript and leave the transcript unedited.
- Is it safe to rely on AI notes for a commitment made on a call?
- No. Write any promise into the place your business keeps promises on the same day, attached to the person it was made to. A summary is a draft, and it is also the document most likely to have paraphrased the terms.
Worth passing along
This serves the lender and the team leader most directly, because they are the two roles running enough calls a week that manual notes stopped being possible years ago and the risk from an unchecked summary is highest. An agent on two listing appointments a week gets the same lesson in a lower dose. Send it to whoever on your referral side quotes their notes back to clients.
Written by Brett K Moore. We build the PAD System, a records structure for people, assets and deals that lives as plain text files on your own computer. Read what it is.
Read next
How do you use Wispr Flow in a real estate business?
Wispr Flow is dictation that types into whatever application you are already in. For agents, lenders and trades whose hands and eyes are busy all day, that is the whole point. Includes the retention setting almost nobody knows exists, and why finding it late costs you something you cannot get back.
How do you decide whether a tool belongs in your business?
A four box audit for a software stack: what is there and should be, what is missing and should be there, what is there and should go, and what is missing for good reason. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.
Where do your business files really live?
Signed disclosures in the transaction system, listing photos in one cloud drive, headshots on a phone, contracts in email, recordings on a drive in a desk. The fix is one file recording where each kind of thing lives, by service name and folder name, never by filesystem path. What that catches, what it costs, and what it deliberately will not do.