What comes in, and what stays inside Grain
On a recorded call, scope is described field by field. Here is the list, as it is written in the connector.
Private notes are excluded for a precise reason. That is where someone writes what they think of the person on the other end while that person is still talking. No work question needs that answer, and the connector offers no way to ask for it.
A marked moment comes in under its label, never under the words spoken
Grain's practice comes down to one gesture. During a call, someone pins a passage, gives it a name, and shares the clip with the team. That pinned passage holds the exact words and the name of whoever said them, which makes it the most sensitive material a company stores.
The connector reads the label of the marked moment and stops there. The passage itself, the speakers attached to it, the call transcript, the audio, the video and any private notes taken during the meeting are never requested from Grain. The request sent to Grain is written into the connector and holds three entries only, the summary, the action items and the marked moments.
Sorting work from personal life takes an AI model, and you supply its key. As long as that key is missing, the connector reads nothing at all, and it says so rather than read anyway.
The six situations where a recording is readable by nobody
A recorded call is only filed when the connector can say, without guessing, which department is entitled to read it. Six situations forbid that, and each one is verifiable in Grain by opening the recording concerned.
- The recording is attached to a team in Grain, whichever team that is.
- Its meeting type carries a sharing value the connector does not recognise as restricted.
- None of your own people took part.
- External people outnumber your own people.
- An internal participant has no usable address, a deactivated account for instance.
- Your people on the call belong to several departments, or one of them is unknown to your organisation.
What the fourth situation changes in practice
A call where three external people talk to two of your colleagues is treated as an external meeting, so it stops there, even when both of your colleagues work in the same department. The rule counts who was present and ignores the intentions.
A participant counts as internal when Grain marks them so, or when their address appears in your organisation. Everything else counts as external, including an entry the connector can say nothing about. That doubt pushes towards waiting and never towards opening up. Fireflies covers the same job with a different threshold, and the catalogue lists the other supported sources.
The questions we get asked about this connector
Will an AI be able to go back over a recorded annual review?
It will have neither the audio, nor the video, nor the words spoken, because the connector does not ask for them. The summary written by Grain first goes through a sort that drops whatever reads as personal, and an annual review carries plenty of those signals. If it is judged work-related all the same, it is filed in the department of the participants, and their colleagues in that department can reach it. On an annual review, the safest control is still not to record it.
Do our external contacts end up inside your system?
Their address is used to work out whether the call is too external to be filed, and then it is not written. On the people side, what gets written about a call comes down to the host and the number of participants. If the summary Grain wrote names someone, that name comes in with the summary, and the connector changes nothing about it.
What happens to a recording deleted in Grain?
It is marked in your Kastel as gone at the source, and it stays kept there, readable by the same people as before. Its actual erasure waits for an administrator's decision. That behaviour holds for every Kastel source, and it makes sense on a recorded call, since a deletion can be a slip of the hand that automatic erasure would make irreversible. A deletion request therefore has to be handled in both places.
Can a plugged-in AI change or delete a recording?
No, no write path exists. The token you supply could technically write into Grain, and that is exactly why the access client exposes no write function at all. There is no read-only setting to forget.
The notes attach after the call, so the sweep is repeated
Grain attaches the summary, the action items and the marked moments after the meeting ends. A recording listed too early therefore carries nothing to write, and nothing is written. A call's title alone never triggers a write.
The connector allows for that by design. It looks at the last ninety days by default, and adds a forty-eight-hour overlap at the edge of that window, to go back over the calls whose notes have just landed. Both values are connector settings, which whoever installs it can adjust.
The enumeration follows a pagination marker. When that marker stops advancing, or the number of pages passes the ceiling, the pass halts with an explicit error. An incomplete list would lead the reconciliation to report calls as gone when they still exist.
A recording that loses its notes, or whose summary becomes personal in the sorter's eyes, has its earlier copy withdrawn at once. Keeping content that vanished at the source and withdrawing content that turned personal are two separate rules, and the connector applies both.
What this connector does not do
The summary and the action items are written by Grain, and the label of a marked moment is typed by a person during the call. That label is free text, sometimes rough or ironic, and it comes in as it stands. The connector chooses the fields it reads, it passes no judgement on what they contain.
It does not give the clip back. A plugged-in AI can know that a moment was marked and under what name, without being able to say what was being said in it. To listen again, you go back into Grain, with the rights Grain gives you.
An administrator's review queue can hold a false positive. A recording that is still live, but whose date has dropped out of the observed window, can stay flagged there as gone. That trade-off is deliberate, because the opposite would lose genuine deletion requests, and because nothing is erased without a decision.
The connector edits no clip, shares nothing on your behalf and never writes into Grain. It files what Grain produced where a plugged-in AI can make use of it, within the limits of what its user is entitled to read. What Kastel does with the rest.
Let us look together at what your recordings actually hold.
On recorded calls, the first question is not a technical one. It is about what your teams agreed to have recorded, and about who can find it again afterwards. Write to us and we will tell you plainly where Kastel stops.
Get in touch
