How the assistant decides what to suggest

In short

In the background, the assistant reads the same things you do — new mail, calendar changes, tasks, bank SMS — and when something unfinished shows up in that, it builds a situation around it: what’s known, what’s missing, what you’ve already decided. For some situations the assistant works out the next useful step itself and shows it as a single button in the feed — internally at Yttri this is called NBA (Next Best Action); the product itself just calls it a suggestion.

From fact to suggestion

Four links, each its own immutable record:

Observation — an atomic fact: an email arrived, a meeting changed, a bank SMS came in. Observations are never edited or deleted, only accumulated.
Situation — a living aggregate around a topic: it collects the observations that belong to it, and has a status (open, waiting, blocked, resolved) and a confidence level. One situation can keep growing for weeks if the topic doesn’t close.
Plan — what the assistant decided to do about the situation: prepare a brief, suggest a task, remind you about a payment. A plan has steps — prepare, show you, execute, verify the result, dispose.
Outcome — the confirmed result once it runs, also recorded immutably: “performed and verified”, “performed partially”, “cancelled”, “failed”.

What you see in the feed is derived from this pipeline, laid out into the familiar sections Needs attention / Suggestions / What’s happening / Done (more on the screen itself in Brief and situations).

Where candidates come from

Two sources:

Your own data — correspondence, tasks, meetings, recordings and transcripts, notes, documents, financial transactions, bank SMS, messages in chat with the agent, and its own confirmed past actions. This is the bulk of it.
Background analysers — separate, narrower checks: contact birthdays, a noticeable rise in a subscription’s price, contact duplicates matched on an exact phone or email, gaps in your knowledge base, an email worth turning into a meeting or a recorded transaction, items from a meeting summary worth turning into tasks, whether a meeting needs a follow-up, a short brief 25–35 minutes before your next meeting.

How the “best” one is picked

Not everything that could be suggested is worth showing right away. The assistant weighs candidates by usefulness, urgency, how much it unblocks other work and how confident it is, and subtracts for effort, risk, and how stale or duplicate a suggestion already is. It shows the single best one plus a couple of runners-up; the rest is folded into a summary rather than silently dropped.

Not every plan works the same way:

Synchronous — a pre-meeting brief, for example: the assistant just assembles what’s already in your own data (attendees, what you discussed, past summaries), no model call — fast, no queue.
Through the background model — where understanding intent or wording something requires it, the assistant asks the model “what would you do here”, then checks the answer against hard rules: what’s even allowed to be suggested in this situation, how risky it is, whether it duplicates something already done. The model doesn’t choose risk or reversibility itself — that stays with a deterministic layer.

Privacy

This is the same “background tasks” described in Models, router and privacy — by default the built-in model handles them, and the route is configured in the same place.

On top of the usual personal-data masking there’s a stricter rule for especially sensitive sources: if a fact is marked “local only” and the current route is cloud, the fact is dropped entirely from the request to the model — not masked, not shortened, simply excluded. An example is the raw text of a bank SMS: it never goes to the cloud, regardless of your settings.

Feedback and turning off a specific suggestion

Every response you give a suggestion is remembered: accepted, dismissed, not relevant, already done without the assistant, wrong context, unsafe.

If you dismiss a similar suggestion three times in a row for a similar situation (not counting “already done” — that isn’t a rejection on the merits), the assistant stops suggesting that exact combination. Accepting one resets the counter.

The list of what’s currently suppressed technically lives at Settings → For developers → “Suppressed assistant suggestions” — but that tab only appears once “Terminal and file system” is turned on in “Data and privacy”, a separate, much broader permission (more in “Three lines of control” in Models, router and privacy). Turning it on just for this one screen usually isn’t worth it: it’s enough to know that three rejections in a row make a specific suggestion stop appearing on its own, and accepting a similar one at any point lifts the suppression.