How Scout uses AI with your data

Scout uses AI models for a narrow set of judgements about text, and for nothing else. This article says which parts of Scout involve a model, what Scout has committed to with the companies that provide those models, and where the line sits between what a model may decide and what only code or a person may decide.

The commitments, first

Zero data retention. Every AI model interaction runs under zero-data-retention terms: the provider does not keep your content after answering.
No training. Those same terms forbid your content being used to train or improve a model. Scout does not train general models on your data either.
Named sub-processors. Every company involved in running Scout, including the model providers, is named in the Privacy Policy. That list is the authoritative one and is kept current.
Your data is never sold, and never shared with another Scout workspace.

What a model is allowed to do

Decide whether something is relevant to your business. A message has to be read to know whether it concerns a customer, a quote, an order, or an invoice. To make that determination, message text may be sent to the model provider under the terms above. A message that does not match your business is not turned into an item and is not shown to anyone.
Pull out typed, quoted facts. Where a message does matter, a model extracts specific values — a date, an amount, a purchase-order number — and must quote the text it took each one from. An unquoted value is discarded.
Propose that two records are the same. Where exact rules cannot settle whether two customers are one customer, a model may propose the match. It is a proposal: a person confirms it or keeps the records separate. See Identity matches.
Write one short assessment of an account's day, built only from facts that were already extracted and cited. These assessments carry no message bodies and no attachments.
Phrase things. The wording of a prepared draft, and the sentence describing a habit Scout learned about an account.

What a model is never allowed to do

Arithmetic about money. Every dollar figure in Scout is computed by code from your own records.
Date math. Business days, cutoffs, and overdue calculations are deterministic. A due date must come from evidence with a supported basis, and Scout would rather show no date than a guessed one.
Deciding what you see. Whether a point becomes visible is decided by one gate in code, which applies the confidence floor and the visibility rules in one place.
Changing a record anywhere. Every connection is read-only. See Scout reads, it never writes.
Sending anything. No model output leaves Scout as a message to an outside party. Scout prepares — you send.
Asserting that something did not happen. "No reply in three days" and "this purchase order was never acknowledged" are computed by code. A model may read such a finding; it may not invent one.

The short version: models judge content, code controls consequences.

Where the work happens

None of this runs while you are looking at a page. Scout does its reading and its thinking on a schedule — mostly overnight on your workspace's own clock — and the app then renders what was already decided. Nothing on your screen is being generated as you read it, which is also why the Inbox is fast and why the same page shows the same thing to two people at once.

Personal mail gets a narrower path

Original message bodies are handled in one tightly bounded place: they are opened only to extract cited facts, and the plaintext is discarded as soon as that finishes. Later stages work from the extracted facts, never from the original message again. The short excerpts that end up as evidence on a point do persist with that point, because evidence is the whole promise — but they are the minimum needed to act, and who may read them is governed by the rules in Who can see what.

Good to know

Scout records what a model was asked to do and what it cost, per workspace. It does not keep the message text to do so.
A model's confidence is part of Scout's own gate, not something you tune. A proposal Scout is unsure about is either withheld or presented as a question rather than a fact.
When you mark a point Wrong, you retire that line of reasoning for that account. That is the most direct control you have over Scout's judgement. See Rating a point.
Automatic sending to your customers is not a feature you can switch on. The setting that would enable it is refused by Scout's own database, not merely hidden in the interface.