Human oversight

The same user and process flow is shown twice, so a governance team can pick the format that suits its review. Both use three definitions:

• In the loop a person reviews and acts on each output before downstream use.
• On the loop people monitor a running process and can intervene.
• Over the loop people set and periodically review system-level configuration without reviewing individual outputs.

Option A · Flowchart in three lanes

Figure 2A. Governance decisions (top) are made by your admin and reviewed periodically. Automated steps (middle) run inside those rules and keep their output on the platform; the optional alert digest is the one automated output people receive, and it is monitored rather than approved item by item. Every request (bottom) ends with a person deciding whether the output is used.

Option B · Sequence diagram

Figure 2B. Reading down: setup messages are governed over the loop; the scheduled block runs automatically and its output stays on the platform; every request ends with the user reviewing the output before anything is saved or shared; the optional alert digest is the one automated message that reaches people, and they monitor it rather than approve each one.

Table 2 · Oversight touchpoints

Touchpoint	Who	Oversight model	What the person does	How it is enforced
Choosing which public figures, handles, newsrooms and feeds are monitored, and how often	Workspace admin	Over the loop	Sets and periodically reviews the source list. Does not review individual retrieved posts.	Only admins can add or change sources; retrieval runs only for configured sources.
Choosing model providers and models per feature	Workspace admin	Over the loop	Selects providers and models; reviews the usage audit.	Per-feature model settings in the workspace; every call logged with provider and model.
Scheduled retrieval, storage, tagging, embedding, persona-context refresh, and newsroom source discovery during onboarding	System	Over the loop	None per item. Output is internal working data. During onboarding, a source-discovery job proposes a company's newsroom URL with a model and web search, verifies each candidate by fetching it and checking that its host belongs to the named company, and attaches the best verified match as a daily crawl source. If no candidate passes, nothing is attached and the Workspace Status page shows the gap. The control is the admin's review of the source list (first row), where a discovered source can be changed or removed.	Runs only inside admin-configured rules; outputs stay on the platform, and the search queries and page fetches involved are listed on the Sub-processors page.
Alert digests and scheduled report emails to Slack, email, webhook or in-app	Recipients your admin chose	On the loop	Read the digest or report as it arrives; pause, retune or disable the schedule. An admin can preview or run an alert by hand before scheduling it.	Both features are off unless enabled for the workspace; admin controls cadence, content and recipients.
Creating a persona or audience	Workspace admin or member	In the loop	Reviews the drafted profile (typed, AI-drafted, imported, or installed from the catalogue) and confirms it before it can be used; edits it at any time afterwards.	Requires the persona-write permission. Exception: personas auto-generated during workspace setup are created directly and reviewed afterwards.
A persona reaction, grade, chat answer, narrative report or content draft	The requesting user	In the loop	Reads the output, judges it, then accepts, refines and reruns, or discards.	Output is only displayed. It is saved to the workspace and can be exported (PDF, PPTX, DOCX) or shared by a link the user creates and can revoke; no feature publishes it externally.
Submitting content for testing (a draft press release, a question)	The requesting user	In the loop	Decides what to submit, knowing it becomes part of the prompt sent to the model provider.	Nothing is submitted on the user's behalf; library sync is opt-in and read-only.