Security → this page

How AI is used

There is exactly one place a model is involved: writing the prose around a status report, when a person asks for it on that report. Everything else in Terranes is ordinary software, and the judgement stays with the person accountable for the work.

Where

One feature, named

Narration of a status report. No other feature calls a model, and no feature requires one to complete.

When

Only on an explicit yes

The checkbox is unticked every time. Absent that tick, nothing is sent and there is no transfer — the report still carries prose, assembled from its own figures inside our infrastructure.

Who

Anthropic, under commercial terms

Named on the sub-processor list. Inputs are not used to train models.

Exactly what is sent, and what is not

Sent, when a person asks

Task counts by status
Dates and milestone names
The titles of overdue and blocked tasks
Schedule drift against a captured baseline

Never sent, on any path

Comment bodies
File contents or file text
Stakeholder notes
Budget lines and cost figures

The second list is asserted by a test against the payload’s shape, not by this sentence. If a change made any of those reachable, the build fails.

The controls

Three switches, at three levels.

Per report, per person

The person generating a report chooses whether to ask for AI wording. The choice is not remembered on their behalf, and no per-person usage is recorded.

The scheduled weekly report

Never calls the provider at all. There is no person present to ask, so it takes the algorithmic path — which is why the weekly report can run without any third party seeing your project.

The organisation

A deployment-wide switch turns the whole path off. With it off, the feature is not hidden or greyed — it is not there, and reports keep their prose.

The questions this usually raises

If I say no, does the report come out blank?

No. Every report carries prose either way. Without AI wording it uses the narrative assembled from the report’s own figures, produced inside our own infrastructure and reaching no third party. The difference is the wording, never whether the report explains itself.

How do I know which sentences a model wrote?

Every AI sentence is labelled wherever it appears — in the app, in the export and in the shared report. A reader never has to guess which claim came from the ledger and which came from a model.

Does AI make any decision in the product?

No. It writes prose around figures that already exist. It does not set a status, close a task, assign work, change a date, flag a risk or decide what goes in a report. Where the product cannot verify something — schedule health with no captured baseline, for instance — the report says so plainly rather than having a model fill the gap.

Is our usage measured?

The quota that protects the shared cost is per organisation and is shown before you generate. It is not a per-person record, and it is not reported to anyone as a measure of how anybody worked — see what we track.

No forced AI, and no AI theatre.

If you want it off, it is off — for a report, for the schedule, or for the whole organisation.

Request early access Sub-processor list