Where the data comes from, where it goes

First-party evidence your team has authored. The kind of writing your organisation already produces every quarter, sitting in a dozen tools nobody has time to cross-reference:

• Sales-call notes, win/loss summaries, deal-review write-ups.
• Customer-research transcripts, interview notes, usability findings.
• Support tickets and the resolution notes attached to them.
• Decision memos, RFCs, ADRs, retrospective outcomes.
• Competitor scans, market notes, pricing-page captures.
• System exports your team has chosen to feed in (HubSpot reports, Linear extracts, Notion pages).

Each piece of evidence is something a person on your team wrote, recorded, or chose to bring in. Nothing arrives via a tracking pixel, an inferred profile, or a behavioural inference. If a sentence doesn't have a human author standing behind it, it doesn't belong in the system.

The boundary is sharp on purpose.

• No clickstream. No tracking pixels, no behavioural events, no session recordings, no heatmaps. The methodology has no use for them; the system has no place to put them.
• No silent ingestion of customer data. Customer-system exports (HubSpot, Salesforce, etc.) only enter the wiki when your team explicitly chooses to feed them in. There is no automated scraper running in the background.
• No third-party data brokers. Nothing is bought, enriched, or cross-referenced against an external profile.
• No model training on your data. Whatever LLM provider you bring, their terms apply — but the methodology itself never aggregates customer wikis to train a model. There is no shared model built from anyone's evidence.

The wiki the methodology builds — insights, problem statements, opportunities, recommendations — is yours. It compiles from your own evidence, it cites that evidence on every claim, and it lives in a workspace your team controls.

Recommendations are surfaced as proposals with cited evidence, expected impact, risks, and confidence stated up-front. They flow outwards into the AI builders where prototypes get made (Claude Code, Cursor, Lovable, etc.) carrying the evidence they came from. Nothing flows outward to us, to a third party, or to a shared training corpus.

Self-hosted Winnow runs on your own infrastructure. Your laptop, your server, your VPC. There is no telemetry, no callback, no remote logging. We have no visibility into what you put into it. The only third parties involved are whichever LLM provider you configure (Anthropic, OpenRouter, etc.) — and your relationship with that provider is yours, on your account, governed by their terms.

Hosted Winnow at usewin.now runs the same software in an isolated per-customer container with an isolated per-customer volume. We don't read your workspace data as part of normal operation. The narrow exceptions are listed in the privacy policy: diagnosing a fault you've reported, restoring from a backup at your request, or a valid legal request we cannot decline. Operator access is logged and auditable.

"Harvest" can pattern-match to data-harvesting — the surveillance-flavoured kind, where someone gathers up signals about your customers without their knowing. That isn't what this is.

Think of a farmer harvesting their own crops. The farmer planted the seeds, tended the field, knows what's in it. Harvest time is the moment to gather what they've grown — not to take from someone else's field. The methodology is named for that move: gathering what your team has already written, what your team already knows, into a place where it compounds.

The verb's grammatical object matters, and we keep it disciplined: throughout the methodology, "harvest" sits next to things your team owns and authored. Never customer data. Never user behaviour. Never inferred signals.