Trust & data
What we measure, what we refuse, what we don't claim.
Buying an AI product means trusting things you can't see. This page is the part you can check.
Six things the AI cannot do
Send anything
The capability doesn't exist. Drafts are approved by a person and sent from your own mailbox.
Spend without asking
Searches, schedules and reveals each need an explicit yes, every time.
Invent a company
Low confidence means discarded, not shown. Unverifiable means discarded, not shown.
Invent a person
Contacts must trace to a real, current, personal profile, or the lead isn't delivered at all.
Invent a number
Every figure Scout states comes from a live query in that conversation, with its source kind named.
Take orders from an advert
Scraped text is data, never instructions. A defence built in from the start, not patched in later.
How quality is kept honest
- Measured before it ships
- Every AI stage is scored against fixed test sets before changes go live, including full end-to-end runs against the real service, not mocked prompts.
- Knowing when to say "don't know"
- The test sets deliberately include cases where the right answer is no answer: anonymous adverts that can't be traced, companies that don't exist. Answering confidently on those counts as a failure. Abstention is graded.
- Audited daily, automatically
- An automated report runs every day across the whole pipeline: what was found, what was discarded and why, and a random sample of the AI's own reasoning re-checked for whether its verdicts were justified.
- Behaviour pinned by tests
- The commitments on this site, the call-script structure, the spend confirmations, the never-send rule, are pinned by regression tests, so a quiet edit can't silently un-make them.
We don't publish a headline accuracy percentage. Small test sets and moving prompts make single numbers flattering and stale. Ask us for current per-stage figures with dataset sizes and run dates, and we'll send them.
Where the data comes from
- Job adverts
- Licensed job-data providers, queried live, limited to adverts posted in the last 30 days.
- Contact details
- Licensed enrichment providers, named in our privacy policy, used only when you ask for a specific person.
- Signals
- Public company-event sources: curated news feeds, public careers boards, and official government filings.
- Company details
- The company's own website and public profile, read at the time of the search, with a timestamp on what was verified.
Compliance, both directions
For your candidates
- You're the controller; OutputX is your Article 28 processor
- Lawful-basis confirmation gates every CV upload, timestamped as evidence
- Deletion destroys the file, not just the record
- Candidate names are anonymised in outreach by default
For the people you contact
- UK GDPR and PECR framing, with a published retention schedule
- Enrichment providers named in the privacy policy
- Emails revealed one named person at a time, no bulk export
- You send from your own mailbox, under your own sending practice
Things we don't claim
Written down now, so nobody discovers them later.
- Not "a verified email on every lead." Every lead has a named person; not every person has a findable address. In our most recent measured window, reveals succeeded on about half of attempts, and every miss was free. We remeasure continuously.
- Not "real-time" signals. The feed refreshes throughout the day. That's fast enough to be early; it isn't a ticker.
- Not "you'll never see an agency." We identify the client behind agency-posted adverts, but an agency hiring its own internal staff is a genuine employer, and will occasionally appear.
- Not global. UK and US. Signals across three desks. We'd rather be accurate here than vague everywhere.
- Not infallible. A few searches in any batch return little. The commonest cause is an anonymous advert we couldn't confidently trace, and we discard rather than guess. The free demand check exists so you can avoid the cold patches entirely.