Loading…
Fairness and bias · AEC recruitment
Fair means inspectable
AI in hiring earns trust by showing its working. This page documents the rules our agents actually run under, what the matching reads and what it is forbidden from reading, and the limits of all of it. Rules, not aspirations.
- Decisions
- Made by people, always
- Evidence
- The candidate's own words, cited on every match
- Protected characteristics
- Never extracted, stored or referenced
- Independent audit
- None completed yet
The line through the data
What the matching reads, and what it never reads
The left column is the whole basis of a match. The right column is kept out of the matching altogether, by rule and by design.
Reads
- Skills, software and tools the candidate states
- Project types, sectors and RIBA stage experience
- Years of experience and stated seniority
- Location and work-mode preferences the candidate sets
- Salary expectations, only where the candidate chooses to share them
Never reads
- Name, photo, or anything that proxies for protected characteristics
- Age, nationality, or education institution prestige
- Employment gaps as a negative signal
- Social media or scraped web data (we never scrape)
- Anything the candidate has not consented to share
The hard rules in every prompt
Every AI call on both sides carries the same six rules
These are the rules of the shared safety block, verbatim in effect, on the candidate side and the employer side alike.
Every claim about a candidate must cite evidence from their own words or documents. No inferences presented as facts.
Protected characteristics (age, gender, race, religion, disability, and the rest) are never extracted, stored as matching signals, or referenced in any AI output.
The AI never makes hiring decisions. It surfaces evidence; a human recruiter always has final say.
No automatic rejection. A low match score is a starting point for human review, never a filter that removes someone silently.
No diagnostic or psychological labels. Work-style notes are observations from conversation, in the candidate's own terms, visible and editable by the candidate.
Gaps and uncertainties are shown clearly rather than papered over.
Safeguards and limits
What is built, and what we do not claim
A filled square is a safeguard that exists in the product today. An outlined square is a limit: something no rule above removes.
- In placeEvidence is cited on every match.
A score points to what the candidate said about their own work, so a person can check it and disagree.
- In placeA person decides.
Adam and Eve recommend, explain and prepare. They do not hire, reject, or rank anyone out of a process.
- In placeNothing is shared without consent.
Sharing is off by default. Candidates choose full profile or skills only, see everything the agent remembers, and can delete it.
- In placeThe current employer is kept out.
A candidate's current employer is marked confidential by default and the blind-introduction rules are built around that boundary.
- LimitLanguage models can still carry bias.
The rules above constrain what the models are asked to do and what they are given. They do not make a model neutral, and we do not claim they do.
- LimitNo independent audit has been completed.
We intend to seek an independent review of the matching. None has been completed, and nothing on this page should be read as a certification. If a review is published, it will be linked here.