Mercu uses artificial intelligence (AI) to enhance recruiter efficiency and candidate experience - without compromising privacy or security. Our AI roadmap is shaped not by hype, but by clear, practical use cases where AI delivers measurable value and can be deployed safely and responsibly.
Fundamentals
AI Model
Mercu uses OpenAI’s foundation models for all AI-powered features. All data is processed under OpenAI’s Enterprise Privacy Agreement, which ensures that data is not used to train OpenAI’s models. Mercu does not train its own models.
Prompting & Fine-Tuning
Depending on the feature, Mercu uses either:
Account-generic prompts: Predefined prompts used across all customers, which cannot be customised at an account-level (e.g. summarising transcripts of asynchronous interviews).
Account-specific prompts: Customisable prompts tailored during onboarding with our engineering and customer success teams (e.g. scoring candidate responses to behavioural questions).
Model Configurations:
Account-generic prompts are run with conservative settings (e.g. very low temperature, low TopP).
Account-specific prompts can be configured per account, including temperature and TopP.
Logging & Monitoring
Every interaction with OpenAI is logged for traceability and debugging. Logs include:
Input: System prompt and user queryOutput: AI-generated responseCreated_at: TimestampModel: Model used (e.g.gpt-4.1-mini)Tokens: Total tokens used (e.g.3021)Temperature: e.g.0.2TopP: e.g.0.2
A subset of logs is accessible via the Mercu portal. Complete internal logs can also be requested via your account manager for deeper debugging if needed.
Visibility & Evaluation
AI-generated content - such as transcripts, summaries, candidate replies, and reply scoring - is available for review directly in the Mercu portal, ensuring continuous transparency and oversight.
Additionally, we regularly evaluate model outputs to:
Assess the impact of new prompts or model versions
Identify and mitigate model drift
A note on bias
AI, like humans, can carry inherent biases. But at Mercu, we’ve designed our systems to actively minimise their impact.
First, AI never makes final hiring decisions; it can score or flag responses, but it cannot reject or advance a candidate on its own.
Second, we always present AI output alongside full context - for example, a score is shown with the complete candidate response - so recruiters can make informed judgments.
Lastly, all AI scoring is based solely on text transcripts, without access to audio, video, or personal information like name, gender, race, or age (unless voluntarily mentioned by the candidate), reducing the risk of discriminatory bias.
A note on hallucinations
To reduce the risk of AI “hallucinations” (i.e. confident but incorrect responses), we apply several safeguards:
Limited Scope: AI is used only in well-defined contexts - like summarising transcripts, checking relevance, or scoring candidate responses - where prompts and expectations are tightly controlled.
Retrieval-Augmented Generation (RAG): For features like the candidate FAQ assistant, we use a RAG pipeline, meaning AI can only answer questions based on predefined sources (e.g. job description, handbook), reducing the chance of it making things up.
No High-Stakes Autonomy: AI never makes final hiring decisions. Its outputs (e.g. scores or summaries) are always shown alongside full context so recruiters can review and override as needed.
Prompt Design & Monitoring: We use carefully constructed prompts - many of which are account-specific - and monitor model behaviour over time to catch and fix unexpected outputs.
These controls ensure Mercu’s AI stays accurate, grounded, and supportive of fair hiring decisions.
AI features
Mercu currently uses AI across the following features:
1. Relevancy Check
Purpose: Prevent irrelevant or lazy answers to short-answer questions.
AI Role: Classifies answers as semantically relevant or irrelevant.
Prompt Type: Account-generic (not customisable)
Modality: Text
2. Conversational Candidate FAQ Assistant
Purpose: Respond to candidate queries using contextual knowledge (e.g. job description, company policies).
AI Role: Retrieves relevant information and generates responses via RAG (retrieval-augmented generation).
Prompt Type: Account-specific (customisable)
Modality: Text-based conversation
3. Transcription of Asynchronous Video/Voice Interviews
Purpose: Convert candidate-submitted audio/video answers into text.
AI Role: Transcribes speech (does not analyze or score it).
Model: OpenAI Whisper
Prompt Type: Account-generic (not customisable)
Modality: Voice/Video → Text
4. Summarising Transcripts
Purpose: Create a preview summary of transcribed candidate responses.
AI Role: Generates concise, readable summaries of interview transcripts.
Prompt Type: Account-generic (not customisable)
Modality: Text → Text
5. Scoring of Candidate Replies
Purpose: Evaluate quality and relevance of candidate replies across modalities.
AI Role: Assigns a score (1–5) and short rationale based on transcript quality and alignment to the question.
Prompt Type: Account-specific (customisable)
Modality: Text/Voice/Video → Text score + rationale