MartialQA uses AI to score customer conversations for correctness. That AI reads real conversations between real people, so we hold it to clear rules. This page is those rules, in plain language: what our AI does, what it never does, and who is accountable for it.
Last updated: August 2026
AI scores conversations and drafts coaching. It never disciplines, ranks for dismissal, or replaces a manager's judgement. Decisions about people are made by people who know them.
Correctness is judged against your own documented policies and answers, not the model's general opinion. If your knowledge base does not cover it, we flag it rather than guess.
We do not use your conversations to train models, and our AI providers are contractually barred from doing so. Your data is processed to produce your scores, and that is all.
No black-box numbers. Each score is traceable to the relevant scorecard criteria and, where applicable, the customer knowledge it was graded against, so anyone can check the work.
Agents see their own scores and the reasoning behind them, and can dispute a score through their manager. Disputes are reviewed by a person, and confirmed errors correct the record and feed back into calibration.
We process conversations for quality assurance and nothing else, in line with POPIA. Data shared for a free audit is kept to an agreed sample and deleted from our systems within 14 days.
Transcription and scoring run on established model providers, each listed with its purpose, the data involved and its processing location on our Subprocessors page. AI providers are marked as such, and we update the page before adding a new one.
We routinely compare AI scores against human reviews, watch for drift when your knowledge base or scorecard changes, and recalibrate when the two disagree. A scoring system nobody audits is exactly the problem we sell against.
MartialQA scores all eligible conversations ingested into the platform against the same scorecard and the same knowledge base. No cherry-picked samples, no reviewer mood, no favourites. Where accents or audio quality affect transcription, that is treated as our problem to fix, not the agent's.
AI-generated transcription and scoring can be incorrect. Accuracy may be affected by audio quality, language, accents, incomplete customer knowledge, ambiguous conversations and model behaviour. Scores should be reviewable by people rather than treated as infallible, which is why agents can see and dispute them.
Conversations are processed under the safeguards in our security programme and Data Processing Addendum: encryption in transit and at rest, EU hosting, access limited by role, and redaction options for audit samples.
Responsible AI at MartialQA is owned by our founding team, reviewed as part of our security programme, and this page changes when our practice changes. Questions or concerns about an AI-generated score go to info@martialqa.co.za and are answered by a person.
MartialQA is not designed to make solely automated employment decisions that produce legal or similarly significant effects on people. AI-generated scores are intended to support quality assurance and coaching, and any decision affecting a person requires human judgement, consistent with the safeguards POPIA section 71 sets out for automated decision-making about a person's performance at work.