How we score

What is scored

The scorer reports 26 criteria, each from 0 to 100. We group them here for reading; the grouping does not affect the calculation.

How the total is built

• The total is not a weighted average of the 26 criteria, and criteria have no fixed weights. It is built from rules that read the transcript.
• It starts from a base of 58 and is adjusted for scenario difficulty: +2 for beginner, 0 for standard, -2 for advanced, -4 for expert and -6 for stress test.
• Rule signals then add or subtract points, for example for a clear opening, needs questions, handling a price objection, a concrete next step, pressure or a forbidden phrase. The result is limited to 0 to 100.
• Safety caps can lower the total regardless of other strengths. For example, ignoring a customer's request not to be called caps it at 35 (62 if the request was softer), and a risky promise caps it at 55.
• Stars are the total divided by 20, rounded up, from 1 to 5. A company administrator may add emphasis to criteria (0.5 to 2). That is shown as a separate weighted number next to the standard total and never replaces it.

Preliminary and final scores

A score is final when it can be compared with other scores. Otherwise it is labelled preliminary.

The AI feedback text in a report explains the result. It does not change the number: the score comes from the deterministic scorer, and a report is checked to confirm the number was not altered after the feedback was added.

Version and scorer artifact

Every report stores the methodology version and the digest of the scorer code that produced it. A score can therefore always be traced to the exact version. Older reports are not recalculated when the version changes.

What the AI does not judge

• Your voice, accent, appearance, emotions or personality. DIRI AI does not recognise emotions: the score works from the transcript and call events of the practice call, and criteria such as empathy look only at what was said.
• Whether a statement is legally or regulatorily correct. The compliance criterion looks for known risky patterns and your own forbidden phrases. It is not legal, insurance or financial advice and not a clearance.
• Your results in real customer calls, your potential, or your suitability for a job.
• Anything the transcript got wrong. Speech recognition errors can change a score.

How the method is checked today

• Developers test the scorer with 15 synthetic conversations, each written in five languages, with expected outcomes that the authors wrote in advance. These expectations are hypotheses, not judgements by human raters.
• There is no comparison with independent human raters yet. It is planned as a separate piece of work, and we intend to publish its results with the methodology version once they exist.
• Until then, treat a score as a coaching signal that helps you choose what to practise next, not as a measurement with a known error.

Intended use and your rights

DIRI AI scores are made for training. They are not a decision about a person.

• Scoring is used to give feedback and suggest practice. DIRI AI does not take decisions about people, and the product is not designed to make decisions based solely on automated processing that produce legal or similarly significant effects (Article 22 GDPR). If an organisation nevertheless based such a decision on a score alone, it would be responsible for that decision.
• DIRI AI is not designed or offered for hiring, pay, promotion, discipline, dismissal, task allocation or for monitoring and evaluating employees' work performance. Using it for those purposes is outside its intended purpose.
• Under the EU AI Act, AI systems intended for those employment purposes are classed as high-risk (Annex III, point 4). DIRI AI has not been assessed or placed on the market as a high-risk system, and we do not claim that it meets those requirements. An organisation that uses it for such purposes acts outside the intended purpose and takes on the obligations that come with that use.
• A participant can see their own report and the explanation behind it. Managers see the reports their role allows. Where your employer runs the workspace, it is usually the controller of your training data and decides who can see reports.
• If you think a transcript, score or report is wrong, you can flag it and ask for a human review at support@diriai.com. Your data-protection rights, including access, rectification and objection, are described in the privacy policy.

Managers: do not use a score as the sole basis for an HR decision. Any decision about a person must rest on human judgement and other evidence, and the person must be able to discuss it.

AI transparency

• The customer in a practice call is an AI-simulated character, not a real person.
• The score is calculated by a deterministic, rule-based scorer from the call transcript (in voice calls the transcript is produced by automatic speech recognition). The explanations in the report are written by an AI model, which can be wrong, so check them against your own call.
• This page is a plain-language summary and is updated with each methodology version.

Related

• Privacy policy
• AI transparency notice
• Ask about the methodology