The role of AI
AI does a few narrow jobs in our work.
What AI does here, and what people do
We use general-purpose AI models, of the kind anyone can use. What is ours is what they work inside: the rubrics, drawn from developmental research and long practice; the interview and its sequence; the instructions we write for each job; and the people who read what the AI produces and decide what to do with it.
One idea runs through how we work with AI, both in the assessment and in how we built it. People come up with what is wanted and specify it. The AI puts it into words. People check what it produced against what was meant, and stay accountable for it. Our colleague Mo Raei's essays on authorship with AI helped us name what we had been doing: "Who Wrote This? Rethinking Authorship" and "Who Wrote This? The Plot Thickens".
Why use AI at all?
Because of what it makes possible. A person can take the interview when they are ready, with no interviewer to schedule. Their interview has been analyzed against every rubric before the mentor begins, so the mentor's time goes into the conversation with them. And for us: without AI, a team of our size could not have built this in the time it has taken.
We do not claim that AI makes the assessment more objective, more consistent or less biased than a person would be. We have no evidence of that, and we say only what we can show.
What a person meets
Movement by movement. Where something is still being built, we say so.
The interview
The interview is conducted by an AI interviewer. The person is told so in writing before it begins, and the interviewer introduces itself as an AI.
The interviewer speaks its questions in a synthetic voice and shows them as text. The person can answer by speaking or by typing. Spoken answers are turned into text as they go, and only the text is assessed.
It follows a sequence we wrote, in ordinary language, one thing per question. It asks follow-up questions to understand what the person means, but it does not react, judge or give feedback, and it scores nothing. The analysis happens afterwards.
The measurement and the analysis
The measurement. Our own software counts features of the transcript: how sentences are built, causal and conditional links, concrete and abstract language, references to time, and how far a line of thought is followed. It uses standard, openly available language-analysis tools. No generative AI is involved: the measurement writes nothing and judges nothing. Its figures go to the analyses that use them.
The analysis. After the interview, a general-purpose language model analyzes the transcript several times, each time for one job: once for each aspect of the structure of thinking, then once for each skill domain. Each analysis works from one rubric and from instructions our people wrote. The analyses of the structure of thinking are brought together before the skill domains are analyzed.
The report. The AI then writes the report from the analyses: a summary first, then each skill domain with examples from the person's own interview, practices to try, and space for their own reflections.
Learning loops
A person can work through a learning loop in text, with no AI, or in audio mode, talking it through with an AI practice partner. The practice partner introduces itself as an AI, takes the person through the loop's stages by voice, and follows instructions we wrote.
How it is bounded
What the AI does is set by people, and kept to that:
- 1
Each job is narrow.
The interviewer asks. Each analysis works from one rubric. The report writing writes from the analyses.
- 2
People write what the AI works from.
The rubrics, the interview's sequence and the instructions are written by our people from developmental research and practice. We do not train the models; we instruct them.
- 3
Only text is assessed.
The assessment does not analyze a person's voice, face or other physical signals, and it does not try to read their emotions.
- 4
Nothing is ranked or recommended.
The assessment informs the judgment of people, and people make the decisions.
- 5
The mentor reads the report before the person does.
- 6
Participants' data is not used to train the providers' models.
How each of these is done in detail (the instructions, the rubrics' level descriptions, how the analyses are made and checked) we share under a confidentiality agreement, or with clients. Talk with us →
The EU AI Act
Our assessment uses AI, and we tell everyone who takes it so. Under the EU AI Act, using AI to inform decisions about hiring or promotion is a high-risk use. The Act's requirements for it apply from 2 December 2027, and we are preparing for them.
Which providers we use, and where they process data: Sub-processors → How we handle data: Data handling →
How we built it
The rubrics and the interview came first, out of years of developmental assessment, debriefing and learning design with leaders.
Our work with AI began with writing. Over time, that practice became how we built the platform.
Mo Raei's distinctions describe what we found. Who has an idea and who puts it into words are different questions. The person specifies what the text must hold, the AI writes, and the person verifies the result against what was meant and stays accountable for it. Over long use the model is not retrained, but it works through a context the person has shaped, which Raei calls an "interactional imprint".
When things began
-
February 2026
The first AI feature on our platform: questions back to a person on an essay they had written.
-
June 2026
The AI interviewer, the measurement, and the first analyses against the rubrics, joined into one sequence that month.
-
July 2026
The AI practice partner in learning loops.
-
September 2026
The fourth version of the interview, with a written reason for each design choice.
-
October 2026
The first full run of a step-by-step analysis of the structure of thinking, in development.
What we are building
- Analyzing the structure of thinking step by step, so that each step can be checked. In development
- Checking each analysis automatically against the measurement. In development
- More versions of the interview, matched to the scope of different roles. In development
- The skill library portal's AI design partner, to help a person build learning loops for the skills they choose. In development
See for yourself what a person meets.
Ask to try the assessment →