Cases from the MBA and the Vanguard · MBA module — using AI to select young innovators when the deadline is faster than the people
Subtitle: Using AI to select young innovators when the deadline is faster than the people
I coordinate programmes for a pan-African network of university innovation hubs, supported by a UN agency. There are 23 hubs in 21 African countries, with five more opening by December. Many of the students and young founders who come through our hubs have never had access to serious training, mentors or funding. One of our newest programmes is meant to change that. It offers an online AI course from MIT, a 14-week entrepreneurship programme, mentoring, and, for some, a place at an in-person bootcamp in Addis Ababa.
For the first cohort, we did the selection the careful, human way. We long-listed 584 applicants. We booked 554 interviews and completed 446 of them over three days, all online. Twelve colleagues, working in six two-person teams, ran the interviews in morning and afternoon shifts, with each team doing 32 interviews a day. At the end, 242 innovators from 22 countries were selected.
It worked, and I am proud of it. But I also saw what it cost. Twelve people set aside their normal work for three full days. Scheduling interviews across time zones, languages and unreliable internet connections took weeks. Some applicants did not show up, and their slots were lost. Afterwards, the selection records had to be cleaned and reconciled, and even our own numbers moved as we checked them. The panel did its best to be consistent, but twelve people will never judge in exactly the same way on day three as on day one.
Now the second cohort is open. Applications run from 28 September to 30 October 2026, and the programme is planned to start in November. That leaves only a few weeks between the day applications close and the day selected innovators need to begin. I have been asked to design a new, logic-based application questionnaire, so that we can judge how applicants think and not only what they write about themselves. The question on my desk is what happens after the questionnaire.
I see three options.
The first is to repeat what we did for Cohort 1: a human panel that interviews everyone who passes the long-list. This is the option people trust. It is also the option I am least sure we can deliver in the time we have, especially if the programme's growing visibility brings more applications than last time.
The second is to let an AI system score the questionnaire answers and produce the long-list, so that people only interview a shorter list. This would save a lot of time. But it means a machine decides who never gets a conversation with a human being. The applicants come from 22 or more countries, write in English or French, and many are writing in their second or third language. I am not sure a scoring model would treat a strong idea written in simple English the same way it treats a weak idea written fluently.
The third option goes further. An AI agent could run the first-round interviews itself: ask questions, follow up, record, transcribe and score. Human panels would only meet the finalists. This is the fastest option, and it would give every applicant the same interview. It also worries me the most. Many of our applicants have never been interviewed by anyone before. I do not know how it would feel to them to be interviewed by a machine, or what it would say about an organisation that exists to open doors.
Underneath the three options is a harder question: accountability. When a human panel says no, I can ask them why. When an AI system says no, who gives the answer? Me, because I designed the process? The vendor? The panel that never saw the application? Our programme is funded on the promise of fairness and inclusion across countries and between women and men. If the selection turns out to favour certain countries, certain universities or a certain kind of English, I will be the one explaining it, and "the model decided" will not be an acceptable answer.
There is also a pressure no one says out loud. Our partners expect the programme to start on time. A delay makes the whole network look slow. A fast, AI-assisted process looks modern and efficient, and it is exactly the kind of innovation we tell young people to embrace. Choosing the slow, human route could look like we do not believe our own message.
I have not decided. I know how to run a human process, and I know it does not scale. I can see how an AI process would scale, and I do not yet know how to stand behind it.
I would open by asking every person at the table to state, before any discussion, which option they would choose: human panel, AI long-list, or AI interviews. I would ask first the person at the table who has most recently hired or selected people at scale, because they will know the real cost of the human route. The fact that would change the room's answer is the number of applications. If Cohort 2 receives about as many as Cohort 1, the human route is hard but possible. If it receives two or three times as many, the room will have to decide what it is willing to give up: time, fairness, or the human conversation.
Today the question is whether AI should help us read applications. Very soon, the AI will not only read and score. It will be able to act: schedule interviews, send rejection emails, invite finalists, and adjust its own questions based on who performs well. At that point, the selection process becomes an agent making hundreds of decisions about young people's futures in a few days, faster than any person can review.
So the real question for a leader is where to draw the line. What can the agent do alone, such as scheduling, reminders and transcription? What must it stop and ask about, such as borderline scores or signs of bias by country or gender? What must always stay with a human, such as the final decision to say no? And what record must the agent keep, so that when an applicant asks "why was I not selected?", someone can give a true answer?
If a machine helps us choose who gets a chance, a human being must still be able to explain every no.
A live case: every round can be improved, and the author's feedback is the next one.
The same two questions, answered twice: first without the mentor's corpus, then from it — the task-to-agent protocol, the Vanguard AI proposal, the Command Layer, the mentor's DSA pilot and HAI5.
Without the mentor's corpus
1. Option two, rebuilt around one rule: the machine may say yes, but only a person may say no. Let the model score the questionnaire and propose the long-list; that part saves the weeks. Then people read, rather than interview, every application the model would drop near the cut-off, plus a random sample of the clear rejections. Reading an answer takes minutes; an interview takes an hour. The safeguard I would not give up is a fairness check before any result is sent: compare pass rates by country, by language and by gender. American employment practice has a simple tripwire for this, the four-fifths rule: if any group's selection rate falls below four-fifths of the best group's, stop and look. Your fear that a strong idea in simple English loses to a weak one written fluently is exactly what that check catches. I would not choose option three for this cohort. For many of these applicants it would be their first interview ever, and an organisation that exists to open doors should not open with a machine. Ask yourself: which no would I be ashamed to explain to the applicant in person?
2. Diane is, and she should say so before anyone asks. Accountability for a process sits with whoever designed and approved it. The vendor answers to her for how its tool performs, and the reviewers answer for the rejections they sign. "The model decided" is not an answer anyone accepts, and data-protection law often does not even allow it. Europe's GDPR (Article 22), Kenya's Data Protection Act (section 35) and South Africa's POPIA (section 71) all limit decisions based solely on automated processing that significantly affect a person. So the rejection itself should tell the applicant three things:
For a programme built to open doors, add one more: what would make the next application stronger. Ask yourself: if an applicant asks "why not me?" a year from now, what record would give a true answer?
Sources: the four-fifths rule, US Uniform Guidelines on Employee Selection Procedures, 29 CFR 1607.4(D); the GDPR, Article 22; Kenya's Data Protection Act 2019, section 35; South Africa's Protection of Personal Information Act, section 71.
From the mentor's corpus
1. The mentor's task-to-agent protocol sorts this selection before any option is chosen. It has four kinds of work:
Read that way, option two is the protocol applied, and option three hands Craft to a machine. The safeguard comes from the mentor's own design for clinical AI agents, the Vanguard AI proposal: audit the agents' performance separately for each subgroup, because averages hide exactly the cases that matter. Here that means country, language and gender, checked before any result leaves. Ask yourself: which part of my selection is craft, and have I handed any of it to the machine?
2. The mentor's Command Layer chapter answers the first half: delegation does not dissolve responsibility. The subordinate answers for the action, and the commander answers for the intent, the boundaries and the authorisation. Here the model is the subordinate and Diane is the commander, with the vendor answering to her. The second half has a working example in the mentor's reference pilot on the Digital Services Act. There, a marketplace's AI removes a listing because a certificate number does not match. The seller contests the decision, and three people reconstruct it from the evidence each party keeps in its own book. The DSA already requires platforms to send a statement of reasons with every such decision: the facts it rests on, whether automated means were used, and how to contest it. Give every rejected applicant the same statement, and keep the evidence so that a person can re-read it. The mentor's Vanguard AI design adds the form: every recommendation arrives as a structured report of rationale, evidence and confidence, so the reviewer sees an applicant, not a score. Ask yourself: could my rejection letter survive being read by the applicant, the funder and a journalist on the same day?
On the NEO Turn. The case draws the line itself, and the mentor's HAI5 framework gives each side of it a name:
What the agent must stop and ask about is written in advance: a borderline score, or a pass rate that drifts by country, language or gender. The mentor's teleology magnifier warns what happens if the agent is given only the deadline: it will magnify speed. Write the purpose into its charter instead: no qualified applicant lost to language, country or a bad connection.
Sources: the Vanguard Task-to-Agent Mapping Protocol; the Vanguard AI proposal (subgroup audit, the structured report); Vanguard Leadership, vol. 2, the Command Layer; the mentor's reference pilot on the Digital Services Act (cotrugli.tech/demo/dsa) and the DSA, Article 17; the HAI5 framework; the mentor's glossary, the teleology magnifier.
My final note. Make the application demanding: case-style tasks for the candidates. When they send their answers, AI can evaluate them.
Then hold one Zoom session with everyone, however many there are, lasting one to four hours. Split it into groups inside the same call, so that each of the people you have available runs a subgroup. If you have twenty people, each can be the assessor for ten to twenty candidates. In that session every candidate has one minute in their subgroup to introduce themselves, and after that there are all kinds of possibilities for tests.
When it is all over, announce an optional one-hour session. For me, those who stay have better chances to go further.
In any case, work to make the process fair, with people making the final decision. I would like to see the final version, and if you want, I can also moderate the process.
The mentor's note turns the three options into a fourth, better than all three: a demanding task scored by AI, then one live session in which every candidate meets a person. This is how I would run it.
Before the session.
The session. One call of one to four hours, in subgroups of ten to twenty with one assessor each. Twenty assessors can see two to four hundred candidates in an afternoon; last time, twelve people needed three days for 446.
The optional hour. Keep the mentor's test of commitment, and make it fair across the continent. Africa's clocks span only UTC−1 to UTC+4, so one slot in the early afternoon, UTC, is reasonable almost everywhere. A candidate whose connection drops gets a written make-up task the same day, not a lost place.
The decision. The panel meets once, with one page per candidate: the task score and its reasons, the live answer, the assessor's note, and whether they stayed for the hour. People sign every yes and every no, and every no goes out with its statement of reasons.
Ask yourself: what would a candidate who did not get in say about how we chose: that it was fast, or that it was fair?
Sources: the mentor's final note above; round 1 of this chapter; the live room at cotrugli.tech.
A question for the table, a disagreement, what you would have done. The case lead reads every comment; the ones the table takes up enter the chapter as questions from the room, with your name.