Pilot programme

Case study: Magyar Telekom Plc.

8,801 answers1,714 lessons

Telekom Pilot Programme

Pilot Thesis

Knowledge is a trajectory, not a snapshot.

Traditional assessment takes one reading of something that never stops moving and calls it the result. Even an accurate reading shows only where a learner stands, and nothing of how they got there, how their skills are developing, or how much confidence the result deserves.

Vergil measures the trajectory itself. As a learner works through an organisation's own material, every interaction updates a running estimate of how well each concept is understood, and how certain that estimate is. Every answer is recorded and every estimate is auditable, so what emerges is not a score to take on faith but a picture of their skill development, with the evidence attached.

This is the thesis Telekom set out to test. Across the pilot, Vergil tracked skill development concept by concept and observation by observation, so the team could see not only who knew what, but why the platform believed it.

“What I found really interesting was that you're no longer looking at just one score. You can actually see how someone's skills develop over time, which is much closer to how learning works in reality.”

— Tamás SzegediHead of Learning and Development, Telekom

One scale, one environment

One scale, one environment: Measurement made possible

Every learner starts from the same neutral point and is measured on the same scale, in the same environment. That is what makes skill levels comparable, between one person and the next and between whole cohorts.

“We deliberately chose two groups with different backgrounds, and the results reflected those differences very clearly. It was clear that the measurement was reflecting the actual capabilities of the participants.”

— Tamás SzegediHead of Learning and Development, Telekom

Eight skills, both teams, one scale

The neutral startFurther along

LLM limits and risks

Group BGroup A

Human-in-the-loop practice

Financial AI integration

Effective prompt writing

Designing agentic AI

Advanced prompt engineering

Applying AI to text tasks

Foundations of language models

Group A · the more technical teamGroup B
Eight skills, each plotted for both groups on one scale, drawn from the neutral state every learner starts in upward. Group A, the more technical team, stands further along than Group B on every one of the eight, and the segment between each pair is the distance between the two groups on that skill. In order of where Group A stood, highest first: llm limits and risks, human-in-the-loop practice, financial ai integration, effective prompt writing, designing agentic ai, advanced prompt engineering, applying ai to text tasks, foundations of language models. This is a described difference rather than a group effect.

Every answer does two jobs

Every answer does two jobs: The flywheel effect

Learning builds skills and helps retain them. At Vergil an answer does a second job at the same time: it sharpens how the course adapts, and how precisely the platform can measure.

Spaced repetition is how a foundation gets solid. When an idea came back in a later lesson, results rose and stayed above where they started, and the platform ended 47% more certain of what each learner understood than for an idea it met only once.

Assessment results as an idea came back · 80–100%

85.7
97.0
92.4
86.1
94.6
96.6
FirstSecondThird
Prompt qualityNext-word prediction
Two knowledge elements, each measured three times as it returned in a later lesson, as the average share of the question answered correctly on an axis running 80 to 100 per cent. The effect of prompt quality on the output: 85.7, 97, 92.4 per cent. How language models predict the next word: 86.1, 94.6, 96.6 per cent. Both rose and stayed above where they started.

Model uncertainty across knowledge elements

The neutral startFurther along

An idea met once

An idea that came back

47% more certain

The wider the band, the less certain the platform was

Two bands on the same scale. An idea met once, and an idea that came back in a later lesson. Each band is centred on the platform’s estimate for that group of knowledge elements, and its width is the uncertainty left on that estimate. Both sit at almost the same place, and the returning one’s band is roughly half as wide: 47% less of the platform’s uncertainty about what each learner understood.

What the learners said

What the learners said: Learner satisfaction

Learning needs the right tools and the right environment behind it. Vergil is built in close contact with the people using it and improves on that footing, so the material in front of a learner is the only thing asking for their attention.

“I'd make it compulsory for everyone. I loved every minute of it.”

— A pilot learner

9.0 / 10

Logical and simple to navigate

7.9 / 10

Confident they could apply what they learned

7.6 / 10

Would recommend it to a colleague

Knowledge measured. Method proven.