Supported by LU Innovation, Lund UniversityKompis · Public Report
Pilot IIReport
6 – 27 July 2026 · 85 participants · Public findings
Period
July 2026
Prepared by
Steve Emmanuel
Founder, Kompis
Cohort
85
newcomers practising Swedish across 21 days
Key takeaways
- 1. Engagement deepened rather than decayed. The median participant had 46 conversations over the pilot — and interactions per user rose every single week, from ~19 on day 1 to 75 by day 20. No novelty drop-off.
- 2. Speaking anxiety fell and confidence rose. Among 29 participants measured at both entry and end, average anxiety dropped from 0.35 to 0.21 on a normalised scale — a 40% reduction — while self-reported speaking confidence reached 3.6 out of 5.
- 3. Retention is won in the first 72 hours. Re-engagement prompts recovered only ~12% of lapsed users across 966 prompts — people who stay, stay early.
The problem
Newcomers in Sweden aren't blocked by lack of learning material. They're blocked by speaking confidence — they understand more Swedish than they can bring themselves to use, and they freeze in real situations. SFI teaches the language. Almost nothing addresses the fear.
The freeze
“Kompis worked perfectly for me — I was able to converse, every answer was timely, and the conversations were practical.”
What we built
Kompis is an AI conversation companion. No lessons, no grammar drills, no scores. You talk — by voice or text — and it talks back, remembers you, and reaches out when you go quiet.
It's deliberately built to be the lowest-pressure place in your life to speak Swedish badly.
The pilot
21 days. Organic waitlist recruitment. Measured at entry, mid-point, and end.
Four things we measured
Four questions, four answers — each one designed to be readable on its own.
Did people keep showing up?
Half the cohort (51%) was still active in the pilot's final 7 days · median 46 conversations per participant · engagement rose across all 21 days, never plateauing.
Did they choose to speak?
~90% used voice or mixed voice+text · a cohort defined by fear of speaking chose to speak anyway.
Did confidence move?
Among 29 participants with matched before-and-after measures: average anxiety fell from 0.35 to 0.21 — a 40% reduction — and day-21 speaking confidence averaged 3.6/5. Gains were broad, not driven by a few.
Did any of it leave the app?
Reported in real life: a participant who understood his forklift trainer and spoke back on a certification retake he'd earlier failed; a spouse who joined after watching her husband improve at home.
Three things that surprised us
Engagement climbed instead of decaying. We planned for novelty to wear off. It never did — interactions per user rose every week to the end.
Voice won in an anxious cohort. ~90% chose to speak rather than type, in a group defined by freezing when they have to talk.
Win-backs barely work. Only ~12% of lapsed users returned after a nudge, across 966 prompts — which is why retention has to be built in the first days.
In their words
“I recently did my truck and forklift training in Gothenburg. I'd failed the first attempt, but this time I could understand the trainer better and speak too, however badly. My truckkort is arriving this week.”
“It is good talking with the bot, but I would love to extend the time.”
“Sometimes the questions the bot asks are repeated, which is a bit annoying.”
“It would be great to get a text while we send voice messages, and a place to see my new words to build vocabulary.”
What we'd do differently — five things
- 1Build in transcripts — the single most-requested feature, raised independently by multiple participants.
- 2Fix onboarding so no one gets trapped at the first step.
- 3Treat cross-session memory as core — users noticed the moment it failed.
- 4Let users choose the voice.
- 5Add a vocabulary tracker so progress is visible.
What's next
Kompis now goes offline. Rather than keep a half-finished product running, we're rebuilding around exactly what participants asked for: transcripts of what they said, memory that actually holds across sessions, a choice of voice, and vocabulary tracking so progress is visible. When those are in place, Kompis returns for Pilot III — with the same open recruitment and a design that measures transfer into real life, not just time spent in the app.
Get in touch: hello@kompis.site
Method & data note
104 participants onboarded; 85 completed a baseline measure. Speaking confidence and anxiety self-reported using validated instruments at entry, mid-point and end. Outcome figures drawn from participants with matched before-and-after measurements; N stated with each claim. Engagement data is pseudonymous; quotes published with consent. Full methodology available on request.