This lesson was rewritten after a strategic review on April 16, 2026. If you were shown an older version, the shape of the bet has changed — fewer bets, sharper thesis, one direction explicitly parked.
The thesis, restated
PPL Spark builds AI that works on real conversations inside defined systems.
"Real" rules out roleplay-in-cloudy-nowhere-land (Amy, CoachHub and similar): we don't simulate conversations for practice, we work on conversations that are actually happening. "Defined systems" means organizations, teams, regulated industries — places where conversations have repeatable structure, measurable outcomes, and named participants. Predictability is what makes the AI useful; structure is what makes the value provable.
Zurich Kundencenter is the sharpest example of the shape: a call center with measured NPS, defined conversation types, a staffed team. Not a niche — a template.
What we parked
The compliance product direction — AI that reads conversations for regulatory violations — is no longer on the roadmap. It's commodity territory owned by cloud giants (Microsoft, AWS, Google) where the winning play is scale and procurement, not craft. Wrong fit for a tailored, niche team. We'd rather lose the sector than wrong-fit into it.
Two live product motions
Two working prototypes are competing for focus. Both fit the thesis. They pull in different directions.
Motion 1 — Live Companion
AI accompanying a live conversation in real time: transcript with speaker diarization, per-speaker intention tracking, agenda progression, next-move suggestions to the host. Tested on internal calls, on a YouTube mock interview, and — most relevantly — on the April 16 review itself, where it tracked a three-way conversation with a single mic.
Why it matters: external AI investors, shown a prototype, flagged this as the strongest value — intelligent real-time intervention in conversations that would otherwise drift.
Competitive pressure: Notion shipped a similar feature in April. Zoom and Teams will follow. The moat has to be the art of dialogue slant — developmental depth, not just transcripts and summaries — and it has to be built quickly.
Expansion ideas under consideration:
- Per-participant agents ("You've been silent for 5 minutes — bring a creative thought").
- A Dojo: teams legitimately record meetings; each member privately replays and improves their own contributions.
- Chairperson-specific functions — flagging decisions taken vs. decisions still vague.
Motion 2 — Agent-based modeling
Feed a document, an issue, a system → extract the actors → build a persona graph → simulate their conversation → produce a report → let the user chat with any agent or poll the group. Inspired by Farmer's Making Sense of Chaos lineage of bottom-up complex-system simulation.
What we've tested:
- MiCA regulation. 49 actors: central banks, journalists, owners, regulators — including Polymarket-style private-vs-public stance modeling.
- Personal-network scenario. 20 people in Max's network reacting to a "pivot from ecological to fintech" prompt. A simulated mother replied about spring weather — plausible real-life noise, not a bug.
- Iran–US (failed). ~1,000 actors — too expensive to simulate at this scale. Useful ceiling.
Why it matters commercially: the immediate deal flow runs through here.
- Convector / Colin (EU chemical directive). Model the directive's decision process, map coalition options, strengthen Colin's lobbying proposal.
- Convector 360 assessment. 25 leaders, cross-evaluated, projected as "who they'd be together in five years." Reframes a proposal as an ongoing leadership modeling base, not a one-off assessment.
- Phoenix Consultancy / Khalud. Leaders preparing for difficult conversations — rehearse the counter-party as an agent before walking in.
The open risk (flagged internally): any predictive claim has to be backtested on known past events before it's credible. We don't ship forecasting without track record.
The tension — and why it's not resolved yet
| Motion | Pull | Evidence |
|---|---|---|
| Live Companion | Investor pull. External AI investors say this is the massive value-add. | Prototype works. Market validation from outside the team. |
| Agent-based modeling | Near-term revenue pull. Maha's network has three named deals. | Prototype works. Active commercial conversations. |
Investor enthusiasm and near-term revenue point at different products. They are not the same product. A decision is pending, and this lesson deliberately doesn't pretend otherwise.
RAIner, in the new frame
RAIner — the developmental-dialogue product covered in lesson 3 — still sits inside the thesis. It's the longest-cycle bet: AI on a real coaching conversation, producing a growth trajectory. The go-to-market is unchanged in shape:
- HR platform. Mid-to-large HR and L&D teams, months-long sales cycles, ~$20/seat/mo or $5K–$25K/year per org, higher LTV.
- Coach enablement. Individual coaches scaling beyond 1:1 — days-to-weeks decisions, lower LTV, higher volume, distribution moat.
Both are still live as GTM tracks for RAIner. They are vehicles within the thesis, not the thesis itself.
Positioning against the field
Across all three product motions, the gap we occupy is the same:
- Static assessments (MBTI, DiSC): cheap, scalable, produce labels — no trajectory and no live signal.
- Human-led enterprise coaching (BetterUp, CoachHub): deep, expensive, hard to scale, detached from the actual conversation.
- Generic AI (ChatGPT, Replika): conversational, no developmental framework, generic advice, no context on the defined system.
- Transcript tools (Notion, Otter, Fireflies): summarize what was said. Don't know what should have been said.
The wedge across all four: developmental AI working on real conversations inside a defined system. Scalable where humans aren't, grounded where static tools aren't, contextual where generic AI isn't.
Sector plays — proof points, not the market
- Zurich Kundencenter. Anchor enterprise pilot. Real call center, measurable NPS, defined conversations. The shape we want to reproduce — not the market itself. Low-hanging fruit for the thesis; not the whole meal.
- Trade Republic / adjacent signals. Financial services call centers reportedly shifting from chatbots back to trained humans. More "Zurich-shaped" prospects in the pipeline.
- Ukraine — Bureviy. RAIner4Army for Ukrainian Army recruitment, battalion → brigade (Azov Corps, 5 brigades) → National Guard, eventual NATO read-across. Funding path via Emerge VC.
The $750K ask
$750K seed to reach 8,000 subscribers and 25 enterprise clients by end of Year 2, with path to profitability projected in Q3 of Year 2. The pitch leads with the team — Rainer v. Leoprechting (developmental psychology), Max Semenchuk (product), Maha Alusi (operations), Robin Sverd (research) — on the theory that the dominant risk at this stage is execution, not idea.
Key takeaways
- The thesis is AI on real conversations in defined systems. Everything else is a vehicle.
- Compliance is parked. Wrong fit.
- Two live product motions: Live Companion (investor pull) and Agent-based modeling (near-term revenue). Decision pending.
- Zurich is a template, not a niche.
- RAIner's HR and coach GTM continues as the long-cycle motion.
- Predictive claims from modeling need backtesting before they ship as claims.
Last lesson: how the team actually works — the practice that made April's reframing possible — and what to do after this course if you want to go deeper.