Meeting Notes — 2026-03-19

Participants: Max, Reiner Duration: ~53 min Video: YouTube

Context

Working call between Max and Reiner to sort out what needs to be built before the Zurich pilot kicks off, and to name the product's actual differentiator in a market that now has lots of AI sales/coaching tools. Reiner had a Zurich contact coming the same day; the previous evening he had pitched to a Phoenix Consultancy partner who got unexpectedly excited — that conversation reframed the thinking here.

Hand gestures and pattern recognition

Reiner opened on how to integrate the hand-gesture vocabulary (dominance, eye-level, transparency, curiosity, lightness, connection, emergence/collaboration) into the prototype. Max corrected the frame: the software doesn't see gestures — it reads language patterns from transcript text and maps them to pictures. Swapping icons is trivial; the hard problem is detecting the right pattern from words alone.

• Transcript-only input is a real constraint. Laughter, pauses, lightness of voice don't show up unless the transcription service captures them (hums, hesitations, speaker diarisation). Same audio, two transcripts (raw vs. Claude/ChatGPT-cleaned with speaker separation) produced very different detection results. Transcription quality is load-bearing.
• Max's framing: you can only talk in probabilities. Start with ~5 good real dialogues, measure hit rate (maybe 70%), iterate.

Training the AI with real Zurich calls

Reiner's strong position: do not ship the current prototype to the learning cohort as-is. Variance is too high, trust will collapse. Must train the model on real anonymised Zurich call recordings first.

• Zurich has been asked for 50 recordings. Even if only ~10 land, go through each: run the AI, validate manually, mark what the system got right/wrong, correct.
• Framing for the pilot: agents know they are pioneering / co-developing; AI is "another fallible voice in the system," not a higher authority. Max added: let the trainees tag AI outputs agree/disagree — that gives them agency and also feeds training data back.
• Gap management: set expectations low (it hallucinates, it's experimental) but overdeliver. Don't let the gap feel unbridgeable — people should experience it as "correcting a few glitches" rather than "doesn't work at all."

Design decision: the feature where AI suggests an alternative sentence ("you could have said X instead") for sub-optimal moments is not yet built. Reiner wants it. Max: technically a few minutes to add; quality assurance is the real question.

The Zurich button

Reiner re-anchored the Zurich button concept already in the slideshow design:

• Click it → the current dialogue + analysis is saved into the library instead of being discarded.
• Used both as user-initiated "this one's worth keeping" and as the admin/training intake.
• Every saved dialogue gets reviewed and validated by the team before it seeds detection; unvalidated material corrupts the base.
• Architecture that emerged: two tiers — raw saved + validated repertory. Start with raw saves, add a simple admin review view, promote to validated set.

Max to wire the DB save + a basic admin view where the team can approve/disapprove AI suggestions line by line. This is the immediate build.

Competition scan

Reiner asked the direct question: with so many AI sales-training tools out there (Second Nature, MindTickle, Salesforce Agentforce, Microsoft Copilot for sales, Amy / Coach Hub on the coaching side, various call-centre trainers), what's our USP?

Answer: relational intelligence. Nobody else is building on it. The others optimise a KPI (conversion, NPS, handle time) with conversational AI; we build on a psychological model of the conversation itself.

Secondary points:

• We also need to match the baseline competitors offer (the standard KPIs) or IT departments won't deploy us at scale.
• Deployment and IT-architecture fit becomes the battleground once the early-adopter uniqueness sale lands — IT-architecture specialists in our network could help there later.
• Market-scan task: Max and Reiner each independently read through competitor marketing/docs this week, compare notes next call. Goal is understanding, not copying.

Max's counterpoint on how to use this information: Zurich is not running a vendor bake-off. There is no comparison purchase happening. So competitor analysis is for us — to sharpen positioning and avoid re-inventing — not for them. Audience for the research is just the team.

Relational intelligence as the actual product

Long thread on what makes this different and how it converts to revenue:

• Zurich relationship is itself relational intelligence in practice — trust first, then product. "Not selling improvement on existing; selling the new iPhone."
• KPI connection can't be claimed directly (scandals, attrition, external factors all affect NPS). Max proposed a KPI tree: conversation quality is one branch of NPS, with a quantified weighting hypothesis (e.g. "we claim conversation quality contributes ~20% of NPS"). We measure conversation quality pre/post with our own tool; the organisation measures its own KPIs independently; both can be held up together.
• Second selling point: relational environments lower attrition. People stay in workplaces where they feel met. That's in nobody else's pitch.
• Scaling path sketched: once we have CRM integrations (automate the import, route feedback to sales reps), the methodology can be offered to everyone using that CRM as a customised training layer inside their existing system.

Phoenix Consultancy — the reframe

Reiner's meeting with a Phoenix Consultancy partner the previous evening delivered an unexpected signal.

• She is a buyer of assessment + leadership-training programs. Came in searching for something "like Amy / Coach Hub but cheaper, with more features" to offer their clients.
• They walked her through a Zurich-style one-pager with People's Park branding.
• She said, in effect: "I came in for one thing, you're offering something completely new." Got excited.
• Why it landed: coaching schools carry old luggage (two humans, classical coach modelling). We come in fresh — given AI is at every workplace, how would learning actually happen? — and answer with something organisationally grounded.
• She's now a candidate to run a pilot through her firm. Founders there are expanding and actively looking for an AI-assisted product to distinguish themselves.

Zurich visit — next steps

Our Zurich contact visiting Uben House soon for a couple of days to co-develop. Immediate asks: show current prototype state, secure access to more real recordings.

Sidebar — AI "ignoring" the person in the room

Reiner's broader observation from a recent Claude-based experiment: they fed a recording of their own sales conversation to the model and asked "how could we have been more professional?" It produced sentence-by-sentence corrective rewrites against a business-development script it applied without being asked. That sort of corrective output — targeted at keeping the conversation on eye-level rather than on a generic script — is exactly the alternative-sentence feature we want in the pilot, just pointed at the relational model instead of a sales script.

Action items