Products AI Lab Manifesto About Talk to the team
AI Lab  ·  Last updated 17 May 2026

Thinking out loud.

The Lab is the unpolished page. In-progress products, research notes, methodology iterations. We publish here while the work is still warm, so you can see where we're going before we get there.

What we're working on this week

Active investigations.

NOTE · 014 Week of 11 May
Agent liability.

When an AI agent posts a journal entry, who signs it? Working through the audit-trail design for Helios 1. Every agent action inside our internal operating system needs a traceable provenance that survives a financial audit when we ship the customer version (Helios 2). Hard problem; we'd rather get it right than get it shipped.

In review
NOTE · 013 Week of 04 May
Dispatch memory.

Senior dispatchers remember the truck that always breaks down on cold mornings. New dispatchers don't. Building the working-memory layer for Helios 2 that gives a new dispatcher senior-dispatcher context on every call. Prototype is live inside Helios 1 first.

Prototype 0.3
NOTE · 012 Week of 28 Apr
Methodology baseline.

Re-running the AI Methodology against three concurrent engagements to see where the variance lives within the 30–40% blended reduction. Hypothesis: the bottleneck moves from discovery (deeply compressed) to change-management (less compressed). Data inconclusive so far.

Mid-study
NOTE · 011 Week of 21 Apr
FarShip extensibility.

Two customers want to extend FarShip in different directions. Designing a "regulated extension" interface that lets customers add modules without breaking the upgrade path. Reference architecture: Stripe Connect.

RFC draft
NOTE · 010 Week of 14 Apr
Voice as form.

Field technicians take 14 minutes on average to close a work order in the D365 mobile app. Voice-to-structured-form on the truck radio compressed it to 90 seconds in a four-tech pilot. Generalization is the open question.

Pilot · 4 techs
NOTE · 009 Week of 07 Apr
Master-data sanity.

Customers ask us for AI agents, but their master data isn't consistent enough for agents to be honest with. Building a pre-flight diagnostic that scores master-data readiness before any agent work begins. Brutal feedback expected; we'll publish it anyway.

Tool draft
Methodology  ·  In iteration

Versioning the methodology.

The AI Implementation Methodology is the operationalized form of the velocity-compression thesis. We version it like software. Each release records what we learned in production: what worked, what didn't, where the bottleneck moved.

v3.2  ·  April 2026

Pre-flight diagnostic.

Added · Master-data readiness scoring

Across the last three engagements, master-data quality emerged as the strongest predictor of timeline variance. v3.2 introduces a pre-flight diagnostic that scores readiness before kickoff, so we can sequence data-cleansing into the project plan instead of discovering it in week 6.

v3.1  ·  February 2026

Voice-first config.

Changed · How we capture customer requirements

Switched from PowerPoint-driven requirements workshops to voice-recorded operator interviews on-site. Two-hour interview with the actual operator beats a six-hour workshop with their boss. AI transcription and structured extraction handle the synthesis.

Essays  ·  Long-form

Reading from the Lab.

Lab  ·  Hiring
2 open roles

We're hiring people who think about the second-order effects of agents in industrial operations. Senior architects, applied researchers, product engineers.

See open roles