Day 51: Complexity and Regulation
Day two of the workshop in Zurich. If day one was discovery, day two was design.
The Onboarding Architecture
The problem we're solving: when a new customer signs up for a scheduling platform, they need to configure their entire work environment from scratch. Who works when, what shift patterns exist, what are the legal and contractual constraints, which employees can do which shifts. This configuration currently takes weeks of back-and-forth with an onboarding specialist.
The vision: a conversational agent that guides new customers through this setup. You upload your old schedule (in whatever format — PDF export from legacy software, Excel spreadsheet, Word document). The agent extracts the structure: shift types, employees, groups, constraints. Asks clarifying questions where the data is ambiguous. Generates an initial configuration. Produces a test schedule to validate the setup. Hands off to the customer when everything looks right.
What emerged from today's design session is a clear separation of concerns. One component handles document analysis and data extraction — the messy, format-dependent part where PDFs from 2005 need to yield structured data. A second component handles the conversational onboarding flow — asking the right questions, validating constraints, generating schedules, explaining tradeoffs. These two things sound related but have completely different failure modes, so keeping them separate turns out to be important.
Where AI Earns Its Keep
The most useful insight from two days of workshops: AI doesn't need to handle the full complexity of workforce scheduling to be valuable. The constraint space is genuinely difficult — decades of operational research, NP-complete optimization problems, regional legal requirements, human preferences that conflict. A general-purpose AI agent will struggle with this and probably hallucinate rules.
But AI is extremely good at the earlier steps: extracting structure from unstructured data, asking the right follow-up questions, detecting when a constraint is contradictory, routing edge cases to a human specialist. The hardest scheduling math stays in specialized software. AI handles the setup conversation that comes before it.
This is a useful frame for thinking about where AI agents belong in complex business processes generally. Not replacing the specialized system, but making it accessible without a specialist present.
The EU Cloud Act
Between workshop sessions, someone mentioned the EU Cloud Act — coming into force later this year. The short version: enterprise cloud software used in the EU will need to demonstrate that EU customer data cannot be compelled by non-EU governments. US CLOUD Act exposure (American authorities compelling US-based cloud providers to hand over data) becomes a compliance risk for EU enterprises.
I hadn't thought about this explicitly, but it's essentially a tailwind for Pinchy. Self-hosted, data stays on your infrastructure, no US-based cloud dependency. Everything Pinchy is already designed to be becomes a compliance checkmark for EU enterprises navigating this regulation.
There's still work to do — proper documentation of the data flows, explicit certifications, ensuring Pinchy's optional cloud-based providers (Anthropic, OpenAI) are properly scoped and can be replaced with EU alternatives. But the architecture is right. The direction is right. The regulation is coming, and Pinchy is better positioned than most alternatives.
Day 50
Fifty days. A complex onboarding automation designed. A regulation that turns out to be good news. And the clearest articulation yet of where AI agents actually add value in enterprise software: not by replacing the hard specialized systems, but by making them accessible without a specialist.