Studio notes — the right data before the agent
This week the agent fleet reworked, for an insurance broker, a policy-comparison agent to keep no file at all, and for a travel agency, turned a decade of scattered spreadsheets into one clean catalog. Two builds, one lesson: know exactly what data your AI touches, and give it the minimum, cleanest version of it.
Every week, the same fleet of agents is at work across several clients at once. This week, two separate builds share the same thread: at an insurance broker, the studio rebuilt an analysis agent to stop keeping the documents it reads; at a travel agency, it turned years of scattered spreadsheets into a single structured catalog. Two opposite moves — retain nothing on one side, organize everything on the other — but the same discipline underneath: know precisely what data your AI touches before you let it touch more of it.
A gap-analysis agent that keeps no file
The agent that compares an insurance broker’s existing policies against a client’s real needs reads uploaded PDF contracts, extracts each clause into a fixed structure — gap category, severity, recommendation, source clause — and returns the report. The studio rebuilt the pipeline as a stateless endpoint: no database, no account system holding onto uploaded contracts “just in case”, a file read once and discarded the moment the analysis is returned. A widely cited 2009 report by Ann Cavoukian, Privacy by Design: The 7 Foundational Principles, frames exactly this choice: privacy protected by default, embedded into the architecture itself, is worth more than a policy promising to delete data later. Lesson for a broker or any advisor handling client contracts: the fewer sensitive documents your systems retain, the smaller your compliance and breach surface — data minimization is not a policy line, it is an architecture decision.
A catalog rebuilt from a decade of spreadsheets
At the travel agency, the actual product knowledge — activities, accommodations, vehicles, pricing — lived in years of Excel files with no shared structure between them, not in any database a booking pipeline could query. Before any AI-driven recommendation could plug into the reservation system, the studio wrote extraction scripts to normalize those spreadsheets into one consistent catalog, validated first against a large synthetic dataset before touching production. Erhard Rahm and Hong Hai Do’s often-cited survey, Data Cleaning: Problems and Current Approaches, documents exactly why this step cannot be skipped: inconsistent, duplicated or poorly structured source data is the single most common cause of failure in any system built on top of it. Lesson for a CIO: a large share of “AI-ready” work in an established business is not model selection — it is excavating and normalizing the legacy data an agent will actually be asked to work from.
What connects both builds this week: an agent is only as trustworthy as the data discipline behind it — sometimes that means storing as little as possible, sometimes it means finally structuring what has piled up for years. Before adding any AI layer, the studio always asks the same question first: what exactly does this agent need to hold onto, and for how long?
Free resource
The self-assessment grid: 20 tasks AI can automate
Sales, admin, support, operations: the 20 tasks AI agents already handle in SMEs — with, for each one, the tell-tale sign that your team is concerned.