What changes when the AI shares your environment

Harmonic Design runs retros at the end of every engagement. The stickies get written, people say the right things, and the next project starts with the same problems. The gap wasn't in the retros — it was in the journey from retro to change. A colleague and I had been working that gap for months: combing Miro boards, pulling transcripts, interviewing people to clarify ambiguous stickies, grouping findings into 40 themes across 9 categories.

Then, in a studio Spotlight, we asked the whole team to rate every theme on two scales: severity (1–4) and frequency (1–7). Twenty-five people, forty themes, two scales — 2,000 possible cells. We got 821 actual ratings. That's a dataset that deserved real statistical handling. I rebuilt it as a synthetic dataset first (so I could use my own tools without exposing internal Harmonic data), then ran it through two AI surfaces to see what would happen.

I started the way most people start. Chat tab open, spreadsheet open, copying and pasting back and forth. Within a couple of hours I had a bubble chart, two heatmaps, a filtered list of the seven critical themes, and an outline for a presentation. Useful. Entirely produced by my keystrokes.

Three moments from that round show what this working mode is actually like. The bubble chart was wrong — Sheets was reading the vote-count column as a data series instead of the size dimension. The chatbot walked me through the menu changes; I clicked them. One bubble remained hidden behind another no matter what the chatbot suggested. It couldn't reach into my chart. Then the chatbot started building a story I had to stop: a heatmap framed around Leads "absorbing all the pain." I stopped and asked why we were showing count instead of average. Count was biased by participation rate, not severity. That catch wasn't the AI. That catch was my own critical reasoning running fast enough to notice before the story set.

What I came out with: a bubble chart that wasn't quite telling a story, two heatmaps in a Google Sheet, and a text document outlining what a presentation might look like. It felt like having a patient helpdesk on the other end of a chat window who could answer any question about Sheets but couldn't actually open Sheets.

For round two I used Claude Code from my terminal, with access to my filesystem, my repos, and a deploy target. Same CSV, same goal. The first thing I noticed was that I stopped describing chart types and started describing what I wanted to learn.

Audience is leadership. Use whatever charts and structure best answer those questions. End with discussion questions, not conclusions. Don't fabricate data. Show me what you compute and where it comes from. When you pick a visualization, briefly explain why you chose it.

That last sentence was the one that mattered. Asking the model to justify each visualization gave me something to push back against — and a running tutorial on what chart shape fits what question. The agent read the CSV directly, computed aggregations in Python, built an interactive HTML site, deployed it to a public URL, and copied both files to my laptop. I never opened a spreadsheet. I never wrote a formula. I have still never opened a Python file.

The prompts that changed the deliverable weren't "make a bubble chart." They were: "The role-average bars aren't telling me anything. Show me what categories each role rated highest, and whether those overlap." "The frequency scale is bucketed labels — an average of 3.5 means 'between most projects and once per project,' not 'daily friction.' Make the prose reflect that." That last edit would have been an afternoon with the chatbot. With the agent, it was one pass across nine sections in seconds.

The difference isn't that the agent produces better outputs, though it often does. The difference is where the bottleneck lives. With the chatbot, the limit is execution: how well I can drive Sheets, how cleanly I can move output between the AI's window and my work. I am the bridge. With the agent, that bridge collapses. We share the same environment. The bottleneck shifts from execution to articulation — how clearly I can describe what I want, what I notice when something's wrong, and when to push back.

That is a different muscle. And it is closer to the muscle service designers already have. Briefing. Critiquing. Naming what's missing. Articulating the shape of a good answer before you have it. The transition from round one to round two isn't a technical leap. It's a framing shift about where your attention goes.