AI Left the Chat Window. Now It's Doing Real Work.

For the past year or so I've written about using AI as a set of advisory teams, for financial management, and even for cooking. Each of those involved me typing questions and getting text back. Useful, even transformative in places, but still fundamentally a conversation. Then I started using Claude Cowork.

Cowork is a desktop application that gives Claude access to your files, your browser, and your local environment - with your explicit permission. If you're looking for a how-to guide, there are good ones available. What I want to share is what it actually feels like to use day-to-day, because the shift from chatbot to desktop assistant is bigger than I expected. As I noted in a recent post, if you haven't used AI lately you don't know what you're missing. This is another exhibit.

One practical note before the examples: give Cowork access to a dedicated folder, not your entire drive. And before handing it any file to work on, put a copy in that folder rather than the original. Simple discipline, eliminates a lot of potential anxiety.

Work

Complex Report Analysis: A friend was building some new business metrics that required amalgamating data from P&L statements, sales reports, and quality records going back several years. I uploaded the various reports to Cowork and described the metrics and their calculations that required input from different reports. Within a few minutes it had crunched the numbers and built a formatted Excel workbook with color visuals, a summary tab, a data and calculations tab, and a tab explaining each metric and its business relevance. It also suggested a couple of additional metrics based on the context I'd described. In the past this would have taken hours of data wrangling across multiple systems before even opening Excel. Easily a five-hour savings, probably more.

Office Pain: The more mundane Excel work is equally painless. We've all inherited spreadsheets that have been copy-pasted together until they're full of duplicate rows, then spent twenty minutes trying to remember how to de-dupe them. Now I just upload the file and describe what I want. Same for pivot tables, complex formulas, or any operation where the real problem is remembering the syntax rather than knowing what you want. PDF-to-Word or PDF-to-Excel conversions follow the same pattern. Microsoft's built-in conversion has always been mediocre, producing something that requires more cleanup than starting from scratch. Cowork handles these in seconds, almost always requiring zero additional work.

Home

Photo archives: I had thousands of old photos as electronic files, accumulated over decades of travel and life events. I put copies of the folders in my Cowork folder and asked it to organize them by topic and date, then rename each file from the generic "IMG_xxxxx.jpg" to something descriptive in three words. I expected something like "mountain-snow-tree.jpg," which would have been a modest improvement. Instead, Claude recognized specific locations, people, and context, producing filenames like "kitzbuhl-ski-kevin.jpg." I then asked it to select the 20 best photos from each folder and put copies in a /Favorites subfolder for a potential photo book. I didn't define "best." What it chose, based on its read of composition, quality, and uniqueness, aligned almost exactly with my own aesthetic judgment. That part was a little unsettling, honestly.

Home maintenance: Our house requires a lot of regular maintenance: tree trimming, fruit tree fertilizing, professional window washing (coastal salt air is not kind to glass), furnace checks, chimney sweep, and so forth. I uploaded three years of maintenance records and receipts and asked Cowork to create a maintenance calendar. It didn't simply copy the prior dates, which were often irregular because I had forgotten. It recognized seasonal patterns and suggested optimal timing for each task, and offered to add them directly to my calendar.

Medical and life history: Medical records are a chronic headache when you're seeing a new doctor and can't quite remember when something happened. With two decades of scanned records, I asked Cowork to build a health timeline for my wife and me. It was quick enough that I kept going: addresses we've lived at, cars we've owned over the years. It brought back some memories, and quietly corrected a few others.

Landscaping and irrigation: The most layered home project involved our landscaping, which is extensive and varied. I used Google Maps for an aerial view of the property, then walked the gardens taking photos from labeled perspectives. Cowork created a property map, recognized and labeled the plants (asking me rather than guessing when it wasn't certain, which I had specifically instructed), and overlaid our ten-zone sprinkler system. From that it suggested a revised watering regimen based on plant types, recommended some changes to zone coverage, and proposed seasonal variations. Next step, once I get further into agentic AI, is tying this into a smart irrigation controller connected to the weather forecast. Baby steps.

The broader point

One smaller but telling example: calorie tracking. For years I used MyFitnessPal, laboriously searching for approximations of home-cooked meals and estimating portions. Now I can take a few seconds to just describe what I ate, including custom or modified recipes that I can upload, and get a complete nutritional breakdown of the entire day or a specific recipe in seconds. No app required.

Which raises a larger issue. Two of these examples, business metrics and nutritional tracking, previously required dedicated software with subscriptions, onboarding, and ongoing friction. That software has a significant problem. AI is increasingly replacing point solutions with a general capability that doesn't need its own interface or business model. The market has noticed, which is part of why software stocks have taken a hit lately.

Cowork is not a chatbot. It reads your files, writes files, browses the web, and connects to your calendar. For me it's been a meaningful jump in what AI can accomplish in a workday.

Agentic AI is the next step, and it's a significant one. Everything described above still requires me to initiate each task. I upload a file, I describe what I want, I review the output. Agentic AI removes that dependency. Instead of me asking Cowork to suggest a watering schedule, an agentic system would monitor the weather forecast, decide when to adjust irrigation, send the command to the controller, and log what it did, all without me asking. Instead of me uploading maintenance receipts to build a calendar, an agent could monitor my email for service confirmations, update the schedule automatically, and remind me when something is overdue. The difference between a capable assistant and an autonomous one is not just efficiency. It's a different relationship with the work entirely.

I'm not there yet. But the irrigation project, the maintenance calendar, the photo organization, all of them point toward what becomes possible when AI doesn't wait to be asked.