On Monday, Anthropic made its S-1 public. That’s the giant document a company files with the SEC before it can sell shares on the stock market. It plans to list on Nasdaq under the ticker ANT, and reports say the IPO could raise around $100 billion, which would make it the biggest ever.

Note that some final numbers, like share price, aren’t in the filing yet. That will come later in an updated filing ahead of the IPO.

1. The growth is real, and it’s getting more efficient to deliver

Anthropic’s revenue went from $1.2 billion to $16.3 billion, comparing the first half of 2025 to the first half of 2026. That’s about 14x in a year. For context, it made $4.6 billion in all of 2025.

The more interesting number is what happened to its biggest cost. Over the same period, spending on computing power (the chips and cloud capacity that train and run Claude) went up only about 4x, from $2.3 billion to $8.9 billion.

In plain terms: revenue is growing much faster than the cost of the thing it sells. However, because Anthropic books much of its computing capacity in advance, this means some hefty upfront spending commitments, especially when it comes to the current demand landscape for compute power.

It also says each new model makes the work cheaper for customers. Its latest model, Opus 5.5, cut customers’ cost per task by about 40%. When the work gets cheaper, people use more of it.

One caveat: Anthropic has already signed up for far more computing power than it uses today (more on that below). The revenue growth outpacing the cost of compute growth story only holds if revenue keeps outrunning that bill.

2. It’s less dependent on a few big customers than previously reported

Earlier this year, reports based on a draft of the filing said about 20% of Anthropic’s revenue came from just two customers.

However, the latest public filing paints a broader picture:

• No single customer made up more than 10% of revenue in the first nine months of 2026.

• About 79% of Q3 2026 revenue came from customers outside its top five, up from about 64% in Q1 2026.

• Customers spending $100,000+ a year grew from more than 950 at the end of September 2025 to more than 11,900 by end of September 2026. Over half the Fortune 500 is in that group.

The filing still lists customer concentration as a risk, and its biggest customers are growing too: each of its top five hit a record high last quarter. But “two customers are a fifth of the business” doesn’t appear to be the story anymore.

3. Every layer of The AI Village I previously illustrated shows up as a documented risk factor

Now, I won’t pretend to have read a ton of S-1’s. But many times throughout this document I found myself thinking “didn’t I just read this section?” Part of that was the structure of the S-1, but part of it was just how often compute was cited as a potential risk factor to the business.

In my AI Village post, I mapped the AI industry as a village with layers: apps on top, then models, then cloud and compute, then chips, then data centers and power at the foundation. Anthropic sits in the model layer. Its filing reads like a tour of every layer below it, and each one shows up as a risk.

The cloud: landlords who are also competitors

Anthropic rents much of its computing power from Amazon, Google and Microsoft. In its own words, these companies “simultaneously serve as our investors, customers, distribution partners, and competitors.”

In the Village post I pointed out that Google’s Gemini competes with Claude while Google sells Anthropic computing power. The filing confirms it and goes further. Google is also an investor as well as a customer. Nearly half of Anthropic’s revenue (48% in the first half of 2026) comes in through Amazon’s, Google’s and Microsoft’s cloud stores, where companies buy Claude on their regular cloud bill.

The catch is who gets the computing power when there isn’t enough to go around. The filing says its ability to get compute “is subject to allocation decisions by our CSPs” (cloud providers), and admits: “we have in the past sought quantities of compute from CSPs in excess of what those partners were willing to provide.”

Chips and memory: a handful of makers

The filing names the companies the whole industry depends on. TSMC makes most of the world’s most advanced chips, and SK hynix, Samsung and Micron make most of the specialized memory AI chips need. It says TSMC has “at times, been unable to fully satisfy market demand,” and as a result Anthropic has “not always been able to obtain sufficient quantities of chips on a timely basis.”

Data centers: building its own

Anthropic used to rent all of its computing power. Now it’s also leasing its own chips and placing them in data centers that partners build and run to its specs. It calls this “Anthropic Direct,” and its first facility came online this year. The goal is partly to depend less on those cloud landlords.

Power: the foundation

These data centers need enormous amounts of electricity, and power is often the hardest part to secure. Anthropic doesn’t generally build power plants itself. Instead, it signs long-term deals with utilities for dedicated power, or agrees to arrangements where power is built for its facilities. Either way, it commits years ahead, and for years at a time, before it knows exactly how much demand there will be.

That timing is part of the risk. Power deals take years to put in place and to bring online, and the filing also notes that local opposition to data centers “has also become more organized and coordinated,” which can slow or block a project even after the money is committed.

The common thread: booking years in advance

Across every layer, suppliers want commitments up front. So Anthropic has to predict how much computing power it’ll need years from now and sign deals in advance, “often with limited ability to adjust commitments if our needs change.”

Those compute forecasts are inherently uncertain. The filing points to research pace, how compute-intensive training new models is, customer demand, and how fast hardware gets more efficient.

Anthropic admits that “In the past, we have significantly underestimated our compute requirements, which left us capacity constrained.”

The filing lists individual deals like at least $110 billion with Amazon, about $114.5 billion with Google and $44.9 billion with Microsoft, many payable “regardless of actual usage.” It leaves the grand total blank for now. The Wall Street Journal says the infrastructure costs could rise to $750 billion based on disclosed deals that include chip leases.

4. If you buy the stock, you technically get a vote, but you don’t get a say

Anthropic’s setup is unusual even by tech-company standards:

• The seven co-founders will share a single special share that carries 50.1% of the votes on regular shareholder matters. Whatever they vote for wins.

• An independent trust of five people with backgrounds in AI safety, national security and public policy picks a majority of the board. Its trustees have no financial stake in the company, and they don’t run it day to day. Their power is choosing directors.

• Public shareholders get one vote per share, but they can’t outvote the founders, and they don’t vote on the trust’s board seats

• Big investors like Amazon and Google mostly hold non-voting shares.

The filing is upfront about the trade-off. The structure “may effectively eliminate the ability of holders of our Class A common stock to influence important decisions, including a change in control.” In practice, nobody can buy Anthropic unless the trust and founders agree.

Why build it this way? Anthropic is a “public benefit corporation” whose stated purpose is “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.” The structure is meant to protect that mission from short-term pressure from shareholders.

There’s also a philanthropic piece: 5% of the company’s shares went to a fund for public-benefit work, and each co-founder pledged to give away 80% of their personal stake.

The short version: buying Anthropic stock means you’d be buying a piece of the business, not a say in how it’s run.

5. Overall velocity, and Claude Cowork Adoption

I feel like Claude Code gets a lot of buzz as a hugely popular coding tool. Admittedly, I was an early Claude Code adopter and have spent a lot more time there than in Claude Cowork. So I was a bit surprised to see the initial adoption numbers of Cowork outpacing those of Claude Code: the filing states that it “outpaced the adoption Claude Code achieved at the equivalent point” after launch.

The pace at which Anthropic releases also seems pretty unparalleled as far as I’m aware. They’ve done 250 launches in nine months. That can mean new models, features, or apps. And that equates to more than one every business day.

6. Not a ton of focus on open-weight model competitors

One risk I thought the filing underplayed, and this part is my own speculation, is the potential risk of cheaper open-weight models eating into Claude’s model business.

Open-weight means the model’s weights (the numbers that make it work) are published, so anyone can download and run them, often much more cheaply. That doesn’t mean everything about the model is public, and it doesn’t mean they’re all from one place. DeepSeek is the best-known example right now, but the category also includes models from Meta, Mistral, and others.

Across more than 90 pages of risk factors, the S-1 gives that specific pricing pressure essentially one sentence, plus a note elsewhere that open models compete for work that prioritizes cost efficiency over intelligence.

Developer platforms like OpenRouter and Vercel AI Gateway do show a good portion of their routed traffic going to open-weight models. But those charts are not representative of the full market picture here: they only measure tokens routed through those platforms, not revenue across the whole market. Anthropic’s enterprise traffic mostly wouldn’t show up there, and Anthropic’s own numbers show those big customers continue to increase their spend on Claude as of end of Q3 2026.

So I’m not calling this an immediate threat. It’s simply a data point I’m keeping an eye on in the long-term.

7. Chip stocks had a rough few weeks, then bounced on filing day

After a rough couple of weeks, the semiconductor ETF SOXX closed up almost 3% on Monday, October 26. Chip stocks like Nvidia, Micron, Broadcom, and AMD also closed up ~2 - 3%.

I’m curious to see how the rest of the AI industry stocks react to Anthropic’s IPO. If Anthropic prices well in November, does it lift the whole sector?

The big picture

First, congratulations to everyone at Anthropic!

So what does this filing actually tell us? For now, revenue is growing much faster than the cost of compute, but Anthropic is locking in that compute years ahead, before it knows what demand will look like. Possible future constraints on compute, chips, and power are real risk factors to the business. And it’s a real-life example of how competitors can also be your investors and your customers in the AI industry.

Disclaimer: This post is for education only, not investment advice. I’m not a financial advisor, and nothing here is a recommendation to buy or sell any stock.

Sources:

1. Anthropic S-1 Filing

2. WSJ: Anthropic’s Infrastructure Costs Could Top $750 Billion in Coming Years

3. OpenRouter dashboard

4. Vercel blog: Open-weight models take 56% of token volume, Astra doubles Fable 5.1 spend

I wanted an easier way to see, at a glance, what each of these companies actually does, how they fit together, and why they keep doing deals with each other. Mostly so I could form my own opinion about their businesses instead of nodding along to headlines.

So I had ChatGPT help me draw one. Welcome to The AI Village 🏡 🤖.

Why this matters (even if you don’t care about AI)

The headlines are full of NVIDIA, OpenAI, Microsoft, Meta, and enormous deals between companies you thought were competitors. It’s hard to tell who makes what, who needs whom, or what any of it has to do with the app on your phone.

If you’re thinking, “Cool, but I mostly use ChatGPT for recipes,” here’s the twist: your money may be more involved than you are. By one JPMorgan analysis, 41 AI-linked companies made up about 49% of the S&P 500’s value as of early October 2026.

So, if you have a 401(k) or an index fund that tracks the S&P 500, you’re probably already more involved with AI than you realized.

One request, five layers

Let’s follow one ordinary request through the village:

Help me plan a weekend in Chicago. Find a cute hotel under $250 a night, near restaurants and an art museum, and suggest an itinerary.

Somewhere between “cute hotel” and “please keep it under $250 a night”, you’ve tapped into model builders, chip designers, cloud providers, and power companies. You weren’t thinking about any of them. You just wanted a hotel.

We’ll start at the top of the village, with the app you actually use, and work down to the foundations.

One thing the map can’t show: this isn’t a relay race where your message climbs down five floors and back. Every request leans on all five layers at once. The app runs on a model, the model runs in the cloud, the cloud runs on chips, and the chips need a building with power and cooling. Each terrace shows what that layer needs from the one below it. And some companies (e.g., Google) own property in more than one neighborhood.

Apps & agents: the storefronts

ChatGPT, Claude, or your AI assistant of choice is the front door. The app is where you type, upload a file, read a response, and use whatever tools it offers.

Like a travel agency’s storefront, it bundles several capabilities into one experience. You don’t need to know which system handles each step to ask for a travel itinerary.

Many AI assistants and agents can also take action for you, from running a search to filling out a form, depending on the tools and permissions they’ve been given. For our Chicago trip, that could mean handling the hotel search and comparison, and, with your permission, the booking.

The storefront and the model behind it sometimes come from different companies. Take Cursor, an AI coding app: you can work inside Cursor while using a model built by Anthropic. Cursor supplies the workspace and tools; Anthropic supplies the model. In village terms, that’s a shop in the apps neighborhood paying the model-builders neighborhood every time you use their model.

Examples of who lives here: assistants like ChatGPT, Gemini, and Claude; Cursor for coding; Perplexity for search; and a fast-growing crowd of agents like Meta’s Muse.

Model builders turn your request into a response

Behind the storefront is the model: the system that reads your request and writes the response. Think of it as the itinerary planner at the desk. For our trip, it’s juggling your budget, location, dates, and the kind of weekend you want.

A model’s own knowledge is frozen at the point it was trained, so it doesn’t carry live hotel prices in its head. It can look them up with tools like web search, but it can still slip up and search the wrong thing, misread a source, or answer from memory when it should have checked for more current information.

When it does search, the handoff goes roughly like this: the model asks for a search, the app runs it, the results come back, and the model picks up from there, sometimes over several rounds. (Simplified, but that’s the shape of it.)

Two terms will come up in this series:

• Training is how a model learns its capabilities in the first place. It happens before you ever get to use the model, and it’s an enormously expensive and compute-intensive process.

• Inference is the model using what it learned in training to answer a new request. Your Chicago request is inference, and every request like it draws computing power from the layers below: the cloud, the chips, and the buildings that house them.

You’ll also see model names that sound like code names: Claude Opus 5.5, GPT-6.1 Sol, Muse Spark, etc. Most are a family name plus a version number. The family usually signals the tier (most capable, balanced, or fast and cheap), and the number marks the generation. Your apps usually pick a default model for you, so you don’t need to track these, unless you’re building software, where the choice can matter a lot.

Examples of who lives here: OpenAI, Anthropic, Google, Meta, Mistral, etc.

Cloud & compute: the rented workspace

The model needs computers to run on: rows of servers packed with processors, memory, and storage, linked by high-speed networking and housed in data centers. “Compute” is the industry’s shorthand for all that capacity.

Cloud providers rent that capacity out. It’s like renting a fully equipped office instead of building and running your own building. In practice, AI companies mix and match: they buy some servers, rent some capacity, and lock in more through long-term contracts.

That capacity isn’t infinite. When demand spikes past what’s available, requests queue up and responses slow down, which is one reason providers set usage limits or sell different levels of access. (Not every limit means the system is maxed out. Usage limits are also product and pricing decisions.) For your Chicago trip, this is the rented desk space where your request actually gets worked on.

Examples of who lives here: the big clouds (AWS, Microsoft Azure, Google Cloud, and Oracle), plus specialists that rent out AI computing power, like CoreWeave, Nebius, and Lambda.

So what’s actually inside that workspace? The next two layers: the chips doing the work, and the buildings, power, and cooling that keep them running.

Chips & hardware: the machines doing the work

When we say “computers,” picture rows of powerful machines in tall cabinets called servers. They have the same basic parts as your laptop, just configured (and multiplied) to run AI:

1. A processor does the calculations.

2. Memory keeps information it’s working on close at hand.

3. Storage holds information longer term.

4. Networking moves it between machines.

Think of a desk: the processor is you doing the work, memory is your desktop, storage is the filing cabinet, and networking connects you to the other desks. A bigger filing cabinet won’t fix a crowded desktop.

You’ve probably heard about NVIDIA’s GPUs: processors that are good at running many calculations at once, which makes them useful for AI. AMD makes competing GPUs, and Cerebras takes a different approach, building one giant chip instead of many small ones. But the hardware story goes well beyond processors. Memory comes from SK hynix Samsung, and Micron. TSMC manufactures many of the chips other companies design. And companies like Broadcom supply the networking.

Most companies specialize in one or two of these, and switching specialties takes years and billions of dollars. Some of the big players keep expanding into their neighbors’ territory: Google, Amazon, and Meta all design their own AI chips, and OpenAI recently unveiled its first, Jalapeño, built with Broadcom.

Examples of who lives here:

• Chip design: NVIDIA, AMD, Cerebras, and Broadcom

• Manufacturing: TSMC and Samsung

• Memory: SK hynix, Samsung, and Micron

• Networking: Broadcom and NVIDIA

Physical infrastructure keeps the village running

All those machines have to live somewhere, and that somewhere is a data center: a building, or a campus of buildings, filled with rows of servers. A data center supplies what those machines can’t run without: reliable electricity, cooling to keep them from overheating, and two kinds of connections. One goes to the power grid. The other goes to the fiber-optic cables that carry your request over the internet and send the answer back to the app on your phone or computer.

Some data centers are built and run by the big cloud companies themselves. Others belong to operators that rent out space, power, and cooling to many customers at once, a bit like a landlord for servers.

Data centers aren’t new. Their roots go back to dedicated computer rooms in the 1940s, and the modern versions of them have quietly handled much of your digital life for decades: websites, email, online shopping, streaming, social media, video calls, and the cloud backup holding your camera roll. Data centers already sit behind most of your digital life.

What’s new is how fast AI is adding to the load. According to the International Energy Agency, data centers’ electricity use worldwide grew 17% in 2025. At AI-focused data centers, it grew 50%. The IEA expects total data center electricity use to roughly double by 2030. (Those figures measure electricity, not new buildings.)

Examples of who lives here:

• Data center operators: Equinix

• Cooling and power-equipment makers: Vertiv

• Power companies: Constellation

Why the same companies keep making deals with each other

Now the headlines get easier to read. When two AI companies announce a deal, one of four things is usually changing hands:

• Equipment: the machines and components that do the work.

• Compute: access to someone else’s computing capacity.

• Model access: the right to use a model inside a product.

• Equity: an ownership stake in the other company.

For example, CoreWeave has an agreement to supply AI infrastructure to OpenAI, while also becoming an investor. OpenAI is the customer, buying computing capacity, and it also took a stake in Coreweave. Amazon’s relationship with Anthropic is similar: an investment plus Anthropic’s use of AWS infrastructure.

Companies can also compete in one neighborhood while doing business in another. Google, for example, builds Gemini, which competes directly with Anthropic’s Claude model. Google has also invested in Anthropic and sells it compute capacity. That’s why the village labels describe roles, rather than assigning every company one permanent address.

When the next deal appears in your feed, ask:

1. What’s changing hands: equipment, compute, model access, or ownership?

2. What does that let the buyer do that they couldn’t before?

Next up: a deep dive into data centers

Your Chicago itinerary is the visible tip of all this activity. Behind it are software decisions, rented capacity, specialized machines, and a lot of electricity. Once you can see the layers, you can make sense of both the tool in front of you and the industry behind it.

Next up in this series, we’ll start at the bottom of the village: data centers, and the power and cooling that keep them running.

Disclaimer: This post is for education only, not investment advice. I’m not a financial advisor, and nothing here is a recommendation to buy or sell any stock. I own shares in some of the companies mentioned here.

I didn’t hire a coach. Someday I probably will, but I’d like to see how far my current AI training system can take me first. I didn’t run 50-mile weeks; my biggest week was 25 miles. What I had was a system: a training plan, a steady stream of data from my Garmin, and an AI I talked to almost every day about all of it.

AI took the thinking load off: the planning, the interpreting, the adjusting, in a way that felt personalized to my goals and running history. Here’s exactly how I used it.

Why I used AI instead of a running app-generated plan

Running apps had served me well. Their plans built a solid base for my first races. But this time I had a specific goal, reflections on previous plans and races, and a recent IT band injury. A template can’t take all of those things into account. Some workout choices felt random; others felt too aggressive.

What I wanted was a plan built from my actual training history: one that could explain its reasoning, and change when I pushed back. So I built it with Claude.

1. A plan built from my own history

It didn’t start from scratch. I exported a summer of Garmin Activities as a .csv and uploaded it to Claude. In went screenshots of my old training plans from Google Sheets, my past race results, and an app-generated plan, annotated with my notes.

That gave Claude a real baseline for my fitness. Together we reshaped the plan around my goal and my gaps: more threshold work, fewer short intervals, and some workouts built into my long runs.

2. Feedback on every workout

After each run, I sent screenshots from my Garmin app and a note on how the workout felt. Back came a read on whether I’d hit the session’s purpose, and what to adjust next. It also gave me heat-adjusted pace ranges for every workout, which was critical for a summer block.

3. Checkpoints instead of vibes

We built time trials into the plan, with clear “if this, then that” rules set in advance. The results of those time trials moved my race paces by rule, not by feel.

4. Making sense of my Garmin

Resting heart rate, heart rate zones, lactate threshold, the race predictor: my watch has a lot of data. Claude helped me understand what each metric means, how much to trust it, and what to do when one looks off, like a resting heart rate creeping up before illness strikes.

5. Adjustments for when I was sick

When I got sick a week before the race, Claude helped me scale back, or skip, my final workouts day by day so I could make it to the start line healthy.

6. Carb-loading math + grocery lists

Carb loading sounds simple, but I’d never quite hit my targets in the past. Claude worked out the numbers specific to me, built a two-day plan around foods I actually like, and kept a running tally as I ate, suggesting what else to eat to close the gap. It’s all stored in Claude’s memory now, so I won’t start from scratch next race.

7. Race day pace plan

In the past, I never really knew what my race pace should be. I’d estimate it from how my long runs were trending, and run it by feel, starting conservatively for the first few miles.

This time, Claude built the plan from my actual training: a pace band for each phase of the race, and heart-rate ranges that told me when to back off. The plan pointed to about 1:58. I ran 1:58:20.

8. The post-race retro

The Product Manager in me needed a retro while the race was still fresh in my mind. I sent Claude my splits, heart rate, fueling, and how I felt throughout the race to determine what went well, and what to change next time.

9. Planning what’s next

No rest for the wicked: the next day, I turned those lessons into next year’s plan. Claude helped structure my goals, a training calendar, and a race schedule, all in one sitting.

I’m running NYRR’s 9+1 program next year (nine races plus a volunteer shift earns a guaranteed entry to the 2028 NYC Marathon), so Claude is helping me prioritize the races that fit my training goals and reshape my weekly plans around them.

Steal My System

I keep one “Running” project in Claude with a living “training context” doc: my races, goal, injury history, heart rate zones, and the fueling that works for me. Claude keeps this up to date, so each conversation starts already knowing me.

Prompts to start with:

1. “Here’s my training history [attach a Garmin CSV export if you have one], my past training plans, and my race results. I’m training for [race] on [date], goal time [X]. I’m also attaching a plan I’m considering, with my notes on it. Build me a plan around my goal, with checkpoints that tell me whether the goal is realistic. Explain your reasoning, and flag anything you’d push back on.”

2. “[Attach Garmin screenshots: total time, average pace, average heart rate, plus the pace, heart rate, and elevation charts.] This run was supposed to be [threshold / easy / long]. Did I execute what it was meant to do? How it felt: [a sentence or two]. Should anything change for the rest of this week, and if so, what, and why?”

3. “I race on Sunday. I weigh [X]. How many grams of carbs per day should I target and how did you get that number? How many days should I carb load? Build me a carb load plan around foods I actually like [list a few], all familiar and low-fiber. Please also build me a grocery list. Then keep a running tally as I tell you what I eat, and adjust what’s left based on the hours remaining in my day.”

4. “Build my pacing plan for [race] on [date]. Base it on this training cycle: the plan, my checkpoint results, and where my fitness actually landed, not on my goal time alone. Break the race into phases, with a pace band and a heart-rate range for each phase, accounting for the course elevation. Give me clear rules for when to back off and when I can press.”

5. “Here are my lap splits and heart rates from the race, plus my fueling [what and when] and how I felt at each stage [notes, including any issues]. Compare it against our race plan: what went right, what didn’t, and why? Then turn it into specific changes for my next training cycle.”

What AI didn’t do

1. I never use AI to replace a doctor or physical therapist. When something doesn’t feel right, I go get it checked out.

2. It didn’t overrule me. I pushed the plan through 12 iterations to accommodate certain aspects I wanted (a 3-week taper, 11 miles as my longest run, etc.)

A quick disclaimer: I’m not a coach, a doctor, or a dietitian. This is what worked for my body, my injury history, and my goal. I still made tweaks to the plan based on what I was comfortable with, and pressed the AI to share its research frequently when I wanted evidence of its decisions.

What are dots?

A dot is an always-on AI agent that lives inside ChatGPT that you can delegate work to. Each dot gets its own computer in the cloud, so it can browse the web and use tools on its own, and it plugs into 4,000+ apps through ChatGPT’s existing connections. You can reach it via ChatGPT, Slack, or Microsoft Teams, including voice calls. OpenAI announced that texting is on the roadmap as well. At launch, you get one dot, with “teams of dots” on the roadmap. You can read more on OpenAI’s blog post announcing dots.

Dots are rolling out gradually to Pro ($100/month and up) and Business Premium users. I updated my app, but my dots haven’t arrived yet. When I get access, I’ll throw my current Muse use cases at it and report back.

Space: The shared workspace

OpenAI also announced Space: “a dedicated space where teammates, ChatGPT, and your dot can build on shared knowledge.” This appears to look like a Notion-style wiki where colleagues and their agents can work on the same docs. Instead of tagging team members in comments with action items, you can tag your dot instead who can immediately start working on the task. It looks like slides and spreadsheets are on the roadmap as well. You can read more on OpenAI’s new webpage detailing Space.

Dots vs. Muse

Both products give the agent a computer of its own. That lets it use a browser and keep working even if you close the app on your phone or computer. They both have similar, cute character avatars you can name.

The core marketing and use cases are where they seem to diverge. Muse marketing is pretty heavily focused on personal and consumer use, which makes total sense given the parent company. OpenAI did include some personal use cases in its launch video, but overall it was much more weighted towards work and productivity use cases, especially considering the launch of Space.

The work use cases OpenAI shared look genuinely exciting. Instead of having to leave comments in a Google Doc for myself with action items to return to later, I can tag my dot in a Space page and ask it to actually take on that task. Or, I can tag my dot in a Slack thread to quickly hand off a task mid-conversation. I can definitely imagine this fitting seamlessly into my daily workflows.

Everyone’s gotta pay the bills!

Another distinction between dots and Muse is the pricing. Always-on agents have a real infrastructure bill. What Meta is currently offering, for free, is a massive infrastructure spend - but they also have a massive Ads business to help subsidize this for their user base. Muse does have paid tiers of $20 and $100 per month if you exceed the free usage limits, but it seems so far like the free tier is really generous with its limits.

Meanwhile, OpenAI dots come at a starting price point of $100 per month - and it’s unclear what the usage limits of dots will be. OpenAI says more info on usage limits will be coming after the first month.

It’s especially interesting to consider Meta’s announcement this week about moving into the enterprise space, along with the launch of Muse for Small Business. How, if at all, could that impact whether Muse for regular users stays mostly free in the long run? Time will tell.

Where to start?

If you’re not currently on the OpenAI Pro plans, I’d suggest waiting until dots fully roll out to Pro users and the reviews roll in to gauge if it’s something you want to invest in trying. In the meantime, maybe we’ll get some signal about future potential roll-outs to the other pricing tiers. Trying Muse for free is still a great way to experiment with agent use cases.

Another tip: Voice mode was used heavily in the OpenAI demo today. If you haven’t tried voice mode for your AI tools, whether that be ChatGPT, Claude, Muse, etc., definitely give it a try. It’s a huge productivity gain not having to type out requests, especially long ones where I’m brainstorming and want to give the AI a lot of context before it starts a task.

What is Meta Muse?

Muse is Meta’s personal AI agent. Meta describes it as “a secure, private personal AI agent that proactively helps with people’s goals and suggests ideas.” You can talk to it in the Muse app, over WhatsApp, or on the web at muse.ai.

Meta really nailed it with the launch of this guy. The user experience is delightful, and it’s really accessible in the sense that you don’t have to think about how or where to do something. You can simply describe what you need in the main chat and let Muse figure out which tools or context belong to the job.

A few tasks I gave it to start with:

1. Booking appointments: I needed to book a massage before an upcoming race. The user experience on that website is pretty cumbersome if you want to check availability across multiple days, so I handed the task over to Muse. He dug through weeks’ worth of options and came back with the best availability based on my schedule. Since the booking required a credit card deposit, he handed me a link to finalize the appointment myself.

2. Replenishing my hair products: I can order the products I use through a number of retailers, but I didn’t want to sift through who had the best price, who would ship the fastest (because I had, naturally, waited until my bottles were almost empty) and who offered the most cash back on Rakuten. Muse compared the options, found me a 10% Rakuten cash back offer, and sent me the link to check out. I’ll definitely reuse this workflow for any online shopping where I’m looking for a specific item.

3. Running my Return Tracker: The tracker reads my order emails, keeps a running record of return deadlines, and nudges me before a window closes. It saves me from having to remember which retailer gives me 14 days versus 30, or which clock started at shipping versus delivery. It also helps start the return for me and reminds me to drop off the package. I shared the full story of how I built it in my previous post, but it has become one of the clearest examples of what I actually want a personal agent to do.

The most useful proactivity I’ve seen so far: On September 19, I asked Muse to do a one-time scan of my inbox for subscriptions. That was permission for one pass, not a standing permission to keep reading my email. The scan surfaced that a free trial of one of my apps would renew on September 27, and Muse held onto that deadline. Then, on September 26, he warned me that the trial was about to turn into a $30 charge, without my having to remember to ask about it again.

Why this matters to me: the proactive part wasn’t knowing about the trial, it was deciding when it was worth bringing back up. He didn’t go rummaging through my inbox looking for new chores. He took something from a job I’d already given him and reminded me about it at the right moment. That’s the kind of proactive I want: attentive enough to catch what I’d forget, restrained enough not to invent errands or create unnecessarily noisy notifications.

Where it got blocked (for now)

Flight monitoring was one limitation for me. Muse can help research publicly available cash fares, but comparing mileage prices requires signed-in airline access, and United and American Airlines both blocked my Muse bot from signing in to my account, even when I took control of the Muse browser to handle the log in myself. I’m hopeful that more companies and retailers work to enable use cases like these for AI agents in the future!

On the plus side, Shopify announced a partnership with Muse! Muse will be able to shop on any Shopify merchant website and check out if you use Shop Pay.

Security seems thoughtfully designed

Meta seems to have put a lot of thought into the security side of Muse, which matters when you’re handing an AI access to systems like your email and calendar.

The way I think about Meta’s approach to Muse security: my Muse doesn’t work at some big shared desk. He has his own private office in the cloud with his own little computer and a door that locks. Everything he knows about me lives in that office, separate from everyone else’s Muse bots, and separate from Meta’s ad business. You can read more about the safety and security details in this Meta Help Center article and Meta’s overview on how they thought about safety

A few things I’ve noticed in practice:

1. Muse can’t see my passwords. Logging into things happens in a secure vault he has no visibility into.

2. Connecting my Gmail still required Muse to ask for my authorization on each new type of email task. For example, I had to authorize permission to search for specific types of emails related to online orders and returns for my Return Tracker. If I wanted to create a task to monitor subscription renewals, Muse would ask for my approval to search for those specific emails for that task. He doesn’t just go rifling through all my emails.

3. I’ve read that Muse won’t take other sensitive actions like making a purchase or sending an email without asking my approval first. I haven’t tested that use case yet.

Why I migrated my Return Tracker

I’ve migrated my Return Tracker to Muse for three main reasons:

1. My Scheduled Task would silently fail: This happened a couple of times when the task got blocked trying to start a return for me. ChatGPT tried to hand the task back to me, I forgot to respond, and that unfinished handoff ended up holding up the daily check for new return deadlines. I didn’t realize the task was actually paused for several days.

   1. The way Muse set the tracker up by default was cleaner: the daily email scan for potential returns runs independently from any active return it is trying to start. If one retailer blocks the return process and Muse needs me to step in, that one return can wait for my action without stopping the monitor for everything else.

2. I personally found the user experience to be more intuitive in Muse: If I need to edit the task, change a rule, or ask about the status of an order, I can just drop the question into the main chat (or even any side chat) of Muse. Because Muse shares context and memory across chats, it can help without requiring me to find the exact conversation where the tracker was originally created.

3. The task structure seems more organized: When I went to edit the task instructions manually in ChatGPT, I noticed it was stuffing rules about individual past returns into the task instructions. This isn’t terrible in the near term, but over time the instructions will become increasingly bloated if the system has to reread a ton of notes about old order statuses every day it runs.

   1. Muse set up the durable rules for how the tracker should work separately from the running record of specific orders. The instructions explain the job; the order ledger holds what happened. It is a small architectural difference that makes the system less error-prone over time.

   2. To be fair, I may have been able to prompt the ChatGPT task into this type of structure. But that further highlights the point about Muse’s accessibility: With Muse, I didn’t have to specify the task architecture. I was able to give it a simple goal, and it separated the quiet daily scan for new return deadlines from the active return attempts by default.

One thing worth noting: While I’ve had better luck so far with Muse being able to successfully start returns on retailer portals, that may simply be luck. As I’ve mentioned earlier, many retailer sites are structured to block bot traffic. I’ve found that it’s pretty hit or miss as to whether AI can successfully navigate a retailer’s return portal on my behalf. I expect this to improve over time as more retailers build support for this type of workflow.

Where Muse fits in my stack

Since migrating my Return Tracker, I’ve been thinking more intentionally about where Muse fits into my life relative to other AI tools.

The key distinction I’ve made is to keep Muse dedicated to personal tasks. Anything related to work still stays in Claude or ChatGPT.

So far, I like that Muse has context and memory across my chats. Claude and ChatGPT offer Projects that create clearer boundaries around a specific body of work. Those project silos are extremely helpful for work tasks, where I may want one set of information and instructions kept neatly separate from another. But for a personal agent, I haven’t encountered the same need for strict separation.

My running training plan, travel preferences, calendar, shopping habits, and return deadlines all belong to the same life. I may organize them into different side chats so I can find things easily, but I don’t want to reintroduce myself or rebuild the relevant context every time I move between them. Muse’s shared memory makes those chats feel like different rooms in the same house rather than completely separate houses.

There are also tasks I’m still relying on ChatGPT and Claude for where built-up historical context really matters. For example, they have months’ worth of training data from my Garmin that gives them far more historical context about my running. I’d have to proactively migrate that data and context over to Muse in order to get the same level of analysis about my running progress.

What I’m evaluating next

I’m excited to see the Agent space evolve further. Full disclosure, I haven’t tested out Grok Bot yet. I wouldn’t be surprised if OpenAI or Anthropic launch their own agent competitor to Muse at some point soon. For now, Muse has earned a real place in my personal stack.

I’ll be trying out new agents as they arrive and look forward to seeing how each one differentiates itself!

Unfortunately, I have missed one return deadline too many. So I built a ChatGPT return tracker. A Scheduled Task scans my inbox for online purchases, flags upcoming return deadlines and sends reminders to my phone. When I decide to send something back, a reusable Start Return skill finds the order, opens the retailer’s return portal and retrieves the QR code or shipping label. This is a quick, real-life example of the kind of Scheduled Task I covered in How to Automate Tasks with ChatGPT or Claude: A Beginner’s Guide.

How my ChatGPT return tracker works

The “Return Monitor” task scans my inbox every morning for new online purchases and relevant order updates. I have it look for order confirmations, shipping notices, delivery emails, and return confirmations. If a retailer starts its return window on the delivery date rather than the purchase date, the delivery email helps it calculate the correct deadline.

I have also configured it to:

• Exclude retailers that I would rarely make returns to including Amazon, Target, grocery stores, etc. I can always refine these filters later if my behavior changes.

• Track the retailer, order number, item, delivery date and return deadline.

• Distinguish between a deadline stated directly in an email and one inferred from a retailer’s published return policy on their website.

• Flag anything ambiguous for my review, such as final sale merchandise or special holiday return windows, rather than guessing.

• Remind me 7 days and 2 days before the return window closes.

• Stop tracking an item as soon as I tell it I’m keeping it.

No spreadsheets. No starred emails. No to-do list. My system finds the purchases and surfaces the decision to me at the moment I actually need to make it.

The most important reminder happens after I start the return

Starting a return and completing it are, unfortunately, two different tasks. Especially those of us who live in NYC and must haul packages to the UPS store on foot. I unfortunately know from experience that I’m capable of requesting a return, generating the QR code, taping the box shut and still leaving it sit in my apartment collecting dust, saying I’ll “return it tomorrow” until the drop-off deadline passes (no judgment please).

This tracker does not consider the job finished once the label exists. It continues reminding me of the drop-off deadline (which is sometimes something like seven days after the QR code was generated) until I confirm I actually dropped it off.

This was the task that finally convinced me to ChatGPT notifications on my phone. A drop-off reminder is only useful if I see it before I leave the house, not the next time I open my computer.

I added a “Start Return” ChatGPT Skill to finish the job

My Scheduled Task itself handles the memory, monitoring, and timing. A custom skill handles the return request itself.

Skills are reusable workflows that teach ChatGPT or Claude how to complete a particular job. You can invoke the skill at any time in a chat. They can include step-by-step instructions, preferences, decision rules, and conditions for when the AI should stop and ask for clarification.

Scheduled tasks are useful for work that should happen repeatedly or at a particular time. Skills are more on-demand. The benefit is, instead of explaining the same process to ChatGPT every time, I can invoke a workflow I have already taught it.

In this case, when my tracker reminds me that a return deadline is approaching, I can respond:

Start Return for the navy dress.

The Start Return skill then:

1. Finds the right order confirmation, shipping, or delivery email.

2. Confirms which item I meant if it cannot confidently identify it within an order containing multiple products.

3. Opens the retailer’s official return portal.

4. Works through the return process.

5. Applies my saved preferences.

6. Retrieves the final QR code or shipping label (and in the future, will send it to my printer!)

If it encounters a choice it cannot safely make from my existing preferences, it will pause and ask for my clarification and approval.

I preconfigured my return preferences

In teaching the system how I usually make return decisions, I can make the process more autonomous rather than asking me a bunch of follow-up questions. I saved the following preferences and instructions:

1. My preferred return methods are: Happy Returns > QR code that I can bring to a drop-off location > Printed shipping label.

2. My preferred refund method is always original payment method instead of store credit (and if original payment method isn’t available, the AI should pause and ask me what to do next).

3. Tell me when returning something in person could avoid a shipping label fee.

4. Choose a return reason related to “fit” of the item. If that’s not available, ask me how to proceed.

5. Pause and ask me for clarification if my preferred options are not available.

These instructions are what actually save me time in the Start Return process. If I had to answer every question for the AI, I might as well have completed the return form myself.

The first attempt wasn’t perfect

The first time I used Start Return, it missed one of the available return methods. It told me Happy Returns wasn’t available and that my only option was a UPS shipping label I would need to print. I could see Happy Returns selected as the default option on my screen, so I sent ChatGPT a screenshot and asked it to try again. The second attempt worked, and it completed the return correctly. Within minutes, I had a Happy Returns QR code in my inbox.

Afterward, I asked ChatGPT to update the skill based on what had gone wrong. It summarized its proposed changes and waited for my approval. When I forgot to approve the changes, it even sent me a push notification the next morning saying the task needed my attention.

That small repair loop of showing it what it missed, asking it to retry, and then folding the lesson back into its instructions, is what got the skill back on track and will make it more durable in the future.

How does ChatGPT click through an online return portal?

ChatGPT and Claude can use a built-in browser to open web pages, click buttons, enter information, inspect what is on the screen, and move through a website.

How can it do this? Computer Use.

Computer Use allows ChatGPT and Claude to see and operate interfaces like web browsers and desktop apps that don’t have a direct integration.

A good way to think about these capabilities is:

• Skills are playbooks

• Computer Use provides the hands

• Scheduled Tasks provides the clock

I don’t own a printer, but I’m planning on getting one to test a final step in this skill that triggers the label printing.

Why ChatGPT cannot complete every online return

Be aware that AI will likely not be able to handle every return. Unfortunately, the world of internet browsers has not fully caught up to AI surfing the web. For years, websites have been built to explicitly block bot traffic. So, sometimes it gets suspicious of AI computer usage as being a bot. Retailer websites may present CAPTCHA challenges, two-factor authentication, email verification links, or other bot detection systems.

My built-in rules for this skill direct the AI to pause and ask me for clarification if it encounters challenges successfully navigating the retailer’s website.

A quick privacy PSA

As AI becomes more connected to our inboxes, files, browsers, and other personal systems, it’s worth spending a few seconds reviewing your data settings in your AI provider.

I personally keep model training turned off on my consumer AI accounts.

For ChatGPT:

Profile → Settings → Data Controls → turn off “Improve the model for everyone.”

OpenAI says new conversations will remain in your chat history but will not be used to train its models after this setting is disabled. You can review the current instructions in OpenAI’s Data Controls FAQ.

For Claude:

Name → Settings → Privacy → turn off “Help Improve Claude.”

Anthropic’s current instructions are available in its Model Improvement privacy settings guide.

Turning off model training is not the same thing as preventing the service from processing or retaining any data required to complete your request. I also recommend reviewing which apps and accounts you have connected and giving each AI system only the access it actually needs.

What this system gives me

Now instead of relying on myself to remember which packages need to go back and when each deadline is, I have a system that keeps track of it for me and helps me take the next step. My alerts even calculate the money at stake for the return. That is the kind of AI system I want to keep building: not something that replaces my judgment, but removes the repetitive tasks that quietly consume my attention and that I’m most likely to defer.

I keep hearing some version of the same thing from my friends and family: “I already use ChatGPT and Claude, and I want to automate more of my work, I just don’t know where to start.”

If that sounds familiar, this guide will walk you through building your first scheduled workflow.

AI workflows can use several different capabilities, depending on what the work requires:

1. Answering: Responding to your questions in ChatGPT or Claude.

2. Working with files: Reading, analyzing, and creating documents, spreadsheets, and other files.

3. Retrieving information: Gather information from the web or other connected sources like email and Google Drive.

4. Taking actions: Creating or changing something in another system.

5. Running automatically: Starting work on a schedule or trigger, completing the workflow independently or pausing for your approval.

These are building blocks, not sequential steps. This guide focuses on how to identify the capabilities a workflow needs and set it to run on a schedule.

Step 1: Map your process

Before deciding how to automate something, map what happens today. Think through each step in detail:

1. What is the trigger for you to do this task? Is it time-based, such as every Friday afternoon, or event-based, such as receiving an email or uploading a document?

2. Where does the work happen? Which systems does it involve?

3. What are the inputs? What information do you need, and where does it live? Be specific: your inbox, a connected app, Google Drive, or a folder on your computer.

4. What are you doing with that information? What do you review, compare, categorize, decide, create, or change?

5. What are the outputs? What does the process produce, and where should that result appear or be sent?

6. Identify exceptions: What changes when information is missing, duplicated, ambiguous, or unusual?

7. Identify reviews: Where is your judgment required? What should AI prepare for you to review, vs. what is safe for it to complete without your approval?

This may seem like overthinking it, but these details determine which tools and type of automation you need. They also help you identify:

1. Where you spend the most time in the process.

2. Which steps depend most on your judgment.

3. Which repetitive, low-risk steps are the best places to start automating.

Here’s an example of a process map for someone handling Customer Support issues at a software company:

You don’t need to automate your entire process at once. Start with one tedious or time-consuming step that can produce value for you without creating much risk. Think about the tasks you find yourself doing repeatedly, either at work or in your personal life. If a task has a consistent set of steps, it may be a good candidate for automation. If you need inspiration, here are a few ideas:

1. Create a return-deadline tracker that searches order and return-confirmation emails and flags purchases approaching their return deadlines

2. Create a subscription monitor that finds upcoming renewals, trial expirations, annual charges, and announced price increases in email

3. Review your inbox and flag messages that appear to need a response

4. Turn meeting notes into tasks in a project management tool with owners and deadlines

5. Search project notes and messages for decisions that are still pending, along with owners and relevant context

Build your automation

Turn your process map into task instructions

Your automation needs instructions. Refer to your process map and describe what should happen in one complete run. At a minimum, specify the following:

1. The source(s): Which connected apps, accounts, files, or websites should the AI use?

2. The inputs: What information should the AI retrieve, and from what time period?

3. The work: What should it analyze, compare, categorize, create, or change?

4. The output: What should the result contain and where should it appear?

5. The exceptions: What should it do if information is missing, ambiguous, duplicated, or unchanged?

6. The review points: Which actions can it complete automatically, and which should pause for your approval?

You will add the task trigger later. For now, focus on describing what should happen each time the task runs.

Connect the systems your workflow needs

The AI needs access to the places where your information lives.

In ChatGPT, connected apps and reusable workflows are commonly installed through Plugins. Think of these like a toolbox for the app you’re connecting. For example, the Gmail plugin gives ChatGPT access to Gmail and may include reusable instructions for tools such as searching your inbox or drafting replies.

Open Plugins from the left navigation, find the application you need, click to install the plugin, and connect your account when prompted.

If you want to understand exactly what each plugin adds, you can open the plugin’s detail page and scroll to Skills. You can select any skill for more detail about the workflows it supports.

In Claude, the most direct way to connect an external application is through a Connector. Connectors give Claude tools for accessing and, when supported and authorized, taking actions in services such as Gmail, Google Drive, Slack, and Google Calendar.

Claude also offers Plugins, which can bundle connectors with skills and other workflow capabilities, but for the straightforward setup covered in this guide, start with Connectors.

Open Customize from the left navigation and select Connectors. Use the Add button to browse the Connector Directory, then Browse connectors to search for the application your workflow needs.

Open a connector to review its use cases and capabilities to see if it supports the information and actions your task requires. When you want to proceed, select Connect and follow the prompts to authorize your account.

Check what the connection can actually do

Just because ChatGPT or Claude can access a system doesn’t mean it can perform every action within it. Before building your workflow, confirm that the connection supports the specific information and actions your task requires.

If the application or capability you need isn’t available through a plugin or connector, the workflow may still be possible through a custom integration or automation platform. That is a more advanced setup I’ll address in a future article.

Determine whether the task needs local files or apps

You can manually upload an individual file through the ChatGPT or Claude apps in your web browser. However, if your task needs direct access to a folder, file, or application that only exists on your computer (and thus is not cloud based), then it must run locally through the ChatGPT or Claude Desktop apps. You will also need to grant them access to the relevant folder or application.

For a scheduled task that needs access to these local files or applications, your computer must also be powered on and awake, the ChatGPT or Claude Desktop app must be running, and the required files or applications must be available.

Is the workflow scheduled or event-triggered?

Look back at your process map. What triggers you to do this workflow? Does it begin at a predictable time, such as every Friday afternoon? If so, it can become a Scheduled Task directly in ChatGPT or Claude.

If your workflow should respond when something changes, such as an email arriving, a file being added to Google Drive, or a task status changing, a connected plugin may allow ChatGPT or Claude to check for that change on a schedule. However, the plugin does not necessarily notify the AI the moment the event occurs. If you need the workflow to begin immediately and the platform does not support that specific trigger, then you may need to leverage additional task-automation tools such as Zapier, n8n, or Make. This guide focuses on scheduled workflows, so I’ll cover those tools and event-driven workflows in a future article.

Test your workflow before creating the scheduled task

Before creating the scheduled task, run the workflow once in a ChatGPT or ChatGPT Work conversation, or in a Claude Cowork session. Provide the complete instructions and ask it to do one test run.

Confirm that it:

1. Accesses the correct sources and finds the intended information.

2. Performs the work and handles exceptions correctly.

3. Produces the output in a useful format.

4. Stops for your review where appropriate.

For workflows that send messages, modify records, or delete information, complete the first test in read-only or draft mode.

Revise your instructions as needed until the task completes as you would expect. Once it’s working, you can move on to scheduling the automation.

Schedule your automation

How to create a scheduled task in ChatGPT

1. You can create a scheduled task in ChatGPT in two ways:

   1. From an existing Chat or ChatGPT Work conversation: The AI will already have context if you’ve been discussing the workflow in a conversation. You can ask it to return to this same thread and share its output there each time it runs the task. The downside is the conversation can become very long over time, making previous results harder to review. You can read more about scheduled tasks inside a chat here.

   2. The Scheduled menu item in the left navigation: This creates a standalone task using self-contained instructions. Each run’s results will appear under Scheduled.

How to create a scheduled task in Claude

1. Select Scheduled in the left menu

2. Click New task

3. Choose Create with Claude if you want help translating your request into a task, or Set up manually if you already have your instructions written out.

4. If you choose Set up manually, enter the task name, instructions, choose an approval mode, select the project, default model, and schedule. You can read more about creating Claude scheduled tasks here.

Refine the task after its first few runs

Even if the workflow performs well during testing, expect to refine it after a few runs. This is especially important when the task involves subjective decisions such as determining what is important, relevant, urgent, or worth including.

Make sure to give specific examples in your feedback such as:

• If it leaves out useful information, define what it should prioritize, and why something was or wasn’t useful.

• If it retrieves too much information, add clearer filters or a threshold for inclusion.

• If the output isn’t organized the way you’d like, provide examples of a specific structure or template.

• If it makes too many assumptions, specify when it should flag uncertainty or request your review.

For instance, asking AI to “find the most important updates from the newsletters in my inbox” leaves significant room for interpretation. You may need to define examples of what is “important” to you and what should be excluded.

Scheduled tasks are only the beginning

One of the first systems I’m building this way is a recurring briefing from newsletters and other sources I already follow. I want it to do more than summarize. I want it to connect ideas across sources, surface counterpoints, and turn what I’m reading into an evolving personal knowledge base that helps me go deeper into the topics I want to learn more about.

I’ll break down how I’m designing that system in a future post, along with more practical guides to AI skills and event-triggered workflows.

In the meantime, I’d love to know: What’s a process you repeat at work or in life that you suspect should be easier by now? What have you tried to automate that succeeded or didn’t quite work the way you expected?

I felt like I’d accomplished something meaningful, until I looked at the clock at 11:30 p.m., a few nights before my trip, and sunk into my couch realizing I still didn’t actually have an itinerary. In my defense, I had hastily made one dinner reservation at Glug, naturally at an hour my friend warned me only American tourists would be eating.

To be fair, I enjoy the research. It gets me excited about a trip and helps me get familiar with a city’s vibe. But I also do it because I don’t want ChatGPT to generate the same itinerary every other tourist received, where we file in a single line through the same churches and eat at the same packed tapas bars in the Times Square of every city.

Some of that is unavoidable. I fully expect to see the major sites. But I also want something that feels more “me.” I want to see local neighborhoods, find quieter spots off the beaten path. So, I sift through TikTok and other social feeds, even though the same bakery appearing on my feed twelve times may be a sign that I’m not quite as far off the beaten path as I think.

I stared at my Google Maps chaos. I had hundreds of pins in Barcelona, with no practical way to distinguish the museums from the coffee shops, croissants, dinner spots, and everything else. I couldn’t easily tell which pins were on my “must-see” list. Then of course I had to consider the factor of proximity to El Born, the neighborhood of my Airbnb.

I was sure AI could help me distill this, but first I had to get the data out of Google Maps. Pasting in a Google Maps URL and connecting my Google account hadn’t given the AI access to my private saved places. Claude pointed me toward Google Takeout, Google’s free data-export tool.

Exporting all 3,500 saved places gave the AI two different kinds of context. My Barcelona pins became the pool of places it could organize into this trip, while my favorite places from other cities helped it understand what I tend to like: walkable neighborhoods, excellent coffee, local bookstores, and places with a little eccentricity. Instead of asking AI to invent a Barcelona itinerary from scratch, I was giving it years of evidence about my taste.

Once I uploaded the files, it took about 30 minutes to group my Barcelona saves geographically, flag logistically awkward attractions, prioritize my remaining reservations, and build a five-day itinerary around my existing plans. Just as importantly, I had a reusable travel taste profile that could make future trips easier to plan.

How to export your Google Maps saved places

Follow the step-by-step below:

1. Go to https://takeout.google.com/

2. Click “Deselect All”

3. Scroll down and select “Saved”. This includes your Google Maps lists, such as “Want to go,” “Favorites",” and any custom lists you’ve created. One exception: If you use Google Maps’ starred places instead of saving them to lists, also select“Maps (your places)” since those starred places live in a separate file and won’t show up in the “Saved” list export.

4. Scroll to the bottom of the list and click “Next Step.”

5. On the next screen, leave the default settings: An email download link, a one-time export, and a .zip file. Click “Create Export.”

6. Google says the export can take anywhere from a few minutes to a few days, but mine arrived within several minutes.

7. Check your inbox after a few minutes for the .zip file.

How to upload Google Maps saved places to ChatGPT or Claude

While you’re waiting for the export to arrive, create a dedicated Project in ChatGPT or Claude. A Project keeps your source files, relevant instructions, and planning conversations together.

Whenever you want to plan a new trip, you can just start a fresh chat in the same project instead of searching for an old chat conversation or re-explaining your basic preferences. In addition to the time savings, Projects are useful for a couple reasons:

• With your Maps data added as Project sources, ChatGPT or Claude can use those files as shared context in each new chat you start within that Project.

• This means you can start a separate chat for each trip instead of keeping one conversation going indefinitely. Your shared travel data remains available, while the dates, reservations, and decisions specific to each trip stay contained in a more focused thread.

Once the email arrives:

1. Download and unzip the export.

2. Open the subfolder titled “Saved.” Inside, you should see a separate CSV file for each of your lists, such as “Want to go” or “Favorites.”

3. Select the Maps-related CSV files and upload them to your new Project. In ChatGPT, open Sources, select “Add sources,” and upload the files. In Claude, select the “+” beside Files and choose “Upload from device.”

4. If you also exported “Maps (your places)” to include starred places, locate and upload that JSON file as well.

How to create an AI travel taste profile

If your export contained a “Favorites” list, it’s worth first asking the AI to analyze your preferences and create a “taste profile” to get a sense of how well it will curate its recommendations for you and give it any initial feedback.

I used a prompt along these lines:

Analyze my saved places and create a travel taste profile that can guide this trip and future recommendations. Treat my “Favorites” list as the strongest evidence of places I genuinely like; treat “Want to go” and other aspirational lists as signals of interest, not proof of preference.

Identify patterns in:

• The kinds of neighborhoods I gravitate toward

• Atmosphere, design, and level of formality

• Food, drink, shopping, and cultural interests

• My apparent tolerance for crowds and traditional tourist attractions

Support each conclusion with examples from my saves, distinguish strong patterns from tentative inferences, and flag any contradictory signals. End with a short list of attributes that make a place more likely to match my taste—and the types of places you would be more skeptical of recommending.

Reviewing this profile before itinerary planning gave me a chance to correct anything it had misread. It also showed me whether the AI was identifying meaningful patterns or simply producing a flattering but generic description.

Once the profile felt accurate, I asked it to use those preferences along with my Barcelona saves and known schedule constraints to begin planning the itinerary.

How to turn saved places into a personalized itinerary

You can use a version of the following prompt to generate your itinerary:

Using my uploaded Google Maps saves and travel taste profile, create a first-pass itinerary for [destination].

My trip details:

• Dates, arrival and departure times: [details]

• Where I’m staying: [address or neighborhood]

• Existing reservations or fixed plans: [details]

• Must-sees versus nice-to-haves: [details]

• Preferred pace and number of activities per day: [details]

• Walking tolerance and primary transportation: [details]

• Food, budget, weather, or other constraints: [details]

As you plan:

• Group places geographically and minimize unnecessary backtracking.

• Prioritize my saved places, but research additional options if there are important gaps.

• Check current opening hours, ticket availability, likely crowd patterns, and the weather forecast when relevant.

• Flag anything that needs to be reserved in advance.

• Identify any must-sees that are logistically inefficient or unrealistic to fit.

• Clearly label anything you couldn’t verify.

Before finalizing the itinerary, show me a rough structure, identify any conflicts or tradeoffs, and ask the follow-up questions you need answered.

Claude responded with a rough structure rather than jumping directly to a rigid hour-by-hour itinerary. It flagged places that were geographically awkward, asked whether I cared enough about them to justify the detour, and separated the reservations I needed to make immediately from places I could approach more spontaneously.

It also noticed details I hadn’t explicitly explained, like the lone running store among my saves, and asked whether I wanted to make time for a run during the trip.

I refined the draft through a few short follow-ups: which afternoon I wanted to leave open, which walking tour I preferred, and which sites I was willing to cut. This is where I wanted to spend my planning time, not evaluating every scheduling possibility, but making the tradeoffs that required my judgment.

From uploading the files to having a workable first draft took about 30 minutes. The itinerary still required my input and verification, but it replaced several hours of manually sorting pins, comparing neighborhoods, calculating the time to get between sites, checking booking requirements, and deciding what could realistically fit.

The surprising part was that the “Saved” export didn’t include separate geographic coordinates. It did include place names, addresses, and Google Maps links, which Claude could use, along with live research, to infer which places belonged in the same neighborhood. That was enough to produce a geographically sensible first draft rather than an itinerary that required me to cross the city repeatedly.

Why this works better than asking AI to plan from scratch

A generic AI-generated itinerary is only as personalized as the information you give it. I could spend 20 minutes explaining that I gravitate toward excellent coffee, quieter neighborhoods, independent shops, and restaurants that feel special without feeling overly formal. Or I could give it years of saved places that already demonstrate those preferences.

I still need to provide the constraints specific to this trip: My dates, walking tolerance, existing reservations, and how much unscheduled time I want. But I don’t have to translate years of taste into a collection of adjectives every time I plan somewhere new.

This doesn’t necessarily mean the AI uncovers “better” places than Google Maps, Tripadvisor, or a travel website. The improvement isn’t primarily in the discovery; it’s in the synthesis. It can turn hundreds of possibilities into a geographically coherent plan, identify what needs to be booked, and help me assess the tradeoffs when everything won’t fit.

It also changes where I spend my planning time. I used to sift through a spreadsheet, calculate transit times, determine how much could comfortably fit into a day, and research the best time to visit each attraction. Now I can focus on the decisions that actually require my judgment: whether I care enough about a particular site to cross the city or whether I’d rather prioritize a low-key dinner over squeezing in another museum.

Plus, since my saves and taste profile live in a dedicated travel Project, that context doesn’t disappear or get forgotten after Barcelona. I can refine it with feedback from my trip and use it as a better starting point for the next one.

Can you trust AI-generated travel recommendations?

It depends on what I’m trusting it to do. I trust AI to help organize my saves, identify patterns, surface logistical conflicts, and produce a strong first draft. I don’t treat it as the final authority on information that can change in the real world.

A saved-places export is only a snapshot. Restaurants close, opening hours shift, timed tickets sell out, and reservation policies change. Even when the AI conducts live research, it can misread an outdated page or present an inference with more confidence than it deserves.

Before finalizing the itinerary, I check official booking pages, current opening hours, reservations, ticket availability, and the final routes in Google Maps myself.

How to improve your taste profile over time

The itinerary saved me time once. The taste profile is the part that can become more useful after every trip, so I saved a concise version of it as a reference source in my travel Project.

I’m improving it over time by:

• Sharing brief feedback on what I loved, disliked, or found overrated and why. For example: “This day was theoretically walkable, but four scheduled stops left me exhausted that day.”

• Asking the AI to revise the taste profile based on that feedback, then replacing the old reference with the updated version.

• Uploading a fresh Google Maps export after a significant round of saving or before planning a trip for which I’ve added many new places.

I’d love to tell you I’ve stopped saving coffee shops in cities I’ve never visited at 11:30 p.m., but if anything, I’m saving more of them. The difference is that those saves now feel less like digital hoarding and more like actual inputs into my next trip.

Your past self already did the research. She just needed better tools to make it useful. The active setup is only a few minutes of checkboxes and uploads, while the payoff is finally being able to use years of accumulated saves.

Google Maps was the easy export. I haven’t yet solved what to do with all the Instagram saves, TikTok videos, and camera-roll screenshots I’ve accumulated, but this experiment is one example of the larger idea I’m building toward.

The Smart Girl Stack is where I test how AI can help us create more intelligent systems for everyday life so we can learn faster, work smarter, improve our finances, take better care of ourselves, and spend more time on the things we actually enjoy. Subscribe to see what works and what ultimately earns a place in my stack.

FAQ

Can you export your Google Maps saved lists?

Yes, through Google Takeout, Google’s free data-export tool.

Can ChatGPT access my Google Maps saved places directly?

Not through ChatGPT’s native Google connections. As of July 2026, connecting Google Drive or your Google account does not give ChatGPT access to your private Maps saved lists. A shared Maps URL may expose some public information, but it isn’t a reliable way to import an entire collection. The most reliable method I found is to export the lists through Google Takeout and upload the resulting files to a ChatGPT or Claude project.

Does the Google Maps export include coordinates?

Not in the “Saved” CSV files I exported. They contain place names, addresses, and Google Maps links, but no latitude and longitude fields. AI tools can often resolve those places and determine which neighborhoods they’re in using the names, addresses, and live research.

What’s the difference between the “Saved” export and “Maps (your places)”?

“Saved” exports your lists as CSV files, which is primarily what you want for this workflow. “Maps (your places)” exports a JSON file containing your starred places, which don’t appear in the “Saved” list export. If you tend to star places, export both “Saved” and “Maps (your places)”.

How do I organize hundreds of Google Maps saved places?

After exporting the data, upload the CSV files to a dedicated travel Project and ask the AI to analyze it. The goal isn’t to create a perfectly organized spreadsheet of all your saved places, it’s to make the relevant saves usable for this specific trip.

Do I need to upload a new export before every trip?

That’s up to you. Your existing files will still be there for the future. I’d upload a fresh export if you’ve saved a significant amount of new places for an upcoming trip, so that the AI has the latest.