I have experienced the frustration of creating heavy formulated spreadsheets using Perplexity Chat earlier this year (think a complete Project Management Tracker with Status Reporting) for Microsoft 365 Excel.
In your example, I am curious why you would not continue the workflow using Claude native connectors for modifying the Google Sheet that have been available since about Feb 2026
Connectors are part of the Claude Pro Subscription
Very frustrating, because our company is not allowing me to install Claude Co-work unfortunately...but at least i have the code function and chat, plus Excel ad-in which has been great
There’s so many new ways to make work easier with AI. This is so helpful! Also though, how can companies maintain their information safe if their workers are all using these tools and potentially uploading critical or confidential data! Some companies simply have a prompt when you log on to AI tools. Barely enough!
Internal tools that might quickly review and advise employees what can be done with the data? There’s only so much that yearly trainings can do to really teach people about what they can do with data and different information.
Interesting read. I work with businesses that rely heavily on spreadsheets for reporting and operations, and I've found that the best spreadsheets are the ones that require the least explanation. If someone needs a 15-minute walkthrough to understand it, the design probably needs work.
Yes. I use AI for formula generation, data analysis, and troubleshooting spreadsheet issues. It definitely speeds up the work, but I still think good spreadsheet design comes down to making the output clear and understandable for the next person who opens the file.
So I’m quite new to AI but wondering if you have a recco for an agent that can help me create a spreadsheet when I don’t necessarily know what the x and y columns should be. For example, if I give the agent the overarching concept with some specifics, can it recommend names rows and columns for things I should probably be thinking about?
Great stuff. What do you suggest about the omnipresent bloated sheets that have been created years ago and are now a mess of redundant, useless data. What’s the best path to clean up?
Great article but you're comparing apples to oranges with Chat vs. Cowork vs. Projects. Chat and Cowork are full operating modes inside Claude and they both have a Projects concept; Projects is not a separate thing. So I'm not sure why you're showing it as an either/or.
@ruben thanks! Have you tried GPTforWork / GPTforExcel? Similar to shortcut.ai it seems, I found it to be very powerful. Uses any LLM you like, has agentic capabilities natively within excel.
Appreciate the article, however I completely disagree with the title!
Reason being is that I am seeing the opposite.
I work for a Big4 firm in the Financial Modelling team and we have reforecasted our business to move away traditional Excel build over the next few years. With Excel review (something we already offer) becoming our primary offering.
This is because we are seeing a growing abundance of clients coming to us having built complex models to support decisions with multi-million/billion consequences, all using AI. They want confirmation that they can trust these models, before they make decisions upon these self built tools. For us, it's never been a better time to master Excel to become the niche, experienced specialists, who provide assurance on AI generated Excel outputs.
The human-review angle is what makes that niche durable. We looked at Booz Allen's AI-fights-AI cyber suite in https://thesynthesisai.substack.com/p/the-mirror: when your verifier shares the generator's architecture, it inherits the same blind spots, confabulation and cascading errors included. A human modeller auditing an AI-built model doesn't share those failure modes, which is why your team's pivot holds while AI-checking-AI keeps tripping over itself.
Thank you, Ruben. This is incredibly comprehensive.
you're welcome, Karo :)
I have experienced the frustration of creating heavy formulated spreadsheets using Perplexity Chat earlier this year (think a complete Project Management Tracker with Status Reporting) for Microsoft 365 Excel.
In your example, I am curious why you would not continue the workflow using Claude native connectors for modifying the Google Sheet that have been available since about Feb 2026
Connectors are part of the Claude Pro Subscription
https://support.claude.com/en/articles/10166901-use-google-workspace-connectors
cowork builds the spreadsheet from scratch. once it's done i click the google drive button and it opens straight in sheets
from there i switch to the chatgpt extension - i love how simple it is for edits
Interesting....will have to test the UX between ChatGPT and Claude Google Workspace.
What OpenAI model were you using with the ChatGPT Extension to get the modification outcomes you were expecting?
ChatGPT 5.5 :)
Spreadsheets have always been powerful but intimidating for most people. AI is finally making that power accessible to everyone💡
intimidating, yes. and time-consuming
open cowork and stop thinking about either
Very frustrating, because our company is not allowing me to install Claude Co-work unfortunately...but at least i have the code function and chat, plus Excel ad-in which has been great
i hear you :) Cowork would've been perfect for you, honestly the best out of all i've tried
There’s so many new ways to make work easier with AI. This is so helpful! Also though, how can companies maintain their information safe if their workers are all using these tools and potentially uploading critical or confidential data! Some companies simply have a prompt when you log on to AI tools. Barely enough!
login banners are pure security theater
leak is an employee pasting a client contract into a free chatbot at 11pm because nobody gave them a better option
fix the tool then write the rules around it
one approved platform with a clear “this data stays in, that stays out” policy works
Internal tools that might quickly review and advise employees what can be done with the data? There’s only so much that yearly trainings can do to really teach people about what they can do with data and different information.
Interesting read. I work with businesses that rely heavily on spreadsheets for reporting and operations, and I've found that the best spreadsheets are the ones that require the least explanation. If someone needs a 15-minute walkthrough to understand it, the design probably needs work.
awesome - do you use AI tools for your spreadsheets?
Yes. I use AI for formula generation, data analysis, and troubleshooting spreadsheet issues. It definitely speeds up the work, but I still think good spreadsheet design comes down to making the output clear and understandable for the next person who opens the file.
Claude Cowork to start. Google Sheets to live. ChatGPT extension to edit. That's the stack.
never knew Excel could be this easy
I think you should give a try to www.gptforwork.com
Despite the name, it lets you pick a different LLM per each job and works very well.
i'll keep this in mind, thanks for sharing :)
The current state of AI seems to be with each tool doing the thing it's actually good at.
No tool doing everything. At least not yet.
a specialist beats a generalist while there’s no super app yet :)
pick the right tool/s per job and move on
The real value here is knowing when to use Claude to build the thing, when to move into Sheets, and when to let ChatGPT handle the edits.
pretty much :) tools are easy, human judgment is the new work
Thanks Ruben, the best article I have read regarding excel. Copilot sucks so using Claude with it is way better
thank you :) how was the Claude Cowork, Google Sheets, ChatGPT’s extension workflow?
Claude is good but hit limit so fast :(
So I’m quite new to AI but wondering if you have a recco for an agent that can help me create a spreadsheet when I don’t necessarily know what the x and y columns should be. For example, if I give the agent the overarching concept with some specifics, can it recommend names rows and columns for things I should probably be thinking about?
Claude Cowork solves exactly your problem :)
Interesting! Can’t wait to dive in. Thanks for all the resources!
of course - hope it all goes well
Great stuff. What do you suggest about the omnipresent bloated sheets that have been created years ago and are now a mess of redundant, useless data. What’s the best path to clean up?
rebuild from scratch with AI and do not clean the bloated legacy sheet, keep a back up
Great article but you're comparing apples to oranges with Chat vs. Cowork vs. Projects. Chat and Cowork are full operating modes inside Claude and they both have a Projects concept; Projects is not a separate thing. So I'm not sure why you're showing it as an either/or.
that’s fair :) i have to share i made that infographic for someone who's never opened Claude before
those three tabs can feel like three different things - its a guide of where to start, when to use each one
@ruben thanks! Have you tried GPTforWork / GPTforExcel? Similar to shortcut.ai it seems, I found it to be very powerful. Uses any LLM you like, has agentic capabilities natively within excel.
i haven’t tested it yet, its my first time to hear about it :)
any pros and cons you can share?
Appreciate the article, however I completely disagree with the title!
Reason being is that I am seeing the opposite.
I work for a Big4 firm in the Financial Modelling team and we have reforecasted our business to move away traditional Excel build over the next few years. With Excel review (something we already offer) becoming our primary offering.
This is because we are seeing a growing abundance of clients coming to us having built complex models to support decisions with multi-million/billion consequences, all using AI. They want confirmation that they can trust these models, before they make decisions upon these self built tools. For us, it's never been a better time to master Excel to become the niche, experienced specialists, who provide assurance on AI generated Excel outputs.
The human-review angle is what makes that niche durable. We looked at Booz Allen's AI-fights-AI cyber suite in https://thesynthesisai.substack.com/p/the-mirror: when your verifier shares the generator's architecture, it inherits the same blind spots, confabulation and cascading errors included. A human modeller auditing an AI-built model doesn't share those failure modes, which is why your team's pivot holds while AI-checking-AI keeps tripping over itself.
Title feels a bit like bait. 😹