I tried automating my job with AI... and failed. I don't think it's a bad thing
Over the last few months I’ve tried to automate my job with AI... and failed. And I don't think it's a bad thing, to be honest.
I'll explain.
Ever since playing around with LMMs focused on coding — like Claude AI — I've been learning basic things like
interacting with APIs
creating my own Github page
creating artifacts and scripts
or even training a local LLM with my own data
This has helped me immensely with my software sales job. Because of this, I've been on a quest to automate as much as I can of my work.
As an account executive, I have an "email job". Most of my work involves receiving and sending emails. Yes, I know, kids at kindergarten who pretend to have real jobs like doctors or policemen don't pretend to have an email job...
To do this email job I need to process a lot of data:
pricing information, quotes, market conditions
negotiations
value props, competitor comparisons
our API's pros and cons
etc.
So I tried training a local LLM (both with Ollama and GPT Everywhere). Mainly using Llama 3.1 8B. My goal would be to feed it certain variables so it would spit out a nice email with a decent pricing proposal... and it didn't work.
It didn't work because for it to retrieve the exact data I wanted, I had to classify it in such a way that I was essentially doing data entry for a database.
So I might as well create a database instead.
When I wanted to train it to make accurate pricing proposals I would only have to check, it also didn't work because it doesn't understand numbers. It can't read a spreadsheet very well either.
I was essentially trying to build a spreadsheet without using a spreadsheet. So I might as well use a spreadsheet.
When I wanted the local LLM to write an email for me after giving it enough variables, I realized I was writing 90% of the email myself in the prompt. And then spending another 10 mins correcting what it came up with.
And more examples like these.
Sure, you could argue I need a more powerful model. I need more parameters. More GPU. More effort into training and data classification. And a large etcetera.
But that’s not the main point.
My main point is that I realized sometimes we want to try out the new shiny object, instead of exhausting our current tools. I realized I can still get out a lot out of tools I already have and use in my day-to-day:
email templates and tools
spreadsheets (brushed-up pivot tables after a few years)
databases
video tools
CRMs
This isn’t to say AI isn’t useful. I use it in other ways in my day-to-day with obvious performance improvements.
But the time it takes me to do manual work like emails, proposals, and data handling has already been reduced by maxing out the above tools. And I’m sure there’s still an extra layer or two of complexity in them I could explore that could benefit me before I turn to AI.