In the AI Era, A Lot of Software Feel Optional

Recently I wanted to answer a pretty simple question: what did I actually spend in the past few years, and what am I likely to spend in the next few years?

Nothing fancy. I just wanted some numbers.

At first I thought this would send me into the usual pile of tools. A budgeting app, some dashboard, OCR software, PDF extraction, spreadsheet cleanup, maybe some forecasting tool at the end. That is usually how this kind of thing goes.

What I didn't expect was this: by the end, it felt like I didn't need a separate app for every step. AI was helping me build the workflow itself.

I don't use cash much, so in theory my spending history was already there. Credit cards, cheques, online payments, pretty much everything. The problem was, I didn't really have raw data. I had years of statements sitting in PDF files.

And those PDFs were messy. Different banks, different statements, different layouts. Some had text I could extract. Some were basically screenshots saved as PDFs. Some looked readable on screen, but once you tried to pull the text out, it came out as garbage. They were made for people to look at, not for anything else.

So before I could even think about spending patterns, I had to get my own data back out of my own documents.

To be fair, a lot of that part could be done manually. I could sit there and type transaction data out of PDFs one by one if I really had to. Painful, slow, and kind of soul crushing, but doable.

If I'm being honest, I'm kind of lazy by nature. If there's a way to skip manual work, I'm going to skip it.

So I used AI to help build a pipeline that could go through the PDFs, pull out text, run OCR when needed, and recover the transaction details. It wasn't clean at the beginning. Around 20 percent of the statements didn't extract properly on the first pass, and another 20 percent had something wrong in the result. Missing rows, bad text, weird formatting, stuff like that.

So the useful part was not only "AI can read a PDF." The useful part was that it could help me notice when the result was wrong, figure out what was missing, and help fix it.

That part took time. I had to test a small batch, adjust things, run it again, fix more edge cases, then run it again. Pretty normal once you get into real world data. But once it got stable enough, the scale changed fast. I was able to import around 1,000 PDFs in minutes. That was when it clicked for me: what looked like a painful cleanup job was really a workflow problem.

After the data was clean, I assumed I would finally move everything into a finance app. But I didn't. I kept the transaction data in CSV. I kept notes and report drafts in Markdown. I kept some structured intermediate files. And I kept using AI to help me work through the messy parts.

The next hard part was deciding what the numbers even meant. What counts as real spending and what is just money moving between accounts. How to treat income, investment, and refunds. Which categories are actually useful. Which spending spikes matter and which ones don't. How to turn old spending into a future estimate that still matches real life.

That is where AI actually helped me. Not in some magic way. More like having someone that could keep working through the data with me, over and over, without getting tired of the mess.

What didn't feel realistic to do by hand was the other part: checking categories across everything, catching the places where my own labeling was off, and noticing patterns I probably would have missed.

One example was a recurring charge with a code like label that didn't mean much when I first saw it. I had dropped it into a vague category just to move on. AI matched it against clearer versions of the same transaction on other statements, and it turned out to be the same regular housing expense. Fixing that one pattern cleaned up a whole batch of transactions, not just one.

That was the part that felt less like automation and more like having a financial consultant, or at least a second set of eyes.

It helped me clean up double counting, review categories, catch my mistakes, spot patterns, and build a forecast that matched actual life changes instead of just stretching the past forward. Things like a vehicle lease ending. One child finishing school. Another starting university. Retirement changing commuting and travel. Those things matter a lot more than a nice dashboard.

By the end, I had something way more useful than I expected. I had a cleaned transaction history, a much clearer view of how money was actually moving, and a 10 year spending model I could adjust when assumptions changed. The baseline version came out to a number that felt pretty big, and the version where I tighten things up came in meaningfully lower. What mattered to me the most was that I could see how the estimate was built, and I could change it when life changed.

That whole experience changed how I think about software.

For a long time, software were valuable because they packaged a workflow for you. That still matters. I'm not saying software doesn't matter anymore. Some software you still really need. But a lot of narrower software now feel less necessary than before.

If I can recover the data, clean it up, ask questions, test assumptions, write notes, and generate a report in one flexible workflow, then I don't always need a separate app for every single step. Sometimes I just need my files, a good process, and AI in the middle helping with the messy parts.

That is also why I like writing the final estimate report in Markdown. It feels simple. I can read it, change it, and work on it again later. AI can help with it too. If I want HTML or PDF at the end, fine, that is just output. The work itself stays in my own files.

I started with a personal finance question.

I ended up feeling that a lot of software is still useful, but not as necessary as it used to be. When the real work is getting information out of messy places, making sense of it, and turning it into a decision, AI can replace a lot more of the old tool chain than I expected.