849 Records. 53 Pages. 4 Days. AI Didn’t Save Me Time.

I started this on Friday.

It’s now Tuesday, 4pm.

All I needed to do was extract 849 records from a 53-page PDF. No analysis. No modelling. Just getting structured data into Excel. It sounded like the perfect use case for AI, so naturally I used it.

My report was about claims so accuracy is of upmost importance.

At the beginning, it actually felt promising. The outputs looked structured, almost usable. It felt like I was just one step away from getting this fully automated. If I could just get the format right, the tool could take over and I wouldn’t have to touch the rest.

So I tried.

I asked it to standardise the format. I asked it to process multiple pages at once. I asked it to combine everything into a single dataset so I wouldn’t have to copy and paste.

Each time, the answer came back with some version of “yes, it can be done.”

And each time, I believed it.

For a moment, it always looked like it was working. The structure was there. The rows looked aligned. It felt like I had finally crossed the line from manual work into automation.

Then I checked more closely.

Something would always be off. A row would split. A column would shift. Plan names would spill into adjacent fields. Numbers would merge with frequencies in ways that made them unusable.

Small errors, but enough to break trust.

So I adjusted. Changed the prompt. Tightened the instructions. Tried again.

And again, it looked like it might work.

Then it didn’t.

This cycle repeated itself five or six times. Each time, I got just close enough to believe the problem was solved. Each time, it fell apart when I tried to actually use the output.

Until eventually, the conclusion became clear, not because it was stated directly, but because I kept running into the same wall.

It couldn’t take over the task reliably.

And the work came back to me.

That was the turning point.

I stopped trying to get it to take over.

I switched to doing it page by page.

For each page, I uploaded the image, took the output, copied it into Excel, checked every row, fixed broken lines, adjusted columns, and made sure the numbers were in the right place.

Then I moved on to the next page.

By the third or fourth page, I already knew what the next one would be like. The pattern was predictable. The errors were familiar. The corrections became routine.

And then suddently AI gave me in a different format of output or it took a long time to process a page and went all quiet. I had to keep asking 'What happened?' in order to get an apology and to 'restart' the enginer again.

And I still had dozens more to go.

So I kept going.

Page after page, repeating the same process, knowing that the output would never be clean enough to use directly, and that I would have to step in every single time.

👉 53 times.

The work never really left me.

It just came back in different forms.

What I thought would be automated became something I had to supervise constantly. Instead of removing effort, it changed the nature of the effort.

From typing, to checking.

From input, to verification.

From doing, to fixing.

And in some ways, that made it more tiring. Because the work was no longer linear. It required constant attention, constant correction, and constant judgement.

If I’m being honest, if I had just done this manually from the start, I would probably have finished in half the time.

That’s the uncomfortable part.

AI did help in small ways. It read the text. It reduced some typing. But it never gave me something I could trust enough to use directly without checking.

So the heavy lifting never disappeared.

It just shifted onto me.

After four days of this, one thing became very clear.

AI today is still excellent at writing. Give it something unstructured, and it will organise it beautifully. It brings clarity to ideas in a way that feels almost effortless.

But when it comes to extracting structured data in bulk, especially when accuracy matters, it still depends heavily on human effort.

And that matters more than we think.

Because this isn’t just about one task.

If we start believing that this kind of productivity gain is enough to reduce manpower, we need to pause and think carefully.

It is easy to remove headcount on a spreadsheet.

It is much harder to bring back the judgement, context, and operational knowledge that experienced people carry. The kind that only shows up when something doesn’t align, when outputs break, or when the process stops behaving as expected.

Because when the tool struggles, it is those same people who step in, fix the process, and keep the work moving.

Looking back, I realised I wasn’t just using AI.

I was trying to get it to take over the work entirely.

But what I was actually using was an assistant.

And assistants don’t replace the work.

They support it.

Final thought

I didn’t save time.

I just spent four days doing the work I thought AI could take over, only to have it handed back to me again and again.