I Cut a 7-Hour Legal Task to 30 Minutes. The AI Couldn't, So I Built the Tool.

Marketing fills the pipeline. Delivery is what keeps clients. This week the highest-leverage thing I did for a law firm wasn't a campaign.

This week I watched a paralegal fight Adobe Acrobat for an afternoon and lose.

A mountain of discovery sat on her desk, half-processed. The software meant to handle it kept choking, clogging, and dragging her machine to a crawl. The fix wasn't a smarter prompt or a better AI model. It was a few hundred lines of Python I wrote by the end of the week.

I've spent 20 years working two sides of the same problem. One side is growth: getting a business found, getting the phone to ring, turning attention into clients. The other side is what happens after the phone rings. The actual work. The delivery.

Most marketers stop at the first side. They fill the pipeline and walk away. I've never understood that. If you flood a team with demand they can't handle, you didn't help them. You made their week worse. Growth a team can't absorb isn't growth. It's a liability with a nice dashboard.

So when I embed with a client, I work both sides. Here's how that played out this week.

The actual job

I'm working with a law firm's litigation department. A new lead attorney came in wanting to modernize how the team works. She'd been using Claude on her own, liked it, and got the firm to approve it for the legal team. Good instinct. AI belongs in that workflow.

My role is the one I always play. I spend the time with the people who do the work. I have them walk me through their process step by step. I find the pain points. I look at the tools they already use. I learn exactly where information piles up and where the hours disappear.

In this case, the process was discovery.

When a case begins, a mountain of information lands on the team, and someone has to catalog all of it. For the paralegal and the assistants, that's where the hours go. The hardest part is the document handling. Gathering PDFs. Combining PDFs. Converting things into PDFs. Then going through every page and applying Bates numbering so each document has a reference for the case. That includes .eml files and the .msg format Outlook spits out, usually long email chains that someone has to convert into clean PDFs by hand, one at a time.

There are third-party services that exist only to organize, index, and prep all of this for a case. But this is a smaller firm. They don't pay for those platforms. They've always done it in-house, with a paralegal and an assistant grinding through it by hand.

Times are changing. But not the way the headlines promise.

Where the AI dream breaks

The fantasy is obvious. Hand it all to Claude or some AI, and the finished, Bates-stamped, perfectly combined record just appears. I had the same thought, so I tried the manual AI route first. I hit the wall fast.

A chat workflow and context box isn't built to be a reliable production line for hundreds of files at a time. It won't sit there and merge a 700 page record of text, images, formats, and stamp every page with consistent Bates labels on demand. And the moment you push real volume through, you hit the plan's usage caps.

Then there's the bill. Every model raises prices over time. The more compute you spend, the more it costs, and those costs stack quietly until someone in finance asks what's going on. Corporate America is starting to feel this. The A.I. usage invoices are arriving, and some companies are quietly deciding it was cheaper to keep people doing the work.

Even Adobe Acrobat struggles here. Throw hundreds of documents at it and it bogs down and drags the whole machine to a crawl. I watched that happen in real time.

That's the moment the job becomes clear. The AI tool isn't the answer for this. The answer is a purpose-built program that does exactly one painful thing, reliably, every time.

What I built

So I built it. A lightweight Python desktop app of tools, each aimed at one specific step in discovery and document processing.

The first combines PDFs. The paralegal drags a folder filled with PDFs and the program merges everything, then generates a report of any errors and isolates the files it couldn't combine. A human only looks at the handful of problem files instead of all of them.

The second ingests .eml and .msg files. Drag in the files or the whole folder, and it outputs them as clean, organized PDFs. No more converting email chains one at a time.

The third is a Bates labeler. It does what you'd get from a paid Adobe setup, applying consistent Bates numbering across the combined record, except it's built for this team and this exact workflow.

And… it was about 20x faster! Adobe has become notoriously bloated and crawls for users.

The part that matters

One case discovery document that used to eat 6 to 7 hours of a paralegal or assistants day now takes about 30 minutes.

Same output. Same Bates numbers. Same court-ready PDFs. A fraction of the time.

That's not a tweak. That's the difference between losing most of a day and losing half an hour. And it trickles. Every matter, every assistant, every week from here forward.

Where the AI actually wins

This isn't an anti-AI story. It's a right-tool-for-the-job story, and the other half proves it.

The same week I was writing Python to kill the mechanical grunt work, we connected the firm's protected SharePoint to the protected Claude environment we'd stood up for the legal team. That's where the model earns its keep. Search and reasoning across specific cases got dramatically faster, and all of it stays inside a protected environment, which for a law firm is the part that makes the whole thing usable in the first place.

That's the distinction that matters. Claude is not the tool for merging 700 PDFs and stamping Bates labels. It is the tool for reading across a case file and reasoning about what's actually in it. Ask it to do the first job and you hit caps and bills. Point it at the second and it's a force multiplier.

We're just getting started, and the early gains are already large.

Why this is the seat I work from

Plenty of marketers can get you clients. Plenty of developers can build you a tool. The number of people who can do both in the same week, sitting at the same desk as the person with the problem, is very small. That's the seat I work from.

I could have run a campaign for that firm. I do that work too. But the highest-leverage move this week wasn't a campaign. It was removing the bottleneck that kept a team from taking on more without drowning. Capacity is growth. A team that gets six hours back per matter can serve more clients, respond faster, and actually deliver on whatever demand I send their way.

It also takes someone who knows which job goes to which tool. When the model is the move, and when it quietly backfires. The skill isn't reaching for the fanciest AI. It's knowing where it wins, where it doesn't, and building the small, boring, reliable thing in the gap.

The math nobody runs

This is just the start. Walking the rest of the litigation department's process, I've already spotted the next round of fixes. I don't have every answer yet. Some I'll solve in five minutes. Some will take four hours of building. It doesn't matter, because the pursuit is what compounds.

And the savings aren't only the hours. Five minutes here, four hours there, it adds up faster than anyone expects. But the real return is what a team does with those hours once they're free. Better work on the matters that count. Faster answers for clients. Room to think instead of grind.

Picture a year of this. Wins stacked across an entire department. A small firm that reins in its own information instead of being buried by it punches several weight classes above its size.

Stay tuned. More coming out of the litigation department.