Last night, I got stuck sorting a pile of client files

As a freelance consultant, I often juggle contracts, interview recordings, and sales materials. Last week, in a shared studio in Hangzhou, I stared at dozens of files with garbled names and felt that knot in my stomach: with this much material, manually digging through it one file at a time could drain my energy before I even sent a quote.

I’ve also been stuck down the wrong path of “buy more expensive software.” In the end, I only moved the mess from one folder into another.

Then I saw that NASA had used 48 AI agents to analyze lidar data in parallel and find a previously undocumented Maya pyramid in 72 hours. My first reaction wasn’t “AI wins again.” It was: could a small team like ours split document review across several agents?

Not a chat tool—a way to divide work in parallel

The NASA-related research was carried out by 48 AI agents supported by the Google SWE points program. They didn’t sit around discussing the same material. Each one handled a specific job: identifying terrain, filtering candidate sites, and cross-checking clues before passing its conclusions to human researchers.

Archaeology consultant Linda Nicholas, who joined the review work on Mexico’s Yucatán Peninsula, said anomalies detected by the AI still had to be assessed against local knowledge. A shape on the scan wasn’t grounds for announcing an archaeological discovery on its own.

The more useful lesson for us is simpler: one agent organizes, another categorizes, another checks for contradictions, and I make the final call. The point isn’t to outsource judgment to a machine. It’s to let the machine handle repetitive work first, leaving more time for communication and decisions.

Today’s cost to replicate it: one subscription, two hours, one human review

Money: Start with the free allowance on an AI tool I already have. If I don’t, a common personal plan costs roughly $20 per month.

Technical barrier: I don’t need programming skills, but I do need to explain three things clearly: what goes in, what gets checked, and what the agents must not do.

Time: Building the first version takes about two hours, followed by a one-week trial with a small sample.

First step: Open the AI tool I already use, create a project, upload 10 redacted client files, and write three rules: categorize them by client type, flag conflicting information, and don’t guess when something can’t be confirmed.

I once dumped an entire folder of material into AI, and it invented a client history that didn’t exist. Accuracy only became stable after I changed the workflow to “extract each document first, then cross-check.” For contracts, identification documents, and contact details, I remove sensitive information first.

This doesn’t require a complicated process for everyone. If handling the material manually takes less time, there’s no need to try it now.

Match it to my stage—there’s no need to copy 48 agents

If I’m just getting started: I’d do one thing: have AI organize my client materials once a week. No rush to build a team or feed it every file at once.

If I already have one or two clients: I’d add a “conflict checker” role that spots mismatched quote versions, requirement notes, and delivery dates, then asks me to confirm them.

If I’m scaling: I’d turn interviews, contracts, and delivery records into a repeatable workflow, while keeping a record of human reviews. At that point, the value isn’t showing off. It’s saving a new colleague from rereading every chat thread when they take over.

I can try it tonight with just 10 files. If the output is messier, I stop. If it saves an hour of repetitive work, I turn it into a weekly process.