Niklas Schmidt runs a patent practice, and the workflow he described was specific: plug into the firm's existing practice management system, and take over the "file it, rename it, tag it" emails — the ones that eat twenty minutes a day and teach nobody anything. Not a bot that answers clients. A filter that keeps the inbox from turning into a landfill.
If you're picking the first AI project for your own inbox, that instinct is the correct one, and it runs backwards from how most firms actually start.
Why drafted replies are the wrong first move
The AI email pitch that gets demoed first is almost always a drafted response — something that reads an inquiry and writes back an answer for you to approve. It feels like the bigger win. It's also the riskier place to put your first automation, because a bad reply is client-facing and a bad file name isn't.
A response that gets the tone wrong, cites the wrong matter, or promises something you didn't mean to promise costs real trust, even with a human reviewing before it sends. Review fatigue is real. The twentieth "looks fine" approval of the day gets less scrutiny than the first one did. The failure mode is visible and it lands on the relationship you're trying to protect.
Compare that to a misfiled email. Worst case, you search for it later and lose ninety seconds. Nobody's relationship takes a hit. Most of the time, nobody even notices. That gap between an expensive mistake and an invisible one is the whole argument for starting with filing.
What's actually eating the twenty minutes
Schmidt's estimate — close to half his inbound email is pure administration — tracks with what we see building this for other professional-services clients. Opposing counsel sends a document that needs to land in the right matter folder under the right name. A client forwards a receipt that needs tagging for billing. A calendar note needs acknowledging and archiving. None of it requires judgment. All of it requires attention, and attention is the resource actually in short supply.
The job isn't "read my email." It's "stop making me read the emails that don't need reading."
Keep your system of record as the trigger, not the AI
The detail in Schmidt's post worth stealing: the automation plugs into the practice management system, it doesn't replace it. That's the difference between a project that survives past month two and one that quietly gets abandoned. If the AI becomes the place filed documents live, you've built a second system nobody trusts as much as the first — and dual systems rot fast once people stop bothering to update whichever one feels optional.
The pattern that holds up: your document or practice management system stays the trigger and the record. An email comes in, a model reads it and decides where it belongs, and the result — a folder, a rename, a tag — gets written back into the system you already run. The AI classifies and routes. It never owns the data.
What this actually costs to run
This is a cheap category of automation, and the pricing backs that up. Claude Haiku 4.5, the model built for classification and routing work, runs $1 per million input tokens and $5 per million output tokens. A typical filing decision — an email body plus a short classification prompt in, a folder name and a couple of tags out — runs somewhere around 400 to 600 tokens total. Run that across 1,000 emails a month and the model cost lands well under a dollar. The expensive part of this project was never the AI. It's the few days of wiring it into whatever document or practice management system you already run.
When you're ready to go further
Filing earns trust fast because it's low-stakes, and you get a month of watching it work before it touches anything client-facing. Once it's proven itself, the next step usually isn't "let it reply." It's "let it draft, and hold for approval past a threshold you set." We wrote up a framework for where those approval gates actually belong once you get there. And before any of it, it's worth running the process through whether it's actually worth automating — filing clears that bar easily; not everything does.
If your inbox has a "file it, rename it, tag it" problem and you want to know what wiring it into your existing systems looks like, tell us what you're running. It's the kind of AI automation build we scope in days, not weeks, and we'll say so if it's smaller than you think.
— Cole
Sources
- Niklas Schmidt, LinkedIn — the post prompting this piece, on plugging AI workflows into an existing practice management system for routine email filing
- Anthropic — Claude API pricing — confirms Claude Haiku 4.5 at $1/$5 per million input/output tokens