From three hours down to 25 minutes — I did the math, and not a single minute saved was judgment

Wednesday afternoon, reconciliation. Fourteen files spread across my desk: three xlsx exports from three systems, five csvs from suppliers, four WeChat screenshots, two PDF statements. Every one in its own format, dates sometimes text sometimes numeric, and one supplier records amounts in cents.

Used to take me two and a half to three hours. This time, from creating the folder to producing the summary table: 25 minutes, 10 of which were me going over it myself.

Two and a half hours saved sounds like a lot. But I broke those two and a half hours down, and the conclusion wasn't quite what I expected going in.

Which chunk of time actually got saved

What I was doing during those three hours before:

• Opening each of the five csvs one by one, checking whether the column names matched, manually shifting the ones that didn't;
• Date columns: some were 2026/9/1 as text, some numeric — unify the format;
• One supplier records amounts in cents — divide everything by 100;
• Delete blank rows, split merged cells, pull the words "tax included" out of the remarks column;
• Finally stitch it into one table, adjust column widths, save.

All that added up to over two hours. The common thread: not one step required thinking, only doing. While doing it I could listen to podcasts.

The remaining twenty-odd minutes were the part that actually needed judgment. That supplier's remark column saying "tax included" is actually a tax-rate marker; whether I'd read the entire remarks field on this PDF; two amounts that don't match — is it a duplicate row or genuinely two orders.

Now those 14 files run through in 25 minutes. The mechanical part is basically gone. The judgment part is still twenty minutes, not a minute less.

What AI ate was the zero-judgment time. Judgment didn't move at all, and a whole new layer of acceptance checking grew on top.

What they announce at launch events isn't the same as month-end reconciliation

I watched this October's wave of model releases. GPT-6 pushing Intelligent UI, saying conversation turns straight into operations, describe a need and it generates a usable interface. Haiku 5.5 cutting API cost to a tenth of the previous generation. Gemini 4 Argon going after long-horizon tasks in finance and law.

All sounds impressive. But sitting there staring at my 14 files, my first reaction was: these capabilities don't solve my problem.

What I need to handle is the pile of dirty work inside these 14 files: one supplier stores dates as text, one records amounts in cents, another writes "tax included" in the remarks column. Generating an interface isn't hard. What's hard is knowing that this supplier's remarks column is actually a tax-rate marker — and that knowledge is in my head, not in any model.

An industry report says enterprise-side AI returns lagged in 2026, with only 4% achieving their transformation goals. That number doesn't surprise me. The further up the stack the tools go, the more conspicuous the dirty work of the last mile becomes — and the small stretch of that dirty work that genuinely needs judgment is exactly the hardest part to hand off.

So my current usage is pretty crude

To make it checkable, I hard-code everything I can.

Three subdirectories at the root: raw/ for original files, tmp/ for intermediate output, out/ for the final table only. Rules hard-coded, so it doesn't have to judge where a file goes.

I don't type instructions on the fly anymore — I wrote them into rule.txt and dropped it at the root, about forty lines: fixed output column order, amounts unified to yuan with two decimals, dates unified to YYYY-MM-DD, remarks column only extracts tax-rate keywords, no adding columns, no renaming columns.

One thing at a time: first read and identify, I check the intermediate result; once confirmed, merge; finally run format validation separately. Three stages, and I can step in and interrupt at each one.

None of this is sophisticated. The only goal: the deader the rules, the faster I can spot-check. Every extra step it thinks, I have to verify one more step — that math doesn't work out.

Acceptance checking is a new cost, and it has to be counted

Plenty of pitfalls, and mostly of this kind: it got something wrong, and it took me time to discover it got it wrong.

The biggest pitfall was me. First time, cutting corners, I tossed in one sentence: "Help me tidy up these tables and merge them into one summary table." It went and planned its own classification logic, built sub-tables by supplier initial, and added two columns I never asked for: "data quality score" and "anomaly flags." Proofreading took longer than doing it myself.

Second was merged cells. Two suppliers had two-row merged headers; first pass read all the column names misaligned. Fix was adding a line to rule.txt: when encountering a merged header, take the bottom row as column names and concatenate the top row as a prefix. Never had a problem after that.

Third was missing fields. It often doesn't read the remarks field on PDF statements completely — still not fully solved — so I always reserve 10 minutes for a manual pass. I don't expect 100% from it.

So of those 25 minutes, 10 are buying insurance. That time won't disappear as models get stronger; it'll only shrink as the rules get clearer.

Right now the top vendors all talk about "delivery," about turning intelligence into revenue. I believe in the direction, but my understanding of delivery may differ from theirs. To me, delivery means letting AI clear out the zero-judgment mechanical time in your work — judgment stays with you, and so does acceptance checking.

So to evaluate whether a tool is worth it, I no longer watch launch events. I look at two things: what share of my work is mechanical time, and how big an acceptance surface it leaves me. The first determines how much it can help me; the second determines how much I still have to put in.

WorkBuddy is still a bit short of this. Sometimes it still wants to show off, still takes an extra step where you didn't ask. Every extra step means one more thing I have to verify. I wish they'd add a switch called "strict mode" — turn it on and it outputs only what's in the instructions, not one character more.

For people who like letting it freestyle, that feature is decoration. For someone like me who reconciles once a month and just wants to verify fast and go home, it's a must-have.

If it actually ships, I'll write another post.