Admin Automation Log: 60% Cut in 90 Days

On a Tuesday in early May, I opened a spreadsheet and did something I had been avoiding for years. I wrote down every admin task I did that day, with a timer. Quote draft for a hotel shoot: 42 minutes. Invoice for a finished delivery: 18 minutes. Chasing a signed contract: 25 minutes. Meeting notes from a Zoom kickoff: 55 minutes. WhatsApp reply to a location scout question I had already answered twice: 12 minutes.

By Friday I had 18 hours logged. In one week. That is roughly 45% of a five-day workweek, and none of it was actual production work. I run a small event and commercial production studio in Taipei, one person plus the occasional freelancer. This is the diary of the next 90 days, in which I cut admin from roughly 18 hours a week to under 7. It broke twice, I rolled back once, and one experiment paid off in a way a client noticed before I did.

If you want the higher-level framing first, the Chinese version of this study lives at ai-workflow-3-months-real-results, and the operating model behind it is written up in Solo Stack Method.

Week 1-2: Baseline (log every task, find the repeat rate)

The first two weeks were pure observation. No automation, no changes. I used a plain Notion table with three columns: task, minutes, tag. The rule: no judging anything until I had 10 working days of data.

Two things showed up. Admin averaged 17-19 hours per week, peaking at 22 during a busy invoicing week. And when I tagged tasks by type, roughly 60% of the total minutes went to work I had done in nearly identical form before. Proposal drafts followed the same six sections. Invoices always needed the same tax fields and payment terms. Kickoff notes had the same shape: scope, deliverables, timeline, blockers.

About that number: it was 58% one week, 63% the next. I have been calling it 60% ever since, because rounding a real measurement is different from inventing one. The tagging took about 90 minutes across two weeks, which felt annoying at the time and turned out to be the most valuable 90 minutes of the quarter.

Week 3-4: First skill (proposal drafter), broke twice

I started with proposal drafting because it was the biggest single category, about 4 hours a week. I built a Claude Code skill that took a short brief (client name, event type, headcount, days, deliverables) and produced a first draft in the format I usually send.

Week 3, attempt one: it worked once, then quoted the wrong day rate because I had two rate cards in the reference file and the skill picked the wrong one. I noticed only because the number felt low. I almost sent it. Bad moment.

Fix: one rate card file per client tier, plus a rule that the skill must state which tier it is using in the first line of its output. A wrong tier now takes three seconds to catch instead of a full read of the draft.

Week 4, attempt two: the skill hallucinated a project timeline with a national holiday as a shoot day. Fix: I fed it my Google Calendar as read-only context and told it never to propose a date without checking. By end of week 4 I was saving roughly 2 hours a week on proposals, still eyeballing every draft carefully.

Week 5-6: Second skill (invoicing), unexpected win

Invoicing felt safer. The template is rigid, the fields are structured, and the failure mode, a wrong number, is easy to catch with a sanity check.

I built a small skill that read the delivered project record and filled the invoice template. Nothing fancy. The time saving was about 1.5 hours a week, less than I expected. The surprise came from elsewhere: two clients said some version of “wow, that was fast” because I sent the invoice the same day I delivered instead of batching for a Friday. One of them paid the same day. Cash flow moved forward by roughly 12 days on that project.

Faster invoicing was never a goal. It fell out of the automation because the friction of writing invoices had been silently making me batch them. You only find this kind of thing by measuring the before state, which is why the boring Week 1 exercise mattered.

By end of Week 6 the two skills had cut roughly 5.5 hours a week. Baseline 18, current around 12.5.

Week 7-8: Meeting notes, failed, rolled back

Confidence was high, so I went after the biggest remaining target: meeting notes. Kickoffs, mid-project check-ins, wrap calls. Roughly 3 hours a week.

The plan was to record calls, transcribe them, and have a skill extract scope, deliverables, blockers, and next actions. Two problems showed up fast. Transcription of Mandarin technical vocabulary was messy enough that the skill worked from a bad source, and the output looked confident while hiding subtle errors that took longer to catch than writing notes from scratch. And some clients were uncomfortable with recording. One asked directly whether the recording was stored anywhere I could not delete, and I had no good answer.

I rolled the whole thing back at the end of Week 8. Unrecoverable time invested: about 6 hours. The lesson: I had jumped straight to the hardest version of the problem, full recording, full transcript, full extraction, instead of starting with the easiest slice.

Week 9-10: Meeting notes v2, different angle

The second attempt was much smaller. Instead of recording the call, I typed a five-line summary the moment it ended: client, date, three bullet points of what was decided, one bullet of what I owe next. The skill turned that into a clean written recap I could send to the client within 30 minutes.

This worked. Not because the AI was doing more, but because it was doing less. The five-line summary took me about 4 minutes to type, and the skill spent about 30 seconds turning it into something that read like a professional recap. Clients started replying with “yes, confirmed,” and scope drift arguments dropped noticeably in the following weeks. I do not have a number for that yet.

Time saved on meeting notes: roughly 1.5 hours a week. Half of the original 3, with zero rollback risk.

Week 11-12: Final measurement

Two weeks of clean measurement at the end, same methodology as Week 1-2. Average admin hours per week: 6.8. Range: 5.5 to 8.1 depending on active project count.

From 18 down to roughly 6.8 is a 62% cut. I call it 60% in conversation for the same reason I called the repeat rate 60% back in Week 2. Real numbers, honestly rounded.

The recovered 11 hours a week went partly to actual production work, which was the point, and partly to writing skills for the next round of targets: client onboarding sequences, delivery follow-up. Some, I will admit, went to reading, walking, and not opening my laptop at 9pm on a Tuesday. That last one might be the real win.

Closing

Ninety days of this comes down to one solo operator measuring the boring parts of the job, automating three of them, breaking two, rolling one back, and ending up with 11 hours a week for work that mattered or a walk. The Solo Stack Method is the framework that came out of doing this on repeat, and the Chinese version of this same 90-day case is at ai-workflow-3-months-real-results if you want the parallel read.

If you are trying this yourself and something breaks in Week 3 the way my proposal drafter did, that is where the weekly QA in Solo Stack Studio actually helps. Members bring the broken skill, we look at it together, and usually the fix is the same shape as mine: split the reference file, force the output to declare which input it used, catch the wrong tier in three seconds instead of reading the whole draft.

FAQ

Was the 60% cut permanent, or did admin creep back?

It is holding at around 7 hours a week five months later. Two skills broke and needed patching (the proposal drafter after I changed my rate card structure, the invoicing skill after a tax rule update). Total patch time so far: about 2 hours over 5 months. The maintenance cost is real but small compared to the ongoing saving.

How long did the skills themselves take to build?

Proposal drafter: about 6 hours across two weeks, including the two breakages. Invoicing: about 3 hours. Meeting notes v2: about 90 minutes. The failed meeting notes v1 cost 6 hours I did not recover. Total build time roughly 16.5 hours to save roughly 11 hours a week ongoing. Break-even was around week 4.

Do I need to code to build these?

The skills I built are mostly plain-English instructions plus a few structured reference files: rate cards, invoice templates, calendar access. I write some code when needed, but the core of each skill is a written specification that Claude Code follows. If you can write a clear one-page brief for a freelancer, you can write a skill.

What would I do differently if I started over?

Start smaller on the hardest problem. My meeting notes v2 approach, five lines in, clean recap out, would have worked in Week 3 if I had tried it before the full-transcript version. I wasted 6 hours assuming more automation was better. Less automation, applied to the actual bottleneck, was the answer.