What to Automate FIRST: 10 Boring Business Jobs, Tested in Claude Code (10/10 Passed)

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: What to Automate FIRST: 10 Boring Business Jobs, Tested in Claude Code (10/10 Passed) · Channel: Hyperautomation Labs · 07:47 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Task and workflow automation, AI evaluation and benchmarks, Claude Code

Summary

The video tests Claude Code on 10 routine small-business jobs, using real files, plain-English instructions, and hidden answer keys to grade the results. All 10 jobs passed in a total of 24 minutes and 51 seconds, with tests using both the creator’s own data and a fictional plumbing company. The creator recommends starting with number-based tasks, followed by writing and drafting tasks, and leaving contract review for last because it requires human judgment. The video also cautions viewers to manually check results during the first month and notes that the fictional-company tests demonstrate the method, not performance on the viewer’s own files.

Key points

Each job was given as a folder of files plus one paragraph describing the desired finished result; Claude Code did not see the answer key.
A script graded each output against an answer key, and all 10 jobs passed.
The tests took 24 minutes and 51 seconds in total.
Three jobs used the creator’s business data with names removed; the other seven used a fictional six-person plumbing company.
The tested jobs covered bank statements, receipts, overdue invoices, email, customer messages, monthly sales, quotes, contracts, a weekly brief, and payroll.
The recommended first tasks are number-based jobs such as payroll, bank statements, invoice chasing, receipts, and monthly close.
Writing tasks can produce useful drafts, but a human still needs to review them and send them.
Contract review is a judgment task: the tool can flag issues, but a person must decide, and major matters still warrant calling a lawyer.
The creator advises checking outputs by hand every time during the first month.
The video says the fictional-company tests demonstrate the method, not how it will perform on a viewer’s own files.

Chapters

Start	Chapter	Summary
00:00	The test setup	The creator explains the 10-job test, its hidden answer keys, grading method, and use of both personal and fictional business data.
01:33	Jobs 1–5: bookkeeping, collections, and messages	Claude Code handles a bank statement, receipts, overdue invoices, an inbox, and real customer messages, passing the reported checks.
04:06	Jobs 6–10: business reports and operations	The remaining tests cover monthly close, quotes, contract review, a weekly brief, and payroll, with the tool flagging items it could not safely determine.
06:13	Recommended order and limitations	The creator recommends starting with number-based jobs, then drafts, and contract review last, while emphasizing human checks and the limits of the sample data.
07:17	Closing and resources	The creator invites viewers to comment for the prompts and scorecard and points to additional guides and social channels.

Quotes

Nobody tells you what to automate. — 00:00

all 10 passed. 24 minutes 51 seconds of claw code time start to finish. — 00:31

Start at the top. Do number one tonight. — 06:45

And for the first month, check the output by hand every time. — 07:17

Concepts and entities

Claude Code, Hyperautomation Labs, Gabarito de avaliação, Automação de tarefas empresariais, Revisão humana

Full transcript (original language)

00:00 Nobody tells you what to automate. That comment has 392 likes under one of the biggest automation videos on YouTube and nobody has answered it with proof. So I gave Claude Code 10 of the most boring jobs a small business has. The bank statement, the receipts, the overdue invoices, the inbox payroll. Every job got one plain English paragraph, one attempt, and an answer key. It never saw

00:31 all 10 passed. 24 minutes 51 seconds of claw code time start to finish. And the order you should do them in is not the order I expected. Here is how the test works because tested is the whole point. Each job is a folder. Inside the real files and one paragraph that describes what done looks like, not steps, just the finished result. Behind that folder sits an answer key. The right category

01:01 for every bank line. The right total for every quote. The right overtime for every employee. Clawed code never sees it. When the run ends, a script grades the output against the key. Three of the 10 jobs run on my own business data with the names removed. The other seven run on a sixperson plumbing company I built for this because I do not own a plumbers’s bank statement. The company is made up. The runs, the times, and the

01:33 scores are real. Job one, the bank statement. Three months of business checking, 131 lines, and a chart of accounts with 15 categories. The prompt asks for every line categorized, a monthly summary, and a list of anything it was unsure about. 1 minute 40 seconds, 131 out of 131 lines in the right category. And it flagged the two identical Home Depot charges on the same

02:03 day as a possible double charge, which is exactly what a bookkeeper would do. Job two, receipts. 27 files, phone photos and PDFs, messy names, and one duplicate I planted. It reads every image, checks each printed total against the line items, and builds the report. 1 minute 43 seconds, every date, vendor, category, and total correct, the grand total to the scent, and it caught the

02:33 duplicate invoice and counted it once. Job three, chasing money, 30 invoices, a client list, and a reminder policy with four tiers. Find what is overdue as of today. Work out the balance and write the right email for each one. 1 minute 45 seconds. 12 overdue invoices found. 12 emails written in the right tone. And it skipped both traps. An invoice that is due today and one that was already

03:05 paid but still marked as sent. Job four, the inbox. 40 emails from one week. New leads, a customer with no hot water, supplier bills, spam, a permit notice. Label everyone and draft a reply for every lead and every emergency using next week’s open slots. 3 minutes 28 seconds, 40 out of 40 labels, and 11 reply drafts that ask the questions a plumber actually needs to ask. Job five

03:35 is real. 40 messages my own customers and viewers sent, names removed. Sort them into question, complaint, praise, request, or spam. Flag anything a human has to handle and draft the reply. 4 minutes 51 seconds. 80 out of 80 checks. And the two abusive messages were flagged for a human with no reply written, which is the right behavior. Five for five so far. If this is useful, subscribe. It is free. And to back the work, hit join here or the supporter

04:06 button on Facebook. Five more and then the order. Job six, the monthly close on my real sales export. 123 orders, gross, fees, net, tax collected, refunds, revenue by product, top countries, and five plain sentences about the month. 1 minute 6 seconds. Every number matched to the scent. And it noticed that the export ends with a totals row and left it out, which is the mistake that doubles your revenue if you

04:38 add up the file blindly. Job seven quotes. A rate card with labor rates and after hours multiplier. A trip fee radius, parts markup, and fixed prices for the common jobs. 15 customer requests. Write the quote for each. 2 minutes 21 seconds. 15 totals correct. Every deposit correct. And the one request that could not be priced without a site visit was flagged instead of guessed. Job 8 contracts, five

05:10 agreements, a vehicle lease, a software contract, a subcontractor deal, a supplier credit line, a maintenance agreement with 20 red flags planted inside them. Read each one against my checklist and tell me what to negotiate. 5 minutes 23 seconds. 20 out of 20 flags found. Zero false alarms. And it told me to walk away from the software contract. Job nine, the Monday brief from my real channel numbers and my real sales this

05:42 week against last week. What moved the top three and three actions 1 minute 28 seconds. Every figure and every percentage correct. Job 10 payroll. Six employees, two weeks of clock punches, overtime past 40 hours, and one salaried person to leave out. 1 minute 7 seconds. 29 out of 29 checks. It found the missing clock out and refused to guess the hours. And it caught a duplicate

06:13 punch that would have pushed one person into overtime. Nothing estimated. So here is the board. 10 jobs, 10 passed, 24 minutes 51 seconds in total. Now the order because this is the part nobody tells you. Sort them by what comes out the other end. numbers first, payroll, the bank statement, invoice chasing, receipts, the monthly close. The output is a file you check once and then trust. And these are the jobs you pay a

06:45 bookkeeper, a payroll add-on, or a collections app to do. Word second, the Monday brief, the inbox quotes, customer replies. Excellent drafts, but a human still hits send, so the saving is smaller. Judgment last contract review. It flags you decide. And for the big ones, you still call a lawyer. Start at the top. Do number one tonight. Three honest limits. The plumbing company is made up. So the seven sample jobs prove

07:17 the method, not your files. The runs use the default model on a paid claude plan. And for the first month, check the output by hand every time. After that, you will know exactly which jobs you can stop checking. Real talk. 90% of you watching haven’t subscribed. On Facebook, subscribe is the supporter button. That is what keeps the lights on here. On YouTube, it’s free and it tells the algorithm to send you more.

07:47 Subscribe to support us. Comment boring on this video and I will send you all 10 prompts exactly as I typed them, the scorecard, and the order. free. If you want the full playbooks, my four paid guides are linked below. Follow along on Instagram and Facebook for byte-size AI and the deep dives live here on YouTube. Become a member here on YouTube and a subscriber on Facebook. Tested with receipts. See you in the next one.