This is a hypothetical business case by SasaDaily.
It is not an official Anthropic client case.
Today, we imagine a:
5-person small air conditioning maintenance company.
The team consists of:
1 owner.
1 administrative staff.
3 field technicians.
What they are truly busy with every day is:
Maintenance.
Cleaning.
Inspection.
Repairs.
But after some time in business, the owner realized:
A lot of time is actually spent not on fixing AC units,
but on:
administrative work before and after the repairs.
Customer sends inquiry at 8 PM: "Office AC isn’t cold. Can you come tomorrow?"
The admin is already off work.
The owner is at a job site.
The technicians are on their way home.
This inquiry:
often won’t be seen until
9 AM the next day.
Then more questions arise:
Which area?
What model?
How many units?
Completely not cold?
Or just less cold?
Any error codes?
Have they been serviced before?
What time do they want to schedule?
Only when all info is complete can scheduling happen.
Second problem: After on-site inspection, turning notes into a proposal
The technician observes on site:
Outdoor unit condition.
Piping.
Drainage.
Parts.
Space constraints.
When returning to the company,
they need to report to the owner:
What issues were found.
What might need replacing.
Estimated work hours.
Then someone has to:
整理报价内容。
Look for similar past cases.
Confirm materials.
Calculate labor.
Write scope of work.
Write warranty terms.
And send it to the customer.
This stage involves a lot of
redundant work.
Third problem: Sending the quote is not the end of the job
They must follow up on:
Has the customer responded?
Which quotes remain unconfirmed?
Which invoices are overdue?
Which customers promised to pay Friday?
Next week’s maintenance schedules?
Which units require advance parts ordering?
In a 5-person company,
there is no dedicated
Operations Analyst.
The owner essentially is the dashboard.
So this company doesn’t ask first: "Can AI run my company?"
That question is too big.
They only asked:
“What repetitive task segments can AI do that we can easily check after?”
They finally selected four areas:
Evening new inquiries.
Proposal after site inspection.
Monday operations brief.
Overdue invoice follow-up.
This precisely matches Claude for Small Business’s
latest workflows and strengths for integration.
First segment: Let AI organize evening inquiries but don’t rush it to quote
Imagine at 9 PM, an inquiry comes in via
website, email, or other channels.
Before, it would wait until the next day.
After implementing Claude, it can first:
Read the inquiry.
Organize:
Who the customer is.
Their needs.
Approximate equipment.
Location.
Urgency.
What info is missing.
Then check the calendar
for possible scheduling times.
And prepare a reply draft.
But in week one, the company does not let it send replies
They only conduct
shadow testing
(as discussed in yesterday’s “AI 1-minute tutorial”).
Claude follows the real process:
Reads data.
Organizes leads.
Prepares replies.
Checks possible slots.
But everything stops at drafts.
Nothing is sent.
No scheduling.
No promises like “We will be there at 2 PM tomorrow”.
The next day, the admin reviews only four things
Missing info:
Did AI miss any important data?
Errors:
Wrong model, location, date, or misunderstood requirements?
Boundary issues:
Did AI treat a mere inquiry as a confirmed appointment?
Timing:
How long would manual review plus AI draft take versus manually doing it completely?
Assuming 25 inquiries in the first week
(These are SasaDaily’s hypothetical numbers.)
Originally, each lead took about 8 minutes
to read, check missing info, reply, coordinate schedule, and record.
25 leads multiplied to about 200 minutes (3 hr 20 min).
After implementation, Claude organizes and drafts first
Admin spends only about 3 minutes per lead
to verify, edit, and decide whether to send.
25 leads take approximately 75 minutes.
In theory, this saves about 125 minutes (2 hours) per week.
However,
this isn’t an official Anthropic ROI,
just a test case estimation.
More importantly: Avoid automatic sending early on
This is because it’s critical to verify whether Claude can correctly distinguish which inquiries
are schedule-able, which are emergencies, which exceed service areas, which need photos first, or cannot be handled by email alone.
Fast AI replies are not always better if accuracy isn’t there.
Second segment: No need for technicians to rewrite reports after inspection
On Wednesday, a technician visits a small office.
Customer reports the AC has been inadequate recently.
The technician:
Examines equipment.
Takes photos.
Records abnormal findings, parts needed, and work constraints.
Back in the van, no computer is opened.
Instead, a 2-minute voice memo is recorded, for example:
Customer info.
Unit details.
Observations.
Suggested work.
Materials to confirm.
Items that are tentative and cannot yet be confirmed.
Claude converts these into a proposal draft
Claude for Small Business’s official Proposal Builder
is designed to handle voice memos, photos, RFPs, meeting transcripts, and similar on-site data
to organize into a proposal.
It can also cross-reference past projects, pricing data, and templates
to create an editable draft.
For this AC company, AI can organize six key sections
1: Customer needs.
2: On-site observations.
3: Suggested scope of work.
4: Possible required parts.
5: Proposed scheduling.
6: Pending pricing and warranty confirmation.
Note:
The last section is
not decided by Claude itself.
A critical human checkpoint: Technical judgment can’t rely on text alone
The voice memo is merely a record of the technician’s observations,
not a full diagnosis.
AI can organize, rewrite, and apply templates,
but a complete proposal should not mistake a “possible cause”
for a “confirmed fault.”
For example, if the technician says, "Possibly low refrigerant and need to check leaks"
AI turning this into “Fault cause: low refrigerant”
would change the meaning.
Proposal review must first check
what is confirmed
versus
what still needs verification,
ensuring they aren’t mixed up.
Second human checkpoint: Pricing
Claude may reference past jobs, proposals, and material prices
to help with drafting.
However, AC maintenance costs vary
based on parts, labor, floors, difficulty, crane usage, height, and site constraints.
Therefore,
final pricing decisions remain human.
AI can prepare but cannot commit on its own.
Third human checkpoint: Scheduling
Calendar availability on Thursday afternoon doesn’t guarantee the job can fit then.
Specific technicians, equipment, parts, or a two-person crew might be needed.
Therefore,
calendar openings
are considered candidate slots,
not firm commitments.
Fourth human checkpoint: Safety
Electrical, equipment, heights, refrigerants, disassembly, and safety protocols
must still follow professional personnel guidelines, factory demands, formal SOPs, and local regulations.
Claude is not the on-site safety officer.
The company’s final proposal workflow becomes:
Technician:
Completes on-site work.
↓
Records voice memo + photos.
↓
Claude:
Prepares proposal draft.
↓
Owner:
Reviews technical details.
↓
Admin:
Confirms scheduling.
↓
Owner:
Confirms pricing and terms.
↓
Finally, send to customer.
AI handles
a large part of organization, transcription, and formatting,
but does not replace
the professional judgment of either side.
If the company prepares 8 proposals weekly
Previously, admin and owner took about 25 minutes after technician reporting
to produce the first draft proposal.
8 proposals totaled about 200 minutes (3 hr 20 min).
After AI drafting, human confirmation takes only 10 minutes each
For scope, pricing, schedule, warranty, and outstanding items.
8 proposals take about 80 minutes.
Theoretically saving another 120 minutes (2 hours).
Again, this is a test estimate, not a guaranteed result from Claude use.
Together, the first two segments save roughly 4 hours weekly
Lead handling:
~2 hours.
Proposal drafting:
~2 hours.
Estimated monthly (4 weeks):
~16 hours.
If the internal cost averages
NT$500/hour,
the theoretical monthly time value is
~NT$8,000/month.
But this can’t just be claimed as “earnings”
There are still costs such as
Claude subscriptions, business tools, connectors, setup, review, and workflow maintenance.
Every company differs in lead volume, proposal complexity, and staff salaries.
This figure is
just useful for deciding
whether to continue experimenting,
not a guaranteed saving.
Third segment: Monday Brief no longer has the owner painstakingly piecing it together Monday morning
Previously, Mondays involved the owner opening
banking, accounting, CRM, calendar, and email
to check how many jobs were done last week, how much money was unpaid, which proposals were stalled, busiest days this week, unresponsive clients, and upcoming big payments.
Claude can consolidate all this into one brief
It doesn’t decide company direction,
but collects scattered operational signals from various systems
onto one page.
Highlighting this week’s top 5 important items,
which 3 invoices are overdue,
which 2 proposals have no response,
which day’s schedule is especially full,
and whether the pipeline increased or decreased from last week.
This workflow suits early automation even better than proposals
Since it is
read-only:
reading data, sorting, generating the brief.
No need to modify CRM.
No need to send emails.
No need for payments or customer promises.
If stable during shadow runs,
Monday Brief might be the earliest workflow to formalize.
The "quick Q&A on AI" permission topic applies here
The owner’s account may see everything,
but Monday Brief doesn’t need access to everything.
If it only needs sales, cash, invoice, pipeline, calendar data,
there is no need to open payroll, employee documents, sensitive contracts, or unrelated folders.
Don’t give AI all permissions just because the owner can see everything
People have dozens of responsibilities a day,
but Monday Brief is only a narrow task.
Since the task is narrower than the person’s full role,
AI permissions should be narrower than people’s full access.
This is called
Least Privilege.
Fourth segment: Overdue invoice follow-up can be drafted by AI but not sent automatically at first
One of the AC company’s biggest headaches
is not lack of revenue,
but
work done but payment not received.
Some customers simply forget.
Some companies have slow payment processes.
Some need additional documents.
Some have disputes.
You can’t send the same collection email to all.
Claude can help categorize
Which invoices just turned overdue, which have been overdue long, which customers have a history of prompt payment, and which frequently need reminders.
It then prepares follow-up drafts with differing tones.
But approval remains in the first phase
Because AI doesn’t know when a customer just spoke to the owner yesterday,
or whether payment was officially promised for next week, or if disputes are ongoing.
If AI automatically sends a harsh demand,
it can ruin the relationship.
Therefore, Claude drafts, humans review and send.
The company does not launch all 43 workflows at once
Claude for Small Business offers 43 workflows.
The worst approach
is installing all on Monday, scheduling all Tuesday, and nobody knowing what each AI agent is doing by Wednesday.
They test one workflow segment at a time
Week 1: Lead shadow run.
Week 2: Add proposal drafting if stable.
Week 3: Test Monday brief.
Week 4: Then invoice follow-up.
Each time asking,
“Does this save time?”
KPI need not be a corporate-grade dashboard
This company tracks five metrics:
New inquiry first reply prep time.
Lead data missing rate.
Proposal first draft prep time.
Significant manual edit rate.
Total AI + review time.
If time decreases and serious errors don’t increase, they continue.
If proposal edits exceed 60%
Don’t prematurely call it an automation success.
Look for causes like
lack of consistent voice memo format, outdated pricing, large service diversity, or insufficient context for Claude.
Sometimes the problem to fix isn’t AI itself
But the company’s original lack of consistent field report formats.
Tech A focuses only on equipment,
Tech B on price,
Tech C forgets work constraints.
AI can only organize input of varying quality.
So the company adds a simple field report card
Each technician’s voice memo must cover at least:
Customer needs.
On-site confirmation.
Pending confirmation items.
Recommended work.
Required parts.
Work constraints.
This isn’t to add admin work for AI,
but to standardize originally scattered verbal handovers.
Many AI implementations improve underlying workflows most
AI just makes problems visible.
Previously, admin thought “technicians never explained clearly,”
while technicians felt “I said everything.”
Now workflows must clearly define
input formats, permissions, stop points, and approvers.
This is the most worthwhile AI workflow improvement for small companies
Not merely adding a chatbot,
but transforming knowledge previously stored only in one person’s head
into visible, repeatable, testable, and transferable processes.
What will this company never let Claude decide on its own?
At least in the first phase, six items:
Formal fault diagnosis.
Official price quoting.
Safety-related judgments.
Customer arrival time commitments.
Refunds or payments.
Disputed customer communications.
AI can prepare, but not decide.
What tasks are most suitable for AI?
Also six,
Organizing new inquiries.
Identifying missing information.
Preparing reply drafts.
Converting site inspection notes to proposal drafts.
Compiling Monday operation briefs.
Drafting invoice follow-ups.
Shared features:
Clear steps.
Repetitive.
Data already exists.
Results easy to review.
Humans involved before impact on customers or money.
This is the most valuable way to use Claude for Small Business
Not telling small owners
"AI will run your company from now on."
But taking repetitive work scattered across email, calendar, CRM, accounting, payment, and documents
step-by-step
and bringing it out for AI assistance.
Then ask four questions for each workflow segment
First:
Where does input come from?
Second:
What can AI accomplish?
Third:
Which step must always stop?
Fourth:
Who is responsible for final approval?
If any of these remain unanswered,
don’t automate yet.
Finally, reconsidering those 16 hours saved monthly
If this company truly saves about 16 hours/month in repetitive tasks,
the best use of time savings is
not immediately laying off anyone,
but returning time
to real value-added work.
Technicians focus more on on-site issues.
Admin focuses more on customer care.
Owner spends more time improving proposal quality, service design, and customer relationships.
Because what customers ultimately pay for is not “how well you can organize CRM”
They pay for
AC units being truly fixed,
correct problem diagnosis,
clear pricing,
punctual service,
and accountability if issues arise.
These define a company’s value.
The best AI role is to
take away those repeatedly time-consuming "side" tasks first.
This case’s real lesson isn’t just that "AC companies can also use Claude"
It’s that a 5-person company
can start AI workflows in very small, manageable ways
without deploying ERP on day one,
without hiring AI engineers,
and without chasing 100% automation.
Pick one weekly repetitive, easily verifiable task,
run it once, then twice,
and prove if it’s really worthwhile,
then gradually let go.
If you want to know which step in your work is best suited for AI first, comment “workflow.”
Today, progress a bit with AI.
Learn one AI tip daily.
Save a bit of time daily.
Improve a bit every day.
SasaDaily, growing together with you.
