A first workflow you can understand

A training company receives: “Can you arrange a remote cloud workshop for eight staff?” Its approved catalog offers a two-hour remote workshop for up to ten people. The useful result is a reply draft and a client-record draft, with no invented price, booking or promised outcome.

This is a good first task because the input, allowed information and reviewer are easy to identify. The person checking the draft can point to the exact catalog entry. An inquiry without an email address stays in a review queue. A request for a guaranteed job stays there too, because the catalog makes no such offer.

You do not need programming knowledge to read the files. To run the optional checker, install Python 3.10 or later and use a terminal. Allow about 20 minutes. The exercise needs no account, package, paid subscription or API key. Every person and inquiry is fictional.

1. Open the complete starter folder

Download the inquiry starter workbook, then extract it into a new folder. It contains README.md, inquiries.csv, services.json, reviews.csv, prompt.txt, workbook.py and expected-intake.txt. The README explains each command and the negative cases. Open the CSV files in a spreadsheet or text editor.

Compare the four inquiries with the catalog before looking at reviews.csv. Decide which can produce a draft and which need staff attention. The supplied review file is a hand-written answer key for this small exercise. It is not a record of model performance.

2. Produce the draft handoff

Open a terminal in the extracted folder and run:

python workbook.py intake

Use python3 if that is the Python command installed on your system. The command creates a results folder containing reply_drafts.csv, client_records.csv and review_queue.csv. Its complete expected stdout is:

inquiries=4 reply_drafts=2 client_record_drafts=2 held=2
HOLD I3 missing_contact
HOLD I4 unsupported_service
files=reply_drafts.csv,client_records.csv,review_queue.csv
data=synthetic model_calls=0 emails_sent=0 crm_writes=0

Open the first reply. It describes the approved workshop and says a staff member must confirm pricing and availability. The client record is marked draft_for_staff. A CRM, or customer relationship management system, is where a team stores client records. This exercise produces a CSV handoff; it does not connect to or update a CRM.

3. Try a mistake deliberately

Change the I4 review to approve,checked. Run python workbook.py intake --output negative-results. The command must fail with ERROR: approved inquiry failed service, contact or group-size checks and create no output folder. An approval in the file cannot invent a service. Restore the reference review before continuing.

The supplied checker catches this contract failure. It does not prove that every possible customer request was understood correctly, or that a supplied email belongs to the named person. A reviewer still compares the inquiry, catalog and draft.

4. Add AI only where it helps

For this tiny structured form, fixed rules are sufficient. A language model may help with varied free-text requests, but there is no reason to assume it improves this task.

To test it, give a model you already use the complete prompt.txt, inquiries.csv and services.json. Keep the answer key hidden until its proposals are recorded. The prompt asks for a proposed decision and an explanation. Review every case yourself and enter only your final decisions into reviews.csv. Then generate drafts in a fresh output folder.

Record the time spent preparing inputs, checking proposals and correcting errors. Compare it with the same task completed without AI. Four familiar examples cannot establish general accuracy or savings. Never paste private customer information into an unapproved tool.

Adapt it and check completion

Replace the fictional catalog and inquiries with approved examples of your own task. Write the acceptance rules before adding more tools. Keep outgoing messages and real client-system updates behind a separate staff action until those integrations have their own tests.

You are finished when you can explain both held cases, reproduce every output line and find no unsupported promise in either reply. If the folder already exists, choose --output results-2. If a file is missing, check that you extracted the ZIP and opened the terminal in its folder.

Next, use the complete published lead-intake pilot below to examine evidence checks and a richer client-record handoff. Then measure the human work left after assistance. Reliability extensions become useful once you can explain the workflow they protect.