Fictional case study

Harbor Lane Clinic

One-line job: draft replies for hours and appointment changes, and stop there. The draft does not send. The tool does not read billing.

What happened

The front desk at this made-up clinic answers the same two kinds of mail every morning: what time are you open, and can this visit move. Nia Okonkwo, the front-desk lead, asked for a draft waiting in a staff queue before the desk opened. A later ask was to connect the send button "so we save a click," and to let the tool see the billing folder "in case the question is about a bill." Both asks were refused. The pilot stayed a draft queue. Staff still send the mail themselves. Billing stayed out of the index.

End to end

1. Sit with the front desk for one morning. Write who does the work today: a person reads the inbox, looks at the public hours page and the appointment book, and types the reply.
2. Write the metric in Nia's words, then in a sentence you can check. Her words: "A draft is waiting before we open, for hours and for moving a visit." The check: for those two message types, a draft is in the staff queue, and a person still sends it. A deflection rate is not this metric. A deflection rate is the share of questions that never reach a person. It can rise while the draft is wrong.
3. Write the pilot boundary. In: those two message types, the public hours page, and the appointment book the desk already uses. Out: sending, billing, clinical notes, prescriptions, and anything that writes to the chart. The written no is on the page: we will not send, and we will not read billing.
4. Build a loop that may only call a read of the hours page and a read of the appointment book. A step limit stops the loop if it never finishes. A tool named send is not on the list. The lab pattern is the same as an unknown tool: record it, do not run it.
5. If the appointment book does not return a time, the draft says the book did not answer. It does not copy a time the sender pasted into the mail.
6. Classify the data before anyone indexes a folder. Public hours may be read. The appointment book is staff-only, and only the fields the draft needs. Billing is a different class and is not in the corpus. Indexing the whole shared drive is the wrong call. It was refused.
7. Test with scripted messages, not with a live mailbox. One message asks for hours. One asks to move a visit and the book has the slot. One asks to move a visit and the book is silent. One asks about a bill. One tries to make the loop call send. The bill case and the send case must not produce a sent message or a billing fact.
8. Hand over a runbook, not a send button. The runbook says how to turn the draft job off, who may open the queue, and that the demo uses an account created for the pilot, not a personal admin login. The old habit stays: staff still type and send their own mail during the parallel run. Adoption is drafts a person edited and then sent themselves. It is not how many times someone opened the queue.

The buyer

Nia Okonkwo, front-desk lead. She runs the morning inbox. She is the buyer for this pilot. The clinicians were not the buyer. A discovery that only quotes a sponsor who never touches the inbox would be a bad discovery. The note names the desk staff who actually type the replies.

The metric she named

A draft is waiting before we open, for hours and for moving a visit.

Nia Okonkwo, front-desk lead. Fictional. Not a measured rate.

That sentence can be checked by looking at the queue. "The demo felt faster" cannot. "Tickets deflected" cannot. If someone later treats a deflection rate as success, say the mistake out loud: the rate can go up while the draft invents a time.

In scope and out

• In: hours questions, reschedule questions, the public hours page, the appointment book fields the desk already uses, a staff queue of drafts.
• Out: sending mail, billing, clinical notes, prescriptions, writing to the chart, and any other department's folder.

The wrong call that was refused

Two wrong calls. First: connect send so the desk saves a click. Refused. The queue is the product. A person sends. Second: index the whole shared drive so billing questions "just work." Refused. That is the named wrong call from search week: indexing the whole folder. Pay and billing stay out.

The security and data constraint

Billing is not in the corpus, so a prompt cannot retrieve it. The send action is not a tool, so a prompt cannot authorize it. Drafts sit in a staff queue. The constraint list ties each rule to a control: no send tool, billing path excluded before rank, and a pilot account that is not an admin login. A passing lab run does not grade this list. You still have to be able to say it.

How it was tested

Scripted messages only. No live clinic mailbox. The hours case must mention the hours page, not a guessed schedule. The silent appointment book must produce "the book did not answer," not a time copied from the mail. The billing case must not contain a balance, a code, or a claim number. The send case must record an unknown tool and stop when the step limit is used. Nothing in the test sends mail.

Handover

This one was handed over as a draft queue beside the existing inbox. It was not shipped as a sender. The runbook has: how the job is down, how to stop it, which version was running, who may run those steps, and the pilot account. Staff keep sending mail the old way during the parallel run. There is no cutover.

Consulting and life-sciences stream

This is the same fictional clinic, used as the short regulated case. It is not a sixth shop. It is not a real life-sciences engagement and not a placement. The pull-quote above, the diagram above, the sample draft below, and the refused reply below are the ones for this stream.

Written no. We will not send mail. We will not read billing. We will not read the chart. The no is on the page before anyone asks for a send button.

Corpus exclusion. The public hours page and the appointment-book fields are in. Billing and chart notes are not upserted. The exclusion is the missing rows, checked by a snippet on week 3, identity and audit. That snippet is not part of the 31 lab checks. A sentence that says "do not mention billing" is not the exclusion.

Parallel run. Staff keep the old inbox and still send the mail themselves. The draft queue sits beside it. A green health check does not turn the old inbox off. This page does not describe a cutover.

Handover their on-call can follow without you on the call:

1. How to stop the draft job, and which version was running.
2. Who may open the queue. The pilot account is not a personal admin login.
3. What the tool will not do: send, read billing, read the chart.
4. If a draft quotes a balance, a code, a claim number, or a chart sentence, stop the job and keep the old inbox.
5. The fictional front-desk lead is Nia Okonkwo. This page has no phone number.

Sample draft

A draft that would sit in the staff queue. It has not been sent.

Thanks for writing. The hours to quote are the hours on the public hours page, not a guess. The appointment book shows a later opening for this visit. This note is a draft. A person at the desk sends it.

Refused reply

What the tool answers when the mail is about a bill, or when someone asks it to send.

This tool does not read billing, and it does not send mail. I will not quote a balance, a code, or a claim number. A person at the desk sends the reply.

What to say in an interview

I studied a fictional clinic inbox. The buyer wanted drafts for two message types before the desk opened. I would describe the metric in her words, the written no on sending and on billing, and the test where a billing question returns nothing from billing. I would not say I was placed at a clinic, and I would not quote a deflection rate.

The same ideas are in week 1 (discovery, the metric, the written no), week 3 (who may read a document, corpus scope, leave pay out), week 4 (their constraints, each with a control), and week 5 (the pilot account and the runbook). Practice questions are on the interview page.