# Session 2 Handout: Guided Task Labs and Sanitizing Inputs

<p class="meta">FAST-06 Occupation-Specific AI Role Labs · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

Real work arrives messy: an email with a client's account number, a voicemail with a patient's name, notes with an employee's medical leave. The skill that separates safe AI use from risky use is cleaning the input **before** it reaches the tool, while keeping what the task needs. This session builds four more procedures and drills that skill.

## Key concepts

### 1. Sanitize before, never with

If you paste a document and ask the AI to "remove the personal details," the data is already in the tool. Sanitize in your head or in a scratch document first, then prompt.

### 2. Strip, keep, or stop

| Decision | When | Example |
| --- | --- | --- |
| **Strip** | Identifiers the task doesn't need | Customer name, order number, account digits |
| **Keep** | Content the task needs, not tied to a person | "Office supplies rose from $1,240 to $2,980; one-time furniture purchase" |
| **Stop** | The whole input is a record that can't be used | Resume, chart note, HR complaint. Use a template task instead or keep it human |

**Indirect identifiers** count: a first name plus an appointment date; a small business's name plus its monthly totals; a job title in a 45-person company.

### 3. Role checks that work

| Track | Typical output | Concrete check |
| --- | --- | --- |
| HR | Job posting | Pay range exact; bias-language list reviewed; requirements job-related |
| Bookkeeping | Variance explanation | Every amount matches; causes only from notes |
| Front office | Plain-language instructions | Clinical text unchanged; dentist approves |
| Customer service | FAQ answer | Every answer in the approved FAQ; gaps routed |
| Projects | Status report | Status, date, owner match notes; TBD if unknown |
| Office support | Inbox rule | Tested on 10 messages; nothing important hidden |

### 4. Reading-level rewrites (front office and anyone writing for the public)

Plain language means short sentences, everyday words, "you," and headings. Ask for about a 6th-grade level for patient materials. But **the meaning can't change**: ask the AI to list any sentence whose meaning might have shifted, and have the clinician approve.

## Worked examples

**Bookkeeping (Back Bay Accounting Partners).** Input: client email with name, business, bank, account digits, SSN. **Sanitized:** "Office supplies rose from $1,240 to $2,980 month over month; client notes cite a one-time furniture purchase." AI draft added "rising supply costs across the industry." **Check result:** removed (not in notes).

**Projects (Bayside Precision Parts).** Notes: "PPAP for BP-2207: late, Kevin chasing, new date TBD." AI status: "PPAP expected Nov 10; on track." **Check result:** corrected to "At risk; owner Kevin; date TBD."

**Customer service (Lowell BioSupply).** Call note with customer mobile and order number. **Sanitized:** "Customer's glove order delayed by supplier; agreed to send revised ship date by Friday." Follow-up draft promised "free expedited shipping." **Check result:** removed (not agreed).

## Checklist: sanitization drill

- [ ] Every direct identifier stripped.
- [ ] Indirect identifiers considered.
- [ ] What the task needs is kept.
- [ ] "Stop" inputs recognized; template alternative chosen.
- [ ] Sanitized before the prompt, never by the AI.

## Do and don't

| Do | Don't |
| --- | --- |
| Rewrite inputs as short sanitized summaries | Paste whole emails or forms |
| Use placeholders in templates | Fill templates with real data in the AI tool |
| Check numbers against the source report | Trust AI-added causes or context |
| Mark unknown dates TBD | Let AI estimate dates |
| Ask the AI to list possible meaning changes | Assume plain-language rewrites are safe |

## Try this at work

1. Take three real inputs you handle (don't paste them anywhere).
2. Write the sanitized version of each on paper.
3. Ask a colleague in your role whether anything identifying remains.

## Glossary

| Term | Meaning |
| --- | --- |
| Sanitize | Remove sensitive details before using content with AI |
| Direct identifier | Name, number, or detail that identifies a person on its own |
| Indirect identifier | Details that identify someone in combination |
| Reading level | Approximate school grade needed to read a text easily |
| PPAP | (Manufacturing) Production Part Approval Process |

## Key takeaways

- Sanitize before the prompt, never with it.
- Strip, keep, or stop.
- Each role has its own concrete check.