Five workplace scenarios, each mixing a legitimate AI task with data that must not be pasted: identify the risk class, redact or synthesize the sensitive parts, complete the task anyway, and log the before/after prompts. The skill is getting AI's help with sensitive-shaped work without the sensitive data ever leaving the building.
You are Tanvi Kulkarni, executive assistant to the COO at Saral Fintech Pvt Ltd, a fictional payments company — a role that touches customer complaints, salary letters, and board numbers before lunch. Saral's policy allows public AI tools for public-shaped content only; the approved enterprise workspace is down for maintenance this week, and the work cannot wait. Each scenario below is a real task worth doing with AI — containing data that must never reach a public tool. The naive answers are both wrong: pasting it all (a reportable incident) or refusing AI entirely (leaving the productivity on the table). The professional answer is the redaction discipline: classify, strip or synthesize, prompt, then re-insert the sensitive parts locally. One scenario contains data so hot that even redaction is the wrong answer — spotting it is part of the drill.
SCENARIO A - Customer complaint needing a response draft
Email received: "I am Rajiv Menon, account holder since 2019. On
2 Aug a UPI transfer of Rs 84,500 from my account xx4471 failed
but was debited. My registered mobile is 98xxx-xx341 and my PAN
is FKPPM1234Q as per your KYC. Your support has been useless for
6 days. Fix this or I go to the banking ombudsman and Twitter.
- Rajiv (rajiv.menon.77@example.com)"
TASK: Draft an empathetic, de-escalating response promising a
specific resolution timeline.
SCENARIO B - HR compensation question
Slack message from the COO: "Tanvi - need a comparison note today.
Our L5 product managers: Ankit is at 34L fixed + 6L bonus, Shreya
at 38L + 4L, new offer to Kunal Deshmukh (from PayNova) proposed
at 42L + 8L. Is the Kunal offer defensible vs market and vs
internal parity? Need talking points for the comp committee."
TASK: Produce a one-page internal-parity and market talking-points
note.
SCENARIO C - Pre-release financials needing a narrative
From the CFO's deck draft (BOARD CONFIDENTIAL, results announce
on the 18th): "Q1 revenue Rs 46.2 Cr (up 31% YoY), take-rate
improved 12 bps to 0.87%, EBITDA loss narrowed to Rs 3.1 Cr from
Rs 7.8 Cr, merchant count 1.42 lakh. Board asked for a crisper
'path to profitability' narrative for the analyst call."
TASK: Draft the 200-word path-to-profitability narrative for the
analyst call script.
SCENARIO D - Meeting invite needing a polish
Draft invite: "Team offsite planning - Goa, 21-23 Nov. Agenda:
H2 targets, product roadmap review, team awards night. Need RSVPs
by 20 Aug. Budget approved at Rs 12L. Contact Tanvi for dietary
preferences and travel dates."
TASK: Rewrite as a warm, well-structured invite email with an
RSVP section.
SCENARIO E - Security incident needing a summary
IT forwarded: "Incident 2201: On 4 Aug, misconfigured S3 bucket
'saral-kyc-docs-prod' exposed approx 18,000 customer KYC documents
(PAN images, Aadhaar-linked address proofs) to unauthenticated
access for ~11 hours. Access logs show 214 external fetches from
3 IPs. Patched 19:40. CERT-In notification assessment pending.
Legal review scheduled."
TASK: The COO wants a plain-language summary and a draft
communication plan by tonight.
Classify before you type. For each scenario, write one line in your log: what categories of sensitive data are present (customer PII, financial identifiers, employee compensation, undisclosed financials, security-incident detail), and the blast radius if it leaked (one customer, three employees, the company, 18,000 customers). Classification drives technique — you cannot pick the right redaction if you have not named what you are protecting.
Learn the three techniques, pick per scenario. Redaction: strip identifiers, keep structure ("customer, 6-day-old failed UPI debit of a mid-five-figure amount"). Synthesis: replace real values with fake-but-shaped ones and a mapping table you keep locally (Ankit→"PM-1 at X", 34L→"base A"). Abstraction: describe the class of problem and ask for a template with placeholders. Rule of thumb: redaction for narrative text, synthesis when relative magnitudes matter to the answer, abstraction when the AI only needs to supply structure and language.
Work Scenario A. The task needs tone and structure — not Rajiv's identity. Strip name, account digits, mobile, PAN, and email; keep the emotional facts (six days, failed-but-debited, escalation threat) and ask for a response template with [NAME], [AMOUNT], [TICKET], [DATE] placeholders. Fill placeholders locally, in your own mail client. Log the before→after prompt pair. Check the AI's draft for invented commitments — a promised timeline must be one you actually have authority to give.
Work Scenario B — and think harder than "remove names". Names are the easy part; the trap is that three data points with roles and a poaching source ("PM-1", "PM-2", "incoming from a named competitor") can still identify people to anyone who knows the team — small-cohort compensation data re-identifies easily. Synthesize: shift all figures by a consistent factor, use generic labels, drop the competitor name, and ask for the framework of a parity argument (what factors justify a premium offer, how to present internal equity). Rebuild the real note locally with real numbers. Log the mapping table location (local only).
Work Scenario C — then stop and re-read it. Classify honestly: these are unpublished, price-sensitive financial results of a company. Redacting the numbers leaves nothing to narrate; synthesizing fake numbers means the AI writes a narrative for figures that are not yours, then you rebuild it anyway. And even shape-preserved figures ("revenue up ~30%, loss halved") are themselves the undisclosed story. The defensible move: abstraction only — ask the public tool for a generic "path-to-profitability narrative structure for a payments company analyst call" with no figures at all, or hold the task entirely for the approved enterprise workspace or a human. Write your ruling and reasoning in the log; this scenario exists to prove that sometimes the answer is no.
Work Scenario D — the calibration case. Classify it: an internal offsite invite with a location, dates, and an approved budget. Nothing here is customer PII or undisclosed financials; the budget line is mildly internal (reasonable to trim on principle of least sharing), but this is essentially public-shaped content. Paste-with-light-trim is a legitimate ruling. The lesson is the mirror image of Scenario C: over-classification has a cost too — if every invite needs a redaction ceremony, people stop following the policy at all. Log your ruling; either "paste as-is" or "trim budget line" passes.
Work Scenario E — the one that is not yours to prompt. An active security incident involving 18,000 customers' KYC data, pre-legal-review, with regulator notification pending. Feeding incident details to an external AI tool is itself a further disclosure of breach information — potentially creating new legal exposure while legal review is underway. Correct handling: no external tool touches this, not even redacted; the plain-language summary gets written by you (it is six sentences — write them), and the only acceptable AI assist is a fully abstract one ("outline a generic breach-communication plan structure per Indian CERT-In norms") with zero incident specifics. Draft your summary manually, log the ruling, and note who should own this task (legal + IT security, with the COO).
Close the log and extract your personal never-paste list. For each scenario, your log now shows: classification, technique chosen (redact / synthesize / abstract / paste / refuse), the before→after prompts, and the ruling rationale. Distill it into your own five-line never-paste list — the data classes you will not put into a public tool regardless of deadline — and pin it where you prompt. This artifact is the lab's deliverable, and it is the one document from this track worth printing.
Redaction protects individuals; it does not declassify information — undisclosed financials and active-incident details stay sensitive even with every name removed.
Classify, then choose: redact narratives, synthesize comparisons, abstract structures, paste the genuinely benign — and refuse the two classes no technique launders: undisclosed price-sensitive information and active-incident detail. The never-paste list you wrote today is company policy made muscle memory.