AI Prompting for Drug Diversion: A Prompt Library

Diversion surveillance generates enormous data volumes — thousands of dispense transactions, waste entries, override events, and count records every day in an active hospital. Manual review is shallow and slow. Automated surveillance software helps, but the questions that matter change week to week and tool-generated alerts still require a human to interpret them.

AI assistants fill the gap between static dashboards and full custom analytics teams. They can:

The prompts in this library work across multiple tool categories with light adaptation. These are illustrative examples — not endorsements of any specific product. Your organization's policy governs which tools are approved for what data.

A strong diversion prompt has four parts:

Generic examples

You are a drug diversion analyst. I am uploading a 90-day dispense log for a hospital unit.
Columns: user_id, medication_name, qty_dispensed, dispense_timestamp, override_flag, waste_flag.
Identify the top 5 users by override rate. For each user, show their override count, total
dispense count, override percentage, and how that compares to the unit average.
Return a table sorted by override percentage descending.

You are a pharmacy compliance officer. I will paste a summary of a suspected diversion case.
Based on the summary, identify what evidence is present, what evidence gaps exist, and draft
5 clarifying questions I should ask during an investigative interview. Format as a numbered list.

Prompting tips

• Name exact fields. Reference column names from your data (e.g., override_flag, waste_timestamp). Generic prompts produce generic output.
• Say "flag" or "highlight" for anomalies. These verbs produce ranked lists and exception tables rather than descriptive paragraphs.
• Request the output shape. Ask for a table, a bullet list, a numbered list, or a drafted document — not just an "analysis."
• Paste a small sample first. For data uploads, paste 5–10 rows of your actual data at the top of the prompt so the AI understands your real column names and data types before analyzing the full file.
• Iterate with follow-ups. The first response is a starting point. Follow up: "Now sort by z-score descending and remove users with fewer than 10 dispenses in the period."
• Test on small samples before scaling. Run on one month before 12; one unit before the entire facility. Catch prompt errors cheaply.

Use these prompts against de-identified dispense logs, waste records, count discrepancy exports, and access logs. Never include patient names or MRNs — work with de-identified or aggregated data. Each prompt includes a one-line rationale.

1. Waste rate by medication, unit, and shift — flag >2 SD above unit mean

Why it works: Peer comparison with a standard deviation threshold is the most reliable method for separating true outliers from normal variation; raw waste counts alone are misleading because high-volume users will always have more waste.

You are a drug diversion analyst. I am uploading a 90-day controlled substance waste log.
Columns: user_id, unit, shift (day/evening/night), medication_name, qty_wasted, waste_timestamp.
For each combination of medication_name and unit, calculate each user's waste rate
(qty_wasted / total qty dispensed for that user). Then compute the mean and standard deviation
of waste rates for each peer group (same unit + same shift). Flag any user whose waste rate
is more than 2 standard deviations above the peer group mean.
Return a table: user_id, unit, shift, medication_name, user_waste_rate, peer_mean, peer_stdev,
z_score — sorted by z_score descending.

2. Overrides not followed by an administered dose

Why it works: A legitimate override means the patient urgently needed the medication — there should be a corresponding administration record. A high rate of overrides without documented administration is a primary diversion signal.

You are a pharmacy diversion analyst. I am uploading two files:
(1) override_dispenses.csv — columns: dispense_id, user_id, patient_token, medication,
    qty, dispense_timestamp
(2) mar_records.csv — columns: dispense_id, admin_timestamp, documented_by

Perform a LEFT JOIN on dispense_id. Identify all override dispenses with no matching
administration record in the MAR. For each user, calculate: total overrides, unmatched overrides,
and unmatched override rate. Flag users whose unmatched override rate exceeds 15%.
Return a ranked table sorted by unmatched override rate descending.

3. Unwitnessed waste entries

Why it works: Policy requires a second person to witness controlled substance waste. An unwitnessed entry creates an unverifiable gap that can conceal removal of the medication prior to disposal.

I am uploading a waste record export. Columns: waste_id, waster_user_id, witness_user_id,
medication, qty_wasted, waste_timestamp.
Identify all records where witness_user_id is null or blank. Group by waster_user_id and
calculate: total waste events, unwitnessed waste events, and unwitnessed rate.
Also flag any record where waster_user_id and witness_user_id are identical (self-witnessed).
Return two tables: (a) unwitnessed summary by user; (b) self-witnessed records list.

4. Partial-dose waste frequency by user

Why it works: Consistently wasting partial doses — e.g., always wasting the last 1 mg of a 2 mg vial — can indicate the user is keeping part of the dose. High partial-waste frequency relative to peers warrants closer review.

I am uploading a 60-day waste log. Columns: user_id, medication, qty_dispensed, qty_administered,
qty_wasted, waste_reason, waste_timestamp.
Calculate partial_waste_ratio = qty_wasted / qty_dispensed for each row.
Identify users whose median partial_waste_ratio for high-risk medications (fentanyl, hydromorphone,
morphine, oxycodone) is more than 1.5 standard deviations above the unit median.
Show each user's median partial_waste_ratio, the unit median, and the deviation. Sort descending.

5. Cycle-count discrepancies — 3 or more consecutive events

Why it works: A single count discrepancy may be a counting error; three or more consecutive discrepancies on the same pocket or user suggests a persistent and unresolved variance, which is a regulatory red flag and a diversion signal.

I am uploading an ADC count log. Columns: count_id, user_id, location, pocket_id,
medication, expected_qty, actual_qty, count_timestamp.
variance = actual_qty - expected_qty.
Identify any pocket_id + medication combination that has 3 or more consecutive count events
with a non-zero variance. For each such sequence, show: pocket_id, medication, number of
consecutive discrepant counts, date range, and total absolute variance.
Also show which user performed each count in the sequence.

6. Expected vs. actual count variance by location and user

Why it works: Absolute variance aggregated by location identifies high-risk pockets or cabinets; aggregating by user reveals whether the same person is consistently present at the point of discrepancy.

Using the count log I uploaded (columns: user_id, location, pocket_id, medication,
expected_qty, actual_qty, count_timestamp):
(a) Rank locations by total absolute variance (sum of ABS(actual - expected)) over the past 90 days.
(b) For the top 5 locations, show variance broken down by user_id — who was performing the
    counts with the most variance at each location.
Return two tables with location, medication, total_absolute_variance, count_of_events.

7. Off-hours cabinet access patterns

Why it works: Access outside normal shift hours is a known risk factor — ASHP guidelines recommend reviewing off-hours ADC access. Clustering of off-hours activity on specific users or specific times of day is especially meaningful.

I am uploading a 30-day access log. Columns: user_id, location, access_timestamp,
duration_minutes, transaction_type.
Define off-hours as: any access between 19:00 and 07:00 on weekdays, or any access on
Saturday or Sunday.
For each user: calculate total access events, off-hours events, and off-hours rate.
Also produce a heatmap table: rows = hour of day (0–23), columns = day of week (Mon–Sun),
values = count of off-hours access events across all users.
Flag any user with an off-hours rate above 25% and more than 5 off-hours events.

8. Duplicate administration timestamps

Why it works: Two administration records for the same patient, same medication, and nearly identical timestamps may indicate a copy-paste documentation error — or may indicate that one administration record was created to justify a dispense that was actually diverted.

I am uploading a MAR extract. Columns: admin_id, patient_token, medication, dose_mg,
admin_timestamp, documented_by.
Find any records where the same patient_token and medication_name have two or more
administration entries within 5 minutes of each other. List all duplicate pairs with:
patient_token, medication, admin_timestamp_1, admin_timestamp_2, documented_by_1,
documented_by_2, time_difference_seconds.

9. Same user + same patient repeated access

Why it works: A single user repeatedly accessing a controlled substance for the same patient at a rate much higher than the patient's clinical orders can support is a classic pattern in investigations involving substitution (replacing drug with saline or other liquid).

Using the dispense log (columns: user_id, patient_token, medication, qty, dispense_timestamp,
order_id):
Find user + patient pairs where the same user_id dispensed the same high-risk medication for
the same patient_token 5 or more times within a 12-hour window. For each identified pair, list
the dispense timestamps, quantities, and whether each dispense matches an active order_id.
Highlight any instance where dispenses lack a matching order (override_flag = true or order_id null).

10. Provider peer comparison — volume, waste rate, override rate

Why it works: No single metric is sufficient; a composite peer comparison table puts each provider in context and surfaces the combination of high volume + high waste + high overrides that is most predictive of diversion.

You are a diversion analyst preparing a monthly surveillance review. I am uploading a
combined dispense and waste file for the past 30 days.
For each user_id in the same role group (e.g., RN on Med-Surg 4E), calculate and display:
- total_dispenses
- total_waste_events
- waste_rate (waste / dispenses)
- override_count
- override_rate (overrides / dispenses)
- unwitnessed_waste_count
- not_charted_rate (dispenses with no MAR match / total dispenses)
Rank users by a composite risk score = (waste_rate_z * 0.4) + (override_rate_z * 0.4)
+ (unwitnessed_rate_z * 0.2). Flag anyone with composite score above 1.5.
Return the full table sorted by composite risk score descending.

11. Not-charted administrations (dispense without MAR match)

Why it works: A controlled substance dispensed from the cabinet but never documented as administered to a patient is the clearest data-driven signal that something was removed for another purpose. This is the dispense-to-MAR gap that drives the most confirmed diversion investigations.

I am providing two tables joined on dispense_id:
- dispense_log: dispense_id, user_id, patient_token, medication, qty, dispense_timestamp
- mar_records: dispense_id, admin_timestamp, documented_by (may be null if not charted)

Identify all dispense records where mar_records.admin_timestamp is null (not charted).
Group by user_id and medication. Show: user_id, medication, total_dispenses, not_charted_count,
not_charted_rate. Also flag any not-charted dispense that occurred more than 4 hours before
or after the patient's documented pain score assessment.

12. High-dose or unusual-strength ordering patterns

Why it works: Ordering or dispensing atypically high doses — or concentrations outside the normal formulary range for a given clinical area — can indicate fabricated or altered orders intended to obtain larger quantities for diversion.

I am uploading an order and dispense file. Columns: order_id, ordering_provider_id, user_id,
medication, strength_mg, dose_mg, patient_weight_kg, unit, order_timestamp.
For each medication, calculate the mg/kg dose for each order. Flag any order where:
(a) dose_mg/kg exceeds the 97.5th percentile for that medication across all orders in this dataset, OR
(b) strength_mg is not a standard formulary strength (list the distinct strengths found and
    highlight any that appear fewer than 5 times in the dataset as potentially non-standard).
Return flagged orders with ordering_provider_id, medication, dose_mg, calculated_mg_per_kg,
and the dataset 97.5th percentile for that medication.

13. Open-ended pattern finding

Why it works: Sometimes you don't know what you're looking for. Asking the AI to surface unusual patterns first — before you define the hypothesis — can reveal signals you hadn't considered.

You are a drug diversion analyst. I am uploading a 90-day controlled substance dispense log
(de-identified). Without me specifying what to look for, identify the top 5 most unusual
patterns in this dataset. For each pattern, explain: (1) what you found, (2) why it is unusual
compared to the rest of the data, (3) which users or locations are involved, and (4) what
additional data you would want to confirm or rule out diversion as an explanation.
Be specific — cite user IDs, timestamps, and quantities from the data.

1. Summarize a user's dispensing history vs. peers

Why it works: A structured summary puts all the key metrics in one place and makes the peer comparison explicit — the foundation of a defensible investigation file.

I am providing a 90-day data summary for user ID [USER_ID], a [role] on [unit].
The summary includes: total dispenses, waste rate, override rate, unwitnessed waste count,
not-charted rate, off-hours access events, and the same metrics for their peer group.
Write a concise 200-word professional summary comparing this user's behavior to their peers.
Identify which metrics are within normal range, which are elevated, and what the combination
of elevated metrics suggests as a surveillance priority. Use neutral, factual language —
no conclusions about intent. This will be reviewed by HR and legal counsel.

2. Build a timeline of events from a transaction log

Why it works: Investigators need a chronological narrative. Raw transaction logs make this tedious; AI can restructure them into a readable timeline with cross-system events interleaved.

I am uploading a transaction export for a 30-day investigation window. The file contains
dispense events, waste events, MAR administration records, and access log entries — all
for a single user (ID: [USER_ID]).
Build a unified chronological timeline. For each event, show: timestamp, event_type
(dispense/waste/admin/access), medication, quantity, relevant flags (override, unwitnessed,
not_charted, off_hours). Highlight in the timeline any 30-minute windows where multiple
suspicious events occur in close succession. Format as a table sorted by timestamp.

3. Draft interview questions for a suspected diversion case

Why it works: Good interview questions are specific to the evidence — not generic. Providing the AI with the factual pattern summary produces targeted questions that cover each anomaly without telegraphing conclusions.

You are assisting a drug diversion investigation team. The subject is a registered nurse with
the following documented anomalies (no names used): (1) waste rate 3.1 SD above unit peers
for fentanyl over 60 days; (2) 18 out of 22 fentanyl waste events unwitnessed; (3) 7 off-hours
cabinet access events not aligned with scheduled shifts.
Draft 10 open-ended, non-leading interview questions designed to explore these specific patterns.
Include 3 behavioral questions, 3 documentation-focused questions, and 4 situational questions.
Format as a numbered list with a brief note on what each question is intended to surface.

4. Draft a DEA Form 106 loss narrative

Why it works: The narrative section of a DEA 106 must be complete, factual, and timeline-accurate. AI can produce a structured draft from your investigation summary faster than starting from a blank page — and the draft surfaces narrative gaps before submission.

You are a compliance officer at a hospital pharmacy. Draft the loss narrative for a
DEA Form 106 based on the following confirmed facts:
- Drug: [medication name], Schedule [II/III/IV]
- Quantity: [X units] confirmed missing after inventory reconciliation
- Discovery date: [date]
- Period of potential loss: [start date] to [discovery date]
- Discovery method: [automated surveillance alert / routine count / tip]
- Steps taken: [list actions taken after discovery]
Write in neutral, factual, third-person language. Do not speculate about method or intent.
Do not include employee names — use role titles only. Aim for 150–200 words suitable for
submission to the DEA. I will review and complete identifying fields before filing.

5. Summarize a policy or regulation into plain language

Why it works: Investigators and managers often need to understand a regulatory requirement quickly without reading the full CFR section. Plain-language summaries speed up decision-making without sacrificing accuracy.

Please summarize 21 CFR Part 1301.76 (employee screening requirements for DEA registrants)
in plain language suitable for a non-attorney compliance manager. Include: what the regulation
requires, who it applies to, what "disqualifying" means under the rule, and three practical
steps a hospital pharmacy should take to comply. Keep it under 300 words. Preserve the
CFR citation so staff can reference the primary source.

6. Identify evidence gaps in an investigation summary

Why it works: Before closing or escalating an investigation, a structured gap analysis prevents premature conclusions and strengthens the file against legal challenge. AI can act as a neutral second reviewer.

You are reviewing a drug diversion investigation summary for completeness before it is
submitted to HR and legal. The summary I am providing covers: timeline of events, surveillance
metrics, interview notes (anonymized), and current evidence on hand.
Identify: (1) which elements of a complete investigation file are missing or incomplete;
(2) what additional records should be obtained (e.g., specific date ranges of transaction logs,
access control records, pharmacy camera footage timestamps); (3) any factual inconsistencies
between the timeline and the metric summary; (4) questions that remain unanswered.
Present findings as a numbered gap list with a recommended action for each gap.

7. Draft a patient notification letter

Why it works: Patient notification letters must be empathetic, legally vetted, and factually precise. An AI draft gives legal and clinical leadership a structured starting point and surfaces language issues before internal review. See Patient Notification After Drug Diversion for the full guidance on when and what to notify.

You are helping draft a patient notification letter for potential exposure through drug
diversion. The confirmed facts are: [medication] was potentially diluted or substituted
during administration in [clinical setting] between [start date] and [end date]. Patients
who received this medication during this window may have received a reduced or ineffective dose.
Draft a patient notification letter that: (1) explains what happened in plain, non-alarming
language; (2) states what the facility is doing in response; (3) specifies what patients
should do (e.g., contact the provided hotline, seek testing); (4) includes a placeholder
for the hotline number and primary contact name. Keep the tone compassionate and clear.
This draft will be reviewed by legal counsel before distribution — mark it DRAFT at the top.

8. Create a corrective action plan from audit findings

Why it works: Corrective action plans need ownership, timelines, and measurable success criteria. AI can convert a list of audit findings into a structured plan template that managers can complete and track.

I am providing a list of 6 audit findings from a controlled substance diversion review.
For each finding, generate a corrective action plan entry that includes:
- Finding description (restate concisely)
- Root cause category (process / training / technology / oversight)
- Recommended corrective action (specific and actionable)
- Responsible role (e.g., Diversion Prevention Coordinator, Nurse Manager, IT)
- Target completion timeframe (30 / 60 / 90 days)
- Success metric (how we will know it's fixed)
Format as a table. I will assign actual names and dates after review.

1. Executive summary of monthly surveillance metrics

Why it works: Leadership needs a one-page narrative, not a data dump. AI converts metric tables into executive-readable prose that highlights what changed and what requires attention.

I am providing monthly surveillance metrics for [Month Year]: waste rate, override rate,
unwitnessed waste count, not-charted rate, off-hours access events, count discrepancy rate,
and the number of active anomaly flags — with the prior month's numbers for comparison.
Write a 250-word executive summary suitable for the pharmacy and therapeutics committee.
Structure as: (1) one-sentence overall status; (2) what improved; (3) what deteriorated or
requires attention; (4) open action items with owners. Use plain language — no jargon.
Do not include individual employee information.

2. Leadership board report bullet points

Why it works: Board members need 3–5 bullets, not a full report. Concise framing of surveillance program status builds institutional trust without overloading non-clinical board members.

Convert the following 6-month surveillance program summary into 4 board-ready bullet points.
Each bullet should: start with a result (not a process), include one specific number, and
end with the so-what implication for patient safety or compliance. Keep each bullet under
30 words. Tone: confident and factual, not defensive.
[Paste your summary here]

3. Policy rewrite in plain language (keep citations intact)

Why it works: Compliance policies written in regulatory language are not read or retained by staff. Plain-language rewrites improve adherence without changing the legal meaning — keeping citations ensures traceability.

Rewrite the following controlled substance waste documentation policy in plain language
suitable for a registered nurse with no pharmacy background. Rules: (1) preserve the meaning
exactly — do not change any requirements; (2) keep all regulatory citations (e.g., 21 CFR,
Joint Commission standard numbers) in the rewritten text; (3) replace passive voice with
active voice; (4) break long paragraphs into bullet lists where appropriate; (5) target an
8th-grade reading level. Mark the output DRAFT — PLAIN LANGUAGE VERSION.
[Paste your current policy text here]

4. Meeting summary with decisions, action items, and owners

Why it works: Surveillance review meetings generate verbal decisions that need to be captured with accountability. A structured AI summary from raw notes converts discussion into an actionable record.

I am providing raw notes from our monthly controlled substance surveillance review meeting.
Generate a structured meeting summary with four sections:
(1) Decisions made — a bullet list of any formal decisions or approvals
(2) Action items — a table with: action, owner role (not name), due date, and success criterion
(3) Items tabled — topics discussed but deferred to next meeting
(4) Next meeting agenda suggestions — based on open items
Do not include individual employee names in the output — use role titles only.

5. Email escalation draft (no patient identifiers)

Why it works: Escalation emails to HR, legal, or compliance must be factual, appropriately urgent, and stripped of any information that should not be in unencrypted email. A structured draft enforces these constraints.

Draft a professional escalation email to HR and the compliance officer notifying them of
an active drug diversion investigation. Include: (1) date the anomaly was identified;
(2) the type of anomaly (e.g., unwitnessed waste pattern, override-without-administration);
(3) the action already taken (e.g., flagged in surveillance system, data preservation initiated);
(4) what you are requesting from HR and compliance (e.g., guidance on next steps, interview
scheduling, access restriction decision). Do NOT include any patient identifiers, employee
names, or specific medication details in this draft — those will be communicated through
secure channels. Keep the email under 200 words.

6. Generate a DAX, SQL, or Python snippet for a specific metric

Why it works: Turning a surveillance question into a working query is often the bottleneck for teams without dedicated data engineers. AI-generated code dramatically reduces the time from question to verified answer.

Write a SQL query that calculates the 30-day rolling waste rate for each user in a hospital unit.
Table: fact_dispense. Columns: user_id, dispense_date, qty_dispensed, qty_wasted, unit.
The waste rate = SUM(qty_wasted) / SUM(qty_dispensed) over the 30 days ending on each date.
Also calculate the unit average waste rate for the same window and the z-score for each user.
Return: user_id, dispense_date, user_waste_rate_30d, unit_avg_waste_rate_30d, z_score.
Use standard ANSI SQL window functions. Add a comment explaining each major clause.

1. Generate red-flag training scenarios

Why it works: Realistic, unit-specific scenarios stick with staff longer than abstract lists. AI can generate tailored scenarios for specific clinical roles and settings in minutes.

You are developing diversion awareness training for a hospital ICU nursing team.
Generate 5 realistic red-flag scenarios that ICU nurses might observe. For each scenario:
(1) describe the observable behavior in 2–3 sentences from a colleague's perspective;
(2) identify which red flags are present (behavioral, documentation, or inventory);
(3) state what the observing nurse should do next (report to whom, how).
Scenarios should involve realistic ICU medications and workflows. Do not use real names.
Do not include patient identifiers. Make scenarios plausible but not accusatory in tone.

2. Create a quiz on waste documentation policy

Why it works: Knowledge checks reinforce training and create a documented competency record. AI can generate varied question types from policy text, saving hours of curriculum development time.

Based on the controlled substance waste documentation policy I am providing, create a
10-question knowledge check with the following format:
- 5 multiple-choice questions (4 options each; one clearly correct)
- 3 true/false questions
- 2 scenario-based questions requiring a written short answer (1–2 sentences)
For each question, include the correct answer and a one-sentence explanation referencing
the specific policy requirement it tests. Format for easy copy-paste into a learning
management system.
[Paste your waste documentation policy here]

3. Draft a new-hire awareness script

Why it works: New-hire diversion awareness orientation sets expectations from day one. A scripted, consistent message ensures every new employee receives the same foundational information regardless of who delivers the training.

Draft a 3-minute spoken script for a pharmacy manager or diversion prevention coordinator
to deliver to new hires during orientation. The script should cover:
(1) what drug diversion is and why it matters to patients and staff;
(2) three specific behaviors that are always monitored (waste documentation, override rates,
    count reconciliation);
(3) how to report a concern (non-punitive, confidential reporting channel);
(4) the personal and professional consequences of diversion.
Tone: welcoming but serious. The speaker should come across as supportive, not threatening.
Write in first-person speaker voice. Include a natural pause after the reporting section
to allow for questions.

4. Role-play a difficult investigative interview conversation

Why it works: Investigative interview skills atrophy without practice. AI can simulate a subject employee's responses — including evasive or emotional replies — allowing investigators to practice maintaining composure and follow-up questioning technique before a real interview.

Act as a role-play partner for an investigative interview practice session. You will play
the role of an employee (a hospital nurse) who is being interviewed about an anomaly in
their waste documentation — specifically, a pattern of frequently unwitnessed waste entries
over 60 days. The employee is cooperative but evasive about specifics, often giving vague
answers like "I don't remember" or changing the subject.
I will play the interviewer. After each of my questions, respond in character. If my
question is leading or closed, point it out after your character response so I can improve
my technique. Continue the role-play until I type "END SESSION", then give me a brief
coaching note on my interview technique: what worked, what to improve.

One-line prompts organized by goal. Paste any of these into your AI assistant as a starting point, then add your data and refine.

AI prompting is one layer in a comprehensive diversion prevention program. It accelerates surveillance analysis and documentation — but it depends on the data quality, governance structure, and human expertise that the rest of your program provides.