Half-day seminar · For licensed investigators

Prompt & Verify

AI for Investigators

A four-hour, hands-on seminar that teaches licensed investigators to use general-purpose AI assistants safely, with more than half the time spent writing, testing, and checking prompts.

Register now See the agenda
4 hoursHalf-day, in person or live virtual
135 minof 240 spent on the prompt-engineering core
12–20participants, working at laptops
Any toolClaude, ChatGPT, Gemini, or Copilot

Core message

AI is a writing and analysis partner, not a data source.

It organizes, drafts, and reasons over information you lawfully obtained. It does not locate people or confirm facts.

Who it’s for

Licensed investigators and agency staff, from solo operators to agency owners.

No AI experience assumed.

Format

Half-day, in person or live virtual, 12–20 participants with laptops.

Also available as a webinar, a full-day course, or a three-part series. See formats

What to bring

A laptop and an account on at least one AI assistant (Claude, ChatGPT, Gemini, or Copilot). Prompts are tool-agnostic, so use whichever you have.

Before class: a five-question survey on your current AI use and biggest concerns.

What you’ll be able to do

By the end of the seminar, participants can:

  1. Explain why AI tools give confident wrong answers, and which investigative tasks that makes risky.
  2. Redact case material and choose a tool setup that respects client confidentiality and data-provider terms.
  3. Write prompts with the CASE framework that produce a usable draft in one or two tries.
  4. Apply AI to core investigative tasks: report writing, records reconciliation, affidavits, document digests, interview prep, and search planning.
  5. Verify AI output against source records before it reaches a client.
  6. Leave with a personal prompt library and a draft firm AI-use policy.

Agenda at a glance

Four hours, built around practice

Modules 3, 4, and 5 are the prompt-engineering core: 135 of 240 minutes.

Prompt-engineering core Foundations, guardrails & workflow Break
ClockBlockMinFormat
0:00Welcome, survey results, “what have you tried?”Discussion10Discussion
0:10Module 1: What AI is and isn’t for investigatorsLecture + live demo25Lecture + live demo
0:35Module 2: Guardrails — confidentiality, law, ethicsLecture + redaction drill35Lecture + redaction drill
1:10Module 3: Prompt engineering with CASELecture + prompt makeovers50Lecture + prompt makeovers
2:00Break—10—
2:10Module 4: Hands-on workshop stationsPairs at laptops60Pairs at laptops
3:10Module 5: Verification and quality controlLecture + error hunt25Lecture + error hunt
3:35Module 6: Workflow and firm policyTemplate walk-through15Template walk-through
3:50Wrap-up, Q&A, evaluationDiscussion10Discussion

Guardrails come before prompting on purpose. Every exercise after Module 2 uses redacted, fictional data.

More than half the clock is hands-on. Participants write, test, and check prompts at their own laptops.

Module by module

Six modules, six takeaways

Each module has one goal participants should be able to state on the way out.

Module125 min

What AI is and isn’t for investigators

Goal: Stop treating AI as a database.

  • A language model predicts likely text. Without a search feature it looks nothing up, and with one it reads public web pages, never licensed data.
  • Strengths: summarizing, organizing, drafting, rewording, translating, spotting inconsistencies in text you supply, and brainstorming angles.
  • Weaknesses: facts about private individuals, exact citations, long arithmetic, and knowing when it is wrong.
  • “Hallucination” in PI terms: an invented address, relative, case number, or statute, delivered in the same confident tone as real information.
  • Cautionary tale: the attorneys sanctioned in Mata v. Avianca (S.D.N.Y. 2023) for filing a brief with AI-invented case citations.

Live demo · 10 minA three-page fictional document and a question whose answer is not in it. Watch the plain prompt guess, then the fenced prompt from Module 3 answer “not found.”

Module235 min

Guardrails — confidentiality, law, ethics

Goal: Every later exercise uses redacted data and a defensible tool setup.

  • Client confidentiality. Assume anything pasted into a consumer AI tool leaves your control. Know whether your plan trains on inputs, how long chats are kept, and whether a business tier with a data agreement exists. Vendor terms change, so check them yourself.
  • Licensed data. Reports from DPPA- and GLBA-regulated databases carry permissible-purpose limits and contract terms. Read your provider agreement before uploading those reports to any third-party tool.
  • Attorney work product. Much PI work is done at counsel’s direction. Ask retaining counsel before running their case material through AI; how privilege applies to AI tools is still being settled.
  • Redaction habit. Replace names, DOBs, SSNs, account numbers, plates, and employers with labels (Subject A, Address 1). Analysis quality barely changes.
  • Off limits with any tool: pretext scripts to obtain phone or financial records (barred by the Telephone Records and Privacy Protection Act and GLBA), impersonation, fabricated or “enhanced” evidence, synthetic media, and facial recognition where state law restricts it.
  • Documentation. Note which tool, what went in, and how the output was verified. Prompts and outputs may be discoverable.
  • Licensing. Board rules differ by state; check yours for guidance on AI use and client disclosure.

Redaction drill · 10 minParticipants redact a one-page fictional case summary, then compare. Common misses: vehicle tags, employer names, nicknames, case numbers, and unusual job titles.

Module350 min

Prompt engineering with CASE Core

Goal: Write prompts that produce usable drafts and resist invention.

  1. The CASE framework, letter by letter (10 min).
  2. Five techniques that matter most in investigative work (15 min).
  3. Prompt makeovers: groups run a weak prompt and a CASE prompt on the same material and compare (20 min).
  4. Iterating instead of restarting: “Shorter.” “Use only the March log.” “What did you assume?” (5 min).
Module460 min

Hands-on workshop stations Core

Goal: Practice on realistic tasks using one shared fictional case file.

  • Pairs rotate through three of six stations, 18 minutes each, then a 6-minute debrief.
  • Every station uses the same fictional matter, so skills build on each other. See the stations
Module525 min

Verification and quality control Core

Goal: Nothing AI-generated reaches a client unverified.

  • Verify, then trust: trace every name, date, address, number, and citation to a source document.
  • Failure patterns to watch: two people with similar names merged into one, date drift, invented “connecting” details in narratives, overconfident inferences, and records silently dropped.
  • Prompt-side checks: ask for each claim with its source line; ask “what in your answer is not directly supported?”; paste the draft into a fresh chat and ask it to review against the sources.
  • The licensed investigator signs the report and owns every word.

Error hunt · 10 minTeams get an AI-drafted report with six planted errors and the source records. First team to find all six wins.

Module615 min

Workflow and firm policy

Goal: Leave with a repeatable setup.

  • Save prompts that worked as templates with [BRACKETED] fields.
  • Use the tool’s saved-instructions or project features for standing context: firm style, report format, redaction rules.
  • Walk through a one-page firm AI-use policy: approved tools, data rules, verification standard, documentation, client disclosure.
  • 30-day challenge: one AI-assisted task a week, logging time saved and errors caught.

The CASE prompt framework

A checklist you can recall mid-case

CASE is a four-part checklist investigators can recall mid-case, named with a word they already use every day.

C

Context

Who you are, the matter type, the purpose, and who will read the output.

“I’m a licensed PI preparing a diligent-search affidavit. The reader is a judge.”

A

Ask

One specific task, starting with a verb.

“Build a chronological address timeline.”

S

Source

The material to work from, fenced off, plus a rule to use only it.

“Use only the three reports between the record tags.”

E

Expectations

Format, length, tone, and what to do when unsure.

“Table with dates, address, source. Flag conflicts. Write NOT FOUND for anything missing.”

The reusable template

Goes on the pocket card and in the prompt library.

CONTEXT: I am a licensed private investigator working on [matter type].
The output will be read by [client / attorney / court].

ASK: [One task, starting with a verb.]

SOURCE: Use only the material between the <source> tags.
Do not add facts from outside it.
<source>
[Redacted documents]
</source>

EXPECTATIONS:
- Format: [table / narrative / bullets], about [length].
- Tone: neutral, factual, [first / third] person.
- For every fact, note which document it came from.
- Label any inference as INFERENCE and give its basis.
- If information is missing or conflicting, say so instead of guessing.

Five techniques to teach with it

  1. Fence the sources. “Use only” plus clear delimiters is the single biggest cut in invented facts.
  2. Make it show its work. Requiring a source for each claim makes errors easy to spot.
  3. Separate fact from inference. Clients, attorneys, and judges care about that line, and the model will blur it unless told not to.
  4. Chain the steps. Extract, then reconcile, then draft, then review, each its own prompt, checking output between steps.
  5. Show the format. Paste a redacted past report and say “match this structure.”
Bonus move: End a complex prompt with “Before you start, ask me any questions you need answered.” It surfaces missing context before the model fills gaps on its own.

Prompt makeovers

Three before-and-after pairs from Module 3. Groups run both versions on the same fictional material and compare.

1. Surveillance notes → report narrative
Before

Write up my surveillance notes.

After 
I'm a licensed PI writing a surveillance report for an insurance defense attorney. Convert the field notes below into a report narrative.

Use only the notes between the <notes> tags. Do not add details, motives, or descriptions that aren't in the notes.

<notes>
[Redacted, timestamped field notes]
</notes>

Write in first person, past tense, chronological order, one paragraph per observation period. Keep every timestamp exactly as written. Refer to the subject as "the Subject." After the narrative, list any gaps of more than 30 minutes in the notes.

Why it works: It names the reader, fences the source, blocks embellishment, protects timestamps, and surfaces the gaps opposing counsel would ask about.

2. Records reconciliation for a skip trace
Before

Where does this guy live now?

After 
I'm a licensed PI locating a witness for subpoena service. Below are address histories for Subject A from three licensed database reports, redacted. Build one address timeline from these records only.

<report_1>...</report_1>
<report_2>...</report_2>
<report_3>...</report_3>

Output a table: address label, earliest date seen, latest date seen, which reports list it. Sort newest first. Then list every conflict between reports and every address that appears in only one report.

You may say which address has the most recent date, labeled INFERENCE, with your basis. I will verify in the field.

Why it works: The model cannot know where anyone lives. It can organize what your licensed sources already say and show you exactly what to verify.

3. Witness interview prep
Before

Give me questions to ask a witness.

After 
I'm interviewing a witness to a commercial vehicle accident for the defense attorney. The witness's police statement (redacted) is below.

<statement>...</statement>

Draft 15 open-ended questions covering vantage point, timing, lighting and weather, what drew their attention, and any contact with either party since. Avoid leading questions. Then flag any statements that are vague or internally inconsistent, quote each one, and suggest a neutral follow-up question for it.

Why it works: It gives the model the case theory’s terrain, bans leading questions, and turns the statement itself into the checklist.

Workshop stations

One case file, six realistic tasks

Each station card carries the scenario, a weak starter prompt to improve, success criteria, and a verification step.

FICTIONAL

Matter 26-114. All six stations draw on one fictional case file: locating a witness, “Subject A,” for a deposition subpoena in a premises-liability case. It includes mock database reports with overlapping and conflicting addresses, field notes, an attempt log, a witness statement, and a deposition excerpt.

Station 1

Report writer

Turn field notes into a report narrative.

Fencing sources, preserving facts

Station 2

Address timeline

Reconcile three mock database reports.

Structured output, conflict flagging

Station 3

Diligence affidavit

Draft an affidavit of diligent search from an attempt log.

Writing for a court, fact vs. inference

Station 4

Document digest

Summarize a 10-page deposition excerpt with page:line cites.

Long documents, citation discipline

Station 5

Interview prep

Build a question list and flag inconsistencies.

Open questions, critique

Station 6

Search planner

Generate name variants, username patterns, and a records checklist.

Brainstorming; AI suggests, the PI searches

Worked example: Station 3, diligence affidavit Scenario, log, prompt, criteria

Scenario

Service attempts on Subject A have failed. Counsel plans a motion for alternative service and needs the factual body of an affidavit.

Attempt log (fictional)

09/08 7:42 PMAddress 1. No answer. Two vehicles in driveway. Left card.
09/10 6:15 AMAddress 1. Adult female resident, declined to give name, stated Subject A moved out “in the summer.” No forwarding address.
09/11Database update lists Address 2, reported 08/2026 (one source).
09/12 5:30 PMAddress 2. Manager confirmed lease in Subject A’s name, Unit 4B. No answer at unit.
09/14 8:05 AMAddress 2, Unit 4B. No answer. Neighbor stated Subject A works night shifts.
09/15 9:50 PMAddress 2, Unit 4B. No answer. Lights off.

Model prompt

CONTEXT: I'm a licensed PI. A process server needs an affidavit of diligent search to support a motion for alternative service. The reader is a judge.

ASK: Draft the factual body of the affidavit from the attempt log.

SOURCE: Use only the log between the <log> tags.
<log>[attempt log]</log>

EXPECTATIONS:
- Numbered paragraphs, first person, past tense, chronological.
- Keep every date, time, and address label exactly as written.
- State what was observed or said. Do not describe the Subject as evading or avoiding service.
- Attribute every third-party statement to its source.
- After the draft, list facts a judge would likely want that the log does not contain.

Success criteria

  • No facts added beyond the log, and no “evading” language.
  • Each third-party statement attributed (“the resident stated”).
  • The gaps list catches items such as databases searched, phone or mail attempts, and who was spoken to.

Verification step

Check every paragraph line by line against the log. Caption, statute references, and format come from counsel, not the model.

Debrief question

Where did the plain prompt add drama or conclusions the log doesn’t support?

What you’ll leave with

CASE pocket card

Framework and template on the front, verification checklist on the back.

Prompt library

10–15 tested templates with [BRACKETED] fields, one per common task.

Firm AI-use policy

A one-page sample: approved tools, data rules, verification standard, documentation, client disclosure.

30-day challenge

One AI-assisted task a week, logging time saved and errors caught.

Format options

Sized to your group

The half-day plan scales down to a webinar or up to a full day without changing its core: CASE, guardrails, and verification.

Webinar

90 min

Modules 1, 2, 3, and 5 condensed; live demos replace the workshop; prompt library sent after.

Featured

Half-day seminar

4 hr

The full program described on this page.

Full-day course

7 hr

All six stations; an advanced module on multi-document chaining, image and PDF analysis, and foreign-language records; capstone mock case.

Three-part series

3 × 90 min

Session 1 foundations and guardrails; Session 2 prompting and stations; Session 3 verification and workflow, with homework between.

Reserve your seat

Ready to put AI to work, safely?

Half a day, hands-on, at your own laptop. Bring any AI assistant and leave with a tested prompt library and a draft firm AI-use policy.

Register now