Teach on Udemy

Turn what you know into an opportunity and reach millions around the world.

Learn More

Your cart is empty.

Keep shopping

AI Safety at Work: Privacy, Hallucinations & AI Agents

Use AI safely at work: protect confidential data, verify outputs, control AI agents and build your AI safety playbook

Created byStuart Wesselby

Last updated 10/2026

English

What you'll learn

Classify any AI use case by business consequence and choose a proportionate level of review using the PAUSE Check
Decide what workplace information can be shared with an AI tool, what must be minimised or anonymised, and what must never be entered
Verify AI-generated citations, numbers and claims, and catch hallucinations before they reach a customer or a decision
Design permission and approval boundaries so AI agents get only the access and autonomy they need
Recognise prompt injection in documents, webpages and email, and apply layered controls that limit the damage
Build an Everyday AI Safety Playbook with approved uses, data rules, owners, an incident route and a 30-day action plan

Course content

10 sections • 44 lectures • 4h 53m total length

Introduction3:22

Welcome: How to Use AI Safely at Work2:55
Establish the learner transformation, course boundaries and applied project
What AI Safety at Work Covers (and What It Doesn’t)5:43
Define the course's practical scope; distinguish it from technical model safety and legal certification
Rate AI Risk by Business Consequence: Low, Medium, High6:10
Compare low-, medium- and high-consequence uses without implying universal risk categories
The PAUSE Check: A 5-Step AI Safety Checklist7:13
Introduce the five questions and show how they connect across the course

Where Your Data Goes When You Use AI Tools8:19
Show where input, files, context, logs and connected sources may enter an AI workflow
Classify Business Data Before You Paste It into AI9:25
Apply public, internal, confidential, personal and restricted categories without replacing company policy
Anonymise, Redact or Minimise Data in AI Prompts8:35
Choose the least information needed and recognise when alteration is insufficient
Check If an AI Tool Is Approved for Work Data8:38
Check organisational approval, account type, access, retention and contractual context without making vendor guarantees
Scenario: Summarise Customer Data Safely with AI9:59
Work through a request to summarise customer data and model a safe alternative

AI Hallucinations: Why Confident Answers Can Be Wrong9:20
Explain hallucination/confabulation in business language without suggesting all outputs are equally unreliable
Match Verification to Business Risk9:05
Build a verification ladder from light review to authoritative or specialist review
Verify Citations, Sources and Dates6:25
Check whether sources exist, support the claim and are sufficiently current
Check Numbers, Policies and Missing Context6:56
Test calculations, assumptions, omissions and version-sensitive material
Make Human Review More Than a Rubber Stamp6:00
Counter automation bias by defining reviewer competence, evidence, authority and challenge

How AI Agents Turn Answers into Actions6:08
Distinguish generating text from reading systems, changing records, contacting people or spending money
Separate Read, Draft, Recommend and Act Permissions6:55
Introduce four levels of delegated capability and their different consequences
Apply Least Privilege to AI Agents6:49
Translate excessive-agency controls into ordinary business decisions
Require Approval Before Consequential Actions6:27
Design confirmation gates for external communication, record changes, payments and other consequential actions
Scenario: Secure a Connected Email Assistant6:41
Redesign an over-permissioned email workflow while preserving useful automation

Prompt Injection in Plain English6:56
Explain direct and indirect prompt injection without teaching offensive exploitation
How Documents, Webpages and Email Can Manipulate AI6:48
Show how apparently ordinary content can become an instruction pathway
Why Prompts Cannot Be Security Boundaries6:52
Correct the misconception that prompt wording can eliminate injection or safely hold secrets
Layer Controls Against Prompt Injection7:06
Apply trusted-source boundaries, constrained tools, allowlists, deterministic checks and human confirmation
Scenario: Detect Injection in a Supplier Document6:39
Identify the untrusted instruction, contain exposure and choose a safe next step

Shadow AI and Unapproved Workplace Tools6:45
Connect unapproved accounts, unmanaged sharing and absent ownership to business risk
Reduce Bias and Automation Bias6:38
Check who may be disadvantaged, which evidence is missing and where judgement must remain human
Check Copyright, Intellectual Property and Provenance6:40
Teach origin, permission, licence and similarity checks while avoiding legal conclusions
Assess AI Vendors, Connectors and Supply-Chain Risk6:43
Review data access, dependencies, ownership, updates, monitoring and exit routes
Complete an AI Use-Case Risk Assessment6:53
Combine privacy, reliability, bias, IP, supplier and operational questions without turning the process into bureaucracy

Recognise AI Phishing, Deepfakes and Voice Clones6:04
Explain AI-assisted phishing, impersonation, deepfakes and voice cloning without fear-based claims
Verify Urgent Requests Through an Independent Channel6:07
Use process controls rather than relying on a person's ability to detect synthetic media
Stop and Contain an AI Safety Incident6:05
Define immediate actions after a suspected disclosure, erroneous action or manipulation
Report, Escalate and Learn from AI Incidents6:14
Establish ownership, communication, records and a route to specialist help
Scenario: Respond to Executive Impersonation6:10
Respond to a voice-and-message payment diversion attempt while maintaining calm and evidence

Write Clear Workplace AI Safety Rules6:00
Replace vague “use AI responsibly” language with observable decisions and behaviours
Define Approved, Conditional and Prohibited Uses5:56
Build a simple use matrix that can evolve without blocking all experimentation
Assign Owners, Reviewers and Escalation Routes5:57
Define accountability proportionate to organisational size
Keep Proportionate Decision and Incident Records5:47
Decide which use cases, approvals, incidents and changes require evidence
Review and Improve Workplace AI Controls5:50
Set review dates, update triggers and short reinforcement activities

Capstone: Assess a Complete AI Workflow5:59
Explain the fictional scenario, required decisions, deliverables and safety-critical conditions
Apply the PAUSE Check to a Business Scenario6:18
Model the reasoning process using a different scenario from the learner's submission
Test Your AI Safety Playbook5:48
Apply the rubric, safety-critical checks and a challenge scenario
Create a 30-Day AI Safety Action Plan5:46
Turn the Playbook into owners, conversations, quick wins and scheduled review

Requirements

This course is for business owners, managers and employees who use or oversee generative AI and want practical guardrails without the jargon. You need no coding, security or legal background. Basic experience with an AI assistant is enough.

Description

“This course contains the use of artificial intelligence.”

AI safety at work comes down to a few everyday decisions: what to paste into a chatbot, whether to trust the answer, and how much you let an AI assistant do on its own.

Most teams handle these by instinct, vague warnings or whatever the product does by default. This course gives you a simple, repeatable way to make those calls with confidence.

By the end, you'll be able to explain and defend your AI decisions in ordinary business work. You'll classify information before sharing it, check outputs in proportion to what's at stake, limit what AI tools can access and do, spot manipulation and deepfake fraud, and know when to stop and escalate. You'll also leave with a finished Everyday AI Safety Playbook for your business or team.

The PAUSE Check: one method used throughout the course

Every section builds on the PAUSE Check, a five-question decision method created for this course:

Purpose and potential consequence: what could happen if the AI is wrong or misused?
Approved tool and approved data: is the tool authorised, and is this information suitable to share?
User authority and AI permissions: does the AI have only the access it needs?
Source-check the output: what evidence or human judgement do you need before you use it?
Escalate, evidence and evaluate: when should work stop, and who needs to know?

What's covered

AI risk by business consequence: tell low-, medium- and high-stakes uses apart and match your controls to each
AI data privacy: follow where data goes, classify it, and minimise, redact or anonymise inputs
AI hallucinations and verification: check sources, dates, numbers and missing context, and make human review a real check rather than a rubber stamp
AI agents and permissions: separate read, draft, recommend and act permissions, apply least privilege and add approval gates before consequential actions
Prompt injection: see how untrusted documents, webpages and emails can manipulate AI, and why prompt wording alone can't be a security boundary
Wider AI risk management: shadow AI, bias, copyright and provenance, and vendor and connector risk
Deepfakes, AI fraud and incident response: verify urgent requests through an independent channel, contain incidents, then report and learn from them
Workplace AI rules and roles: write clear rules, define approved, conditional and prohibited uses, and assign owners and reviewers

Who this course is for:

Small business owners whose staff already use ChatGPT, Copilot or similar tools and who want clear rules before something goes wrong
Managers and team leaders who are asked to approve AI use but don't have a reliable way to judge what is safe
Employees in HR, marketing, finance, operations or customer service who handle customer or company data and want to use AI without exposing it
Operations, compliance and governance leads who need a practical AI policy their team will follow, not one that just sits in a folder
Anyone setting up connected AI assistants or agents that can read email, change records or act on their behalf
Not for security engineers or developers looking for technical defences or penetration testing. This course is non-technical and needs no coding

AI Safety at Work: Privacy, Hallucinations & AI Agents

What you'll learn

Explore related topics

Course content

Introduction1 lecture • 3min

Section 1: AI Safety at Work: Risk Levels and the PAUSE Check4 lectures • 22min

Section 2 - AI Data Privacy and Confidential Information5 lectures • 45min

Section 3 - AI Hallucinations, Verification and Human Review5 lectures • 38min

Section 4 - AI Agents, Permissions and Approval Boundaries5 lectures • 33min

Section 5 - Prompt Injection and Untrusted Content5 lectures • 34min

Section 6 - Shadow AI, Bias, Copyright and Vendor Risk5 lectures • 34min

Section 7 - Deepfakes, AI Fraud and Incident Response5 lectures • 31min

Section 8 - Workplace AI Rules, Roles and Risk Controls5 lectures • 30min

Section 9 - Create Your Everyday AI Safety Playbook4 lectures • 24min

Requirements

Description

Who this course is for: