LLaManchaAI Enablement
Checking access…
← All modules
Module 1Audience: all

AI-assisted work fundamentals

Covers what AI is good at, where it fails, and why human accountability still owns the outcome.

Outcomes

What you will be able to do

  • Separate useful AI assistance from risky delegation.
  • Recognize hallucination and overconfidence patterns.
  • Pick the right tool and tier for a task's data sensitivity.
  • Place a task on a Reversibility × Sensitivity grid.
Completion check

How this module is approved

Submit one work task AI can help with now and one task AI should not touch yet. Place each on the Reversibility × Sensitivity grid and explain why.

Pass criteria

  • • Includes one safe task with quadrant
  • • Includes one unsafe/not-yet task with quadrant
  • • Explains the risk difference
  • • Names the human review gate
Loading personalization context…
Lesson35–45 minutes self-guided, or 55–70 minutes with discussion

What you should take away

By the end of this module, participants should be able to decide whether AI is a good assistant for a specific work task, explain what could go wrong, and choose a low-risk first use case that still saves real time.

Part 1

AI is assistance, not delegation

For workplace adoption, the most useful mental model is not 'AI does my job.' It is 'AI helps me prepare, compare, draft, explain, organize, and check parts of my work faster.' The human remains accountable for the facts, decisions, tone, quality, and consequences.

  • Good assistance: drafting a first version, summarizing non-sensitive notes, creating checklists, explaining unfamiliar concepts, comparing options, or suggesting test cases.
  • Risky delegation: asking AI to make final decisions, approve work, interpret sensitive records, produce client-facing conclusions without review, or operate on confidential data in an unapproved tool.
  • The question is not 'Can AI do this?' The better question is 'Can AI safely assist this, and what must a human verify?'

Quick check · <30 sec

Reframe 'Can AI do this?' into the better question.

Show answer
'Can AI safely assist this, and what must a human verify?' — the better question keeps accountability with the person and forces a verification step.

Part 2

Where AI is usually strong

Current AI tools are strongest when the task has clear context, a reviewable output, and a human who can judge whether the result is useful. They are especially helpful when the first draft or first structure is the bottleneck.

  • Turning messy notes into a clean summary or action list.
  • Rewriting a technical explanation for a different audience.
  • Generating first-pass documentation, FAQs, training outlines, or meeting agendas.
  • Brainstorming options, risks, test cases, edge cases, or implementation steps.
  • Explaining code, logs, policy language, or unfamiliar domain concepts at a high level.

Quick check · <30 sec

Which task plays most to AI's strengths?

  • A. Deciding who to promote
  • B. Turning your own messy meeting notes into a clean action list
  • C. Confirming today's production incident status
  • D. Approving a contract
Show answer
Clear context, reviewable output, low stakes — the messy-notes-to-action-list task. The others need current facts, authority, or accountability AI does not have.

Part 3

Tooling reality: which AI, and which tier

Treating 'AI' as one thing is the most common 2026 mistake. The tools differ, and — more importantly for safety — the same tool has a consumer tier and an enterprise tier with very different data handling. The practical rule for this module: the tool and tier you may use is decided by the data's sensitivity and your company's approvals, not by which model is 'best'.

  • General-purpose assistants (e.g., Claude, ChatGPT, Gemini, Copilot) overlap heavily for everyday drafting and summarizing; differences matter more at the edges than for a status update.
  • Consumer vs enterprise tier is the safety-relevant distinction: enterprise/work tiers typically come with data-processing terms and no training on your input; consumer tiers may not. Module 3 covers this in depth.
  • Verify-then-use: documented behavior changes over time and by plan — confirm your company's approved tools and tiers rather than assuming.
  • Choose by: data sensitivity → approved tool/tier → capability needed → cost. In that order.

Quick check · <30 sec

What primarily decides which AI tool/tier you may use for a task?

  • A. Which model benchmarks highest
  • B. Personal preference
  • C. The data's sensitivity and your company's approvals
  • D. Whichever is cheapest
Show answer
Capability and cost are tie-breakers. Data sensitivity and approval status are the gate.

Part 4

Where AI commonly fails — and how you would know

AI can sound confident while being wrong. It may invent details, flatten nuance, miss policy boundaries, produce plausible but broken code, or give advice that ignores your company's risk tolerance. The more specialized, current, confidential, regulated, or high-stakes the task, the more careful the review. The skill is not memorizing failure types — it is asking, every time, 'how would I know if this were wrong?'

  • Hallucination: the model fabricates facts, citations, policies, commands, or requirements.
  • Overgeneralization: the answer sounds reasonable but ignores your actual context.
  • False completeness: the output looks finished but skipped important edge cases or constraints.
  • Data leakage risk: the user pasted information into a tool that was not approved for that data.
  • Automation bias: people accept the AI answer because it sounds polished, not because it was verified.
Activity · ~5 minspot hallucination

Three sanitized AI answers to the same prompt: 'Summarize the key points of our standard internal status-update format.' Two are reasonable; one fabricates a confident specific. Find the fabrication and say how you knew.

  • Answer A: 'A status update covers progress, risks, and next steps, kept brief and audience-appropriate.'
  • Answer B: 'Per the ISO 19ablestatus standard, updates must contain exactly seven fields including a RAG code and a mandatory 24-hour SLA clause.'
  • Answer C: 'Typically: what got done, what is at risk, what is next, and any decisions needed — exact format varies by team.'

Your task

Identify the fabricated answer and name the two tells that gave it away.

Show a hint
Confident, oddly specific, and citing a standard you cannot verify is the classic hallucination signature.
Compare with a strong answer
B is fabricated. Tells: (1) it invents a precise authority ('ISO 19ablestatus') that does not exist and you cannot verify; (2) false specificity — 'exactly seven fields', a 'mandatory 24-hour SLA clause' — precision with no source. A and C hedge appropriately; B asserts. The check that catches it: 'how would I verify this claim?' — you cannot, so you do not trust it.

Why this matters: People rarely see a hallucination demonstrated; they only read that it happens. Spotting one builds the reflex the lesson is about.

Quick check · <30 sec

What single question best exposes most of these failure modes?

  • A. Is this well written?
  • B. How would I know if this were wrong?
  • C. Is this fast?
  • D. Did it answer quickly?
Show answer
'How would I know if this were wrong?' forces a verification path and surfaces hallucination, false completeness, and overgeneralization at once.

Part 5

Give it the document, don't try to 'train' it

A frequent misconception is that to get AI help with your domain you must somehow train the model on company knowledge. For almost all workplace use the dominant 2026 pattern is the opposite and far simpler: put the right sanitized information in front of the model at the moment you ask. Retrieval beats training for everyday work — provide the relevant approved document or context, ask your question against it, and the answer is grounded and checkable.

  • Bring the source: paste or attach the relevant sanitized text and ask the model to answer only from it.
  • Ask it to cite which part it used, and to say what is not covered by the source.
  • This is the non-technical idea behind retrieval-augmented generation (RAG): the answer is anchored to a document you control.
  • It also contains risk: the document you bring must itself be approved for that tool and tier.

Quick check · <30 sec

For everyday workplace tasks, the dominant pattern is to…

  • A. Train a custom model on company data
  • B. Give the model the right sanitized document and ask against it
  • C. Never give the model any context
  • D. Use only memorized knowledge
Show answer
Retrieval — bring the approved source and ask against it — is simpler, grounded, and checkable. Training is rarely the right tool for everyday work.

Part 6

The Reversibility × Sensitivity grid

Earlier programs taught a binary: 'safe' or 'not yet'. Reality has gradients. A better tool is a 2×2: how reversible is a mistake, and how sensitive is the data? It tells you not just whether to proceed but how heavy the review must be.

  • Low sensitivity + reversible → green: proceed, light self-review (e.g., draft your own status update).
  • Low sensitivity + irreversible → amber: proceed, but verify before the irreversible step (e.g., a message that goes to a wide audience).
  • High sensitivity + reversible → amber: sanitize or use an approved tool first (e.g., summarizing notes that contain identifiers).
  • High sensitivity + irreversible → red: stop; this needs approval, expertise, or an approved system (e.g., anything customer- or person-deciding).
Activity · ~6 minclassify scenario

Place four sanitized tasks on the Reversibility × Sensitivity grid and state the review weight each needs.

  • Tasks: (1) draft your own weekly status (2) AI-drafted message to all customers (3) summarize notes containing names (4) AI sets a customer's refund amount
  • Axes: Reversible? · Sensitive data?

Your task

Assign each task a quadrant (green/amber/red) and one sentence on the review it needs.

Show a hint
Two of these are amber for different reasons — one for irreversibility, one for sensitivity.
Compare with a strong answer
1 = green (low sensitivity, reversible): self-review before sending. 2 = amber (low sensitivity, irreversible at scale): human verify before it goes out. 3 = amber (high sensitivity, reversible): sanitize or use an approved tool first. 4 = red (high sensitivity, irreversible, decides about a person): stop — needs an approved system and human authority.

Why this matters: The grid replaces a brittle yes/no with a decision that also tells you how much review to apply — which is the real workplace question.

Quick check · <30 sec

A high-sensitivity, irreversible task lands in which quadrant?

  • A. Green — proceed
  • B. Amber — verify first
  • C. Amber — sanitize first
  • D. Red — stop
Show answer
High sensitivity plus irreversibility is red: stop and route to approval, expertise, or an approved system.

Part 7

Choosing your first use case

Pick something boring enough to be safe and annoying enough to be worth improving. The best early wins are not moonshots. They are small repeatable workflows where AI removes friction and the employee still controls the final result — squarely in the green quadrant.

  • Good first use case: draft a weekly status update from sanitized bullet points, then edit before sending.
  • Good first use case: ask AI to create a checklist for reviewing a document, pull request, or project plan.
  • Good first use case: turn a technical note into a plain-language explanation for a stakeholder.
  • Not a good first use case: ask AI to make a hiring, medical, legal, financial, security, or customer-impacting decision.

Quick check · <30 sec

Why is 'draft a weekly status from sanitized bullets' a strong first use case?

Show answer
It is green-quadrant: low sensitivity and reversible, frequent enough to matter, and you can easily tell if the output is wrong before it leaves you.

End-of-module quick check

Five short retrieval questions. Answer from memory first, then reveal each explanation.

  1. 1. The better question before using AI on a task is…

    • A. Can AI do this?
    • B. Can AI safely assist this, and what must a human verify?
    • C. Is AI fast enough?
    • D. Which model is best?
    Show answer
    Accountability stays with the person; the question must include the verification step.

  2. 2. What primarily decides which AI tool/tier you may use?

    • A. Benchmarks
    • B. Data sensitivity and company approvals
    • C. Habit
    • D. Cost alone
    Show answer
    Sensitivity and approval gate the choice; capability and cost are tie-breakers.

  3. 3. The classic hallucination signature is…

    • A. Hedged, vague answers
    • B. Confident, oddly specific claims citing an authority you cannot verify
    • C. Short answers
    • D. Answers that ask clarifying questions
    Show answer
    Unverifiable precision asserted with confidence is the tell — always ask how you would verify it.

  4. 4. True or false: to get domain help you usually must train a model on company data.

    • A. True
    • B. False
    Show answer
    False. For everyday work, bring the approved sanitized document and ask against it — retrieval beats training.

  5. 5. Which quadrant means 'stop — needs approval or an approved system'?

    Show answer
    High sensitivity + irreversible = red. Stop and route to approval, expertise, or an approved system.

Further reading

Worked examples by role

Operations manager

Assist vs. delegate: operations manager

Safe assist: draft a weekly status from sanitized bullets — green quadrant, reviewable, I correct figures before posting. Not yet: AI independently sets incident severity — red, high-stakes and needs current operational context plus the incident lead's authority.

Sales coordinator

Assist vs. delegate: sales coordinator

Safe assist: first-draft a follow-up email from sanitized account context — amber, I fact-check competitor/pricing claims before the owner sends. Not yet: AI decides a customer's discount — red, decides about a customer and is hard to reverse.

Software developer

Assist vs. delegate: software developer

Safe assist: first-pass release note from sanitized implementation bullets — green, easy to correct before publishing. Not yet: AI approves a production deployment — red, irreversible and needs human accountability and formal gates.

Completion artifact

Submit one task where AI can safely assist your work now and one AI should not handle yet. Place each on the Reversibility × Sensitivity grid and explain the risk difference and the human review gate.

ExercisePilot-ready artifact

Classify two real-but-sanitized tasks: one where AI can assist now and one where AI should not touch the work yet or should not handle it independently.

Participant template

  • Safe assist task:
  • Quadrant (green/amber/red) + why:
  • Why AI can help:
  • What I would verify (how I'd know it's wrong):
  • Not-yet / do-not-use task:
  • Quadrant + risk reason:
  • Human review or approval gate:

Example submission

Safe assist task: draft a status update from sanitized project bullets — green (low sensitivity, reversible). Why AI can help: output is low sensitivity and easy to edit. Verify: dates, owners, risks, tone — I'd catch a wrong figure by tracing it to a source bullet. Not-yet task: approve a customer-impacting outage communication — red (high impact, irreversible). Risk reason: high business impact and current operational facts matter. Review gate: incident lead and communications owner approve before sending.

Role-flavored variants

Same exercise, framed for different roles. Use the one closest to your work.

Front office coordinator

Pick an intake/scheduling task as the safe one and a judgment call as the not-yet. Name the quadrant and how you would detect a wrong output before it reaches anyone.

See a sample submission
Front office submission: Safe: draft confirmation wording from sanitized details — green, reversible, low sensitivity; verify names/times against the source before sending. Not-yet: AI decides who is seen first — red, decides about a person and is high-stakes. Gate: front-desk lead reviews; clinical questions go to a nurse.

Training manager

Pick a drafting task as safe and an evaluation decision as not-yet. The verification step must include a bias check.

See a sample submission
Training manager submission: Safe: draft comprehension questions from objectives — amber (low sensitivity, but published), verify for bias and reading level before use. Not-yet: AI scores learner submissions — red, decides about people. Gate: I review and a second reviewer calibrates flagged items.

Learner checklist

Use this as a final check before submitting. Program leads use a separate review guide when they approve or coach submissions.

  • • Includes one safe task with quadrant
  • • Includes one unsafe/not-yet task with quadrant
  • • Explains the risk difference
  • • Names the human review gate
Loading your module status…

Previous module

← Orientation and kickoff

Review the prior step in the path.

Next module

Prompting with context and constraints →

Keep momentum with the next completion check.