The Human EdgeUsing and delivering AI with judgement
Same day. Better name.
Everyone in this room has the same tools. Your judgement is the edge.
Dr Michael Borck · michael.borck@curtin.edu.au
Course site
Class chat · today only
The day
Four sessions. Four tools. Your own work.
01 · 9:00
Judge
How much should I trust this?
Trust tool
A ten-second habit for every output
02 · 11:00
Brief
How do I ask for what I need?
RTCF
Your own prompt library
03 · 1:15
Check
How do I know it's right?
Check loop
A verified analysis, and your decision rules
04 · 3:00
Deliver
Will it work every week?
Opportunity Finder
A delivery plan for one workflow, and an action plan
Morning tea 10:30 · Lunch 12:30 · Afternoon tea 2:30 · Close 4:30The rhythm: a few slides, then you do it.
First, your bearings
AI is a bigger pie. Today is one slice.
What it actually does
AI gives you the convincing average.
It predicts the most plausible next words. Fluent, well-shaped, usually about right.
Your alarm
AI's mistakes don't look like mistakes.
You spot weak work by its signals. AI's wrong answers don't carry them.
The signals you rely on
A person's weak draft
AI's wrong answer
Typos and slips
Gaps, TODOs, missing pieces
Hedging: "I think", "probably"
Uneven formatting, a rushed ending
Your alarm
Rings
Stays quiet
Both are wrong. Only one of them looks it.
The second trap
Sharp where you know. Trusting where you don't.
In your own field
You catch its errors straight away.
"That's not how our contracts work." "That figure's from last year." You know enough to see it.
Outside it
The market, the law, the numbers.
Same tool, same error rate. But it reads as fact, because you can't see the mistakes.
The defence: be most sceptical where you know least."Gell-Mann amnesia": Michael Crichton, 2002, after the physicist Murray Gell-Mann
Exercise 125 MIN · SOLO, THEN TABLES
Your pass first. Then AI's.
Sketch your own plan. Five minutes, no AI.Bullets in the workbook. Rough is the point.
Give AI the same task, unedited.Two tools if you have them. The class chat has two.
Compare side by side.Where did you beat it? Where did it beat you? What did it get confidently wrong?
The task
I manage a team of 8 in a mid-sized company. One of my team, a strong performer for three years, has been late and missing deadlines for the past month. Others are starting to notice. Write me a plan to address this.
OUTPUTWhere you beat it, where it beat you, and one thing it got confidently wrong.
Debrief
Where did you beat it? That's your value. Name it.
Where did it beat you?
Did either tool ask what you already know?
No tool was best. Every answer needed you.
Tool 1 · The trust tool
Two questions before you act on it.
Question 1 · Precision
How exact must it be?
Average
The ballpark is fine. A draft, a summary, ideas to react to.
Precise
One right answer. A figure, a name, a date, a citation, a clause.
Question 2 · Consequences
What happens if it's wrong?
Small
Cheap to fix, easy to reverse, nobody harmed.
Large
Money, safety, reputation, someone's job or wellbeing.
Two questions, ten seconds, four corners.
Tool 1 · The trust tool
Four corners. Four levels of checking.
Tool 1 · The trust tool
Check from outside the tool.
Better models move the corners. They never remove the check.
Exercise 222 MIN · ROOM, THEN SOLO
Place it. Then argue about it.
Blind vote. I read a scenario. Decide in silence, then vote on "three".Five scenarios. A split room is the point.
Play the game. Choose your domain and place all ten cards.Read the explanation on any card you miss.
Place three tasks from your own week in your workbook.Swap with a neighbour. Where do you disagree, and why?
The game
trusttool.borck.education
Free, no login, works on a phone.
How to vote
ClassPulse on your phone or laptop: scan the code on screen. No signal? Fingers, 1 to 4:
1 · A Light review 2 · B Load-bearing bits 3 · C Before you ship 4 · D Human in the loop
OUTPUTThree of your own tasks, placed, with the check each one needs.
Debrief
Where did the room split?
Was the argument about precision, or about stakes?
Which of your own tasks was hardest to place?
Stakes usually depend on who sees it next.
Before the break
Three rules before you paste.
01
Classify it, then pick the tool.
Public, internal, confidential or regulated? Anything internal or above goes only into tools your organisation has approved for it.
02
Leave out what identifies people.
Names, health details, client data, anything under NDA. Describe the shape of the data instead of pasting it.
03
Know what it remembers.
Saved memories, inferred memories and chat history are kept separately. Deleting a chat may not delete what it learned. Use a temporary chat for anything sensitive.
Check yours now, two minutes. Open your tool's data settings. Is it training on what you type? Is memory on?
And: your name goes on the output, not the tool's.
Session 2 · Brief
Brief it like a capable new colleague.
No magic words. Smart, fast, and brand new to your job.
What you know
Who it's for, and what they care about
What happened last time
The politics
What good looks like here
What it knows
How emails, reports and plans usually go
A great deal of public information
Nothing about your client, your team or your standards
Tool 2 · RTCF
Four parts of a good brief.
R
Role
Who should it write as?
A senior HR adviser in a mid-sized Australian firm
T
Task
What exactly, and how much?
Draft a reply. Under 200 words.
C
Context
What does it need to know?
18-month client. Their board shifted priorities. I want a meeting, not a rewrite.
Does most of the work
F
Format
What should it look like?
Email. Acknowledge, explain, propose, close.
Not every prompt needs all four. Anything that matters does.
Tool 2 · RTCF
One minute of briefing. Thirty minutes saved.
Before
Write an email about the project delay.
Generic, polite, and unusable without a rewrite.
After
ROLE Senior project manager. Plain-spoken and warm. TASK Write to the client about a two-week delivery delay. CONTEXT A vendor's API integration ran late. The client values transparency and the relationship is good. New date: 15 March. FORMAT Under 200 words. Acknowledge, explain, new date, recommit.
Close to something you'd send.
Exercise 325 MIN · SOLO
Brief it. Twice.
Pick a real task from your week.Confidential? A scenario card: course site, Exercises, at the bottom.
Round 1: write it the way you normally would. Run it. Score it 1 to 5.Score on one question: how close is this to usable?
Round 2: rewrite it with RTCF. Run it in a fresh chat. Score again.
Note which part moved the score most.
RTCF template
ROLE You are a ... TASK Write / summarise / compare ... (how much?) CONTEXT The audience is ... The situation is ... Good looks like ... FORMAT Length ... Structure ... Tone ...
Prompt builder on the course site.
OUTPUTA before and after for one real task, with scores.
Debrief
Who jumped two points or more?
Which part made the difference?
Did anyone get a worse result in Round 2?
It is usually Context. And a worse result usually means the context was wrong, not missing.
Tool 2 · RTCF upgrades
Weak models can't infer. So leave nothing to infer.
Four upgrades. They help smaller, older and free-tier models most, and the best models still do better with them.
+ Example
Show it one you liked.
Match the length and tone of this example: [paste]
+ Done
Say what done looks like.
Done means under 200 words, figures only from my notes, and anything you're unsure of marked [CHECK].
+ Questions
Let it ask first.
Before you start, ask me up to five questions.
+ Steps
Split the job.
Give me an outline only. We'll draft one section at a time.
Agreeable is not honest
Push back. Three ways.
Models are tuned to please. Disagree with no evidence and many will fold.
01 · Set a guardrail
Keep it honest.
Only change your assessment if I give you a new fact. Name the fact that changed it.
02 · Devil's advocate
Make it attack.
Don't flatter me. Give me three reasons this fails, and name the assumption it all rests on.
03 · Board of directors
Hear a room, not a voice.
Respond as a board: a CFO, a sceptical customer and a frontline manager. One paragraph each, then where they disagree.
You decide what survives. And for an independent read, start a fresh chat: long chats drift towards your view.
Exercise 425 MIN · SOLO
Upgrade it. Then save it.
Take your best Round 2 prompt. Add one upgrade and run it.Example, done, questions or steps. Same task or scenario card as Exercise 3.
Push back with no new evidence. Does it fold or hold?Then try the guardrail, devil's advocate or the board.
Save three prompts to your library.Name, when to use it, the prompt with [blanks], and the check it needs.
Stretch: the voice pass.Have it rewrite its output in your voice. Steps in the workbook.
A library entry
NAME Friday status update USE Raw notes to steering committee PROMPT You are a project manager... Context: [paste notes]... Format: RAG status, wins, risks table, next week. Under 250 words. CHECK Verify before you ship: every date and figure.
OUTPUTYour prompt library: three prompts you'll reuse this month.
Debrief
Which upgrade helped most?
Whose model folded? Which push-back move helped?
What check did you attach to each prompt?
A saved prompt with its check attached is the first draft of a skill. More on that next.
Beyond chat
Agents run the loop. You set the standard.
Session 3 · Check · Tool 3
Fast analysis. Slow trust.
The check loop. Most of the value is in step three.
Tool 3 · The check loop
Five ways a smooth answer goes wrong.
"Are you sure?" doesn't count. The check comes from outside the tool.
01
Miscounts
Especially after filtering or grouping.
02
Correlation sold as cause
"Sunday tickets are slower, so Sundays cause complaints."
03
Silent gaps
Skips missing or messy rows without saying so.
04
Numbers that can't all be true
Shares that add to 112%. An average above the maximum.
05
Invented detail
A data point, quote or source that isn't in the file.
The fix · VET it
Verify. Explain. Test.
Verify the source yourself. Explain it back in your own words. Test the edges: when would it be wrong?
Check for skew too
Bias hasn't gone. It has gone quiet.
Models are trained hard against the obvious stereotypes. The skew lives in the small choices.
Try it · 3 minutes
Write a one-paragraph scene of a respected senior manager on a busy Monday.
New chat: the same, for a respected administrative officer.
Read them side by side
What is each one called? Who is warm, who is capable? Who holds authority in the scene?
That picture does quiet work every time you draft a position description, a reference or a performance review.
The generic output inherits the generic average. Noticing the skew is your job.
Exercise 535 MIN · SOLO OR PAIRS, THEN BLIND VOTE
Karen needs answers by Monday.
Frame. Write Karen's three questions down. Start a fresh chat.
Explore. Upload the CSV and ask it to describe the data first.Rows, columns, date range. Does it match what you were told?
Shape a first brief for Karen, under 300 words.Keep it in the workbook. Don't send it yet.
Verify. Three claims about this data are doing the rounds.Check them against the data, fix your brief, and end it with one line: What I checked.
From Karen, customer service manager, RetailFlow
1. What are the top three drivers of customer dissatisfaction?
2. When do our response times blow out?
3. What should I do about it?
retailflow-customer-service-data.csv 260 tickets · July to December 2024 · on the course site
OUTPUTA brief Karen could act on Monday, saying what you checked.
Exercise 5 · Verify8 MIN · PAIRS, THEN BLIND VOTE
Three claims. At least one is wrong.
Claim A
"Email is roughly three times slower than Live Chat."
Claim B
"Account Issues has the lowest satisfaction of any category, which makes sense, because account problems are the most complex to resolve."
Claim C
"Sunday response times are more than double the overall average."
Check outside the tool: the data, the source, your own sum. Is any of them in your brief?VerifiedOverstatedRefuted
Decide like me · SIM
Teach it your rules. Keep the call.
One question picks the method: could you tell a new starter, out loud, why you decide the way you do?
S · Say
Yes? Talk it through.
Most people trim the exceptions when they type. Out loud, they stay in. Then let it ask you up to five follow-up questions.
I · Interview
"It depends"? Let it ask.
It questions you in turn until every "it depends" is an if-then rule. Each branch is a call you've been making without writing it down.
M · Mine
On file somewhere? Mine it.
Past requests and what you decided. It works out the rules you were following, including the ones you never wrote down.
All three end the same way: a page of if-then rules. Read it like a brief. Would these rules make the calls you make?
Save · test · fix
Don't take its word. Take the test.
Save it
Same chat: "Write these rules up so I can reuse them."
Rules, not stories: take out the names, clients and figures. Save it where your tool keeps standing instructions. Agent tools call a saved brief like this a skill, as on the agents slide before lunch.
Then the blind test
Three past cases. Inputs only. Your calls written first.
It makes each call and names the rule it used. A mismatch is a rule you've never put into words. Add it, then re-test in a new chat: the chat you built it in already knows your answers.
Your past decisions are the check from outside the tool. It can't agree with an answer it hasn't seen.
Exercise 624 MIN · SOLO, THEN TABLES
Teach it to decide like you.
Pick one decision you make over and over.Small and repeatable. Can't use real work? A decision card: Exercises page, at the bottom.
Get the rules out.Talk it through, or let it interview you. Turn every "it depends" into an if-then.
Test it blind on three past cases.Your calls written first, inputs only. Compare. Fix one mismatch.
Debrief at your table.Where did it first disagree? Which rule was never written anywhere?
The interview prompt
Help me write down the rules I use for one decision: [the decision]. Question me one at a time, choosing each question from my last answer. Each time I say "it depends", dig until it becomes an if-then rule. Stop after ten questions and list my rules, every branch included.
Describe your past cases. Don't paste anything confidential.
OUTPUTYour decision rules, tested blind, and a list of what stays yours.
Debrief
Where did it first disagree with you?
Which rule had never been written down anywhere?
What would you never let it decide, and why?
The mismatches show where your judgement lives.
Session 4 · Deliver · Opportunity Finder · 1 Spot
Four signals in your week.
Automate
You do it often, the same way.
Status reports. Sorting requests. Reformatting.
Assist
You read, gather and compare.
Vendor proposals. Meeting prep. Long documents.
Generate
You write from a blank page.
Client emails. Job ads. First drafts.
Analyse
You hunt for meaning in numbers.
Monthly KPIs. Survey comments. Budget variance.
Start with the task you do every week that bores you most.
Then prioritise: start with a quick win. Real time saved, cheap to check. The trust tool sets how hard you check each step.
Deliver · Why AI delivery is different
AI projects don't fail more. They fail differently.
You usually assume
With AI
So you
"Done" can be specified.
Outputs vary. There's a spread, not one right answer.
Decide what "good enough" means, up front.
A working demo means nearly there.
The demo is a trap: it shows the happy path.
Treat the demo as the start of the hard part.
The data is a known input.
The data is the uncertainty. You find out by working with it.
Plan for discovery, not slippage.
The build is the product.
Checking is the product.
Design the checking in, and ship it.
Capability is our edge.
Everyone has the same models.
Compete on your people's judgement.
Deliver · Plan it like a project
Design the human in. Don't bolt it on.
01 · Scope
One workflow, one owner, one measure.
Write down today's time and quality. That's your baseline.
02 · Checkpoints
A person at every amber or red step.
Someone who knows what right looks like, with a specific check.
03 · People
Who's affected, who agrees, who checks.
Tell them before you start, not after something breaks.
04 · Gate
Set the date and the criteria now.
What would make you scale it, pivot it or kill it?
A redesigned workflow is a small AI project. The same judgement scales to big ones.
Deliver · The 30-day gate
Scale. Pivot. Or kill.
Judge the workflow against the criteria you set before you started.
Scale
It met the measure with a light review.
Hand it over for good. Keep the check its corner calls for.
Pivot
Close, but the check costs too much.
Change the brief, the step, or who leads. Run another month.
Kill
It missed, or the errors mattered.
Keep it human. Revisit next quarter; the tools will have moved.
Killing is a good outcome. You found where the line is, cheaply.
Exercise 735 MIN · SOLO, THEN TABLES
Plan the delivery of one real workflow.
Map a quick win: a weekly task of 30 minutes or more.Five to eight steps, honest minutes. Can't use real work? Scenario card 3 or 6.
Assign each step: who leads, and its trust corner.
Design the checkpoints.Who checks each amber or red step, and how? Who needs to know?
Test one AI step live. Then set your 30-day gate.The test this plan must pass: what result means scale, pivot or kill?
Example · weekly team update
STEP MIN WHO LEADS CHECKPOINT
Gather updates 15 You lead -
Draft the summary 20 AI leads Me: light review
Calculate metrics 15 You lead Me: recount totals
Send to sponsor 5 Human only -Gate, 30 days: scale if two edits or fewer;kill if a figure is wrong twice
OUTPUTA one-page delivery plan you'll run this month.
Debrief
Where did you put a person, and why?
What's your kill criterion?
Who needs to know before you start?
Credible beats impressive. Nobody saves 90% in week one.
Deliver · Responsible by design
Australia's six essential practices. You've done five today.
The practice
Where you've already done it
1 · Decide who is accountable
A named owner for each workflow. Your name goes on the output.
2 · Understand impacts and plan
Who's affected? The quiet defaults: who could be treated unfairly?
3 · Measure and manage risks
Classify the data before you paste. Place every output on the trust tool.
4 · Share essential information
New today: tell people when AI drafted or decided something that affects them.
5 · Test and monitor
The check loop, VET, and your 30-day gate.
6 · Maintain human control
Human checkpoints at every amber or red step. Some work stays human only.
For depth on ethics, governance and risk: Ethics of artificial intelligence, Curtin Executive Education.Guidance for AI Adoption, Australian Government, October 2025
Why practice matters now
Fewer junior hires. Not mass replacement.
Australia · DEWR, July 2026
No broad AI-driven upheaval. Youth outcomes have mostly held up.
But jobs more exposed to AI automation have grown more slowly.
Australia · Jobs and Skills Australia, 2025
Some employers now hire more seniors and fewer juniors.