The S Curve
Tue, 22 SeptMelbourne · Singapore · New York
Book a call
The S Curvepowered by Collective Campus
// Guide

AI training playbook

A two-day sequence for AI literacy across a value chain: workflows, owners, and hours returned.

The job of the two days

This is a two-day AI literacy sequence on a hated workflow. It is not a slide titled Transformation. It is not a model zoo. It is not a licence dump.

Pick the workflow. Pair a domain lead with a builder. Write the guardrail page. Practice on live work. Leave a play. Measure hours returned at day 14 and day 30. Keep, kill, or constrain.

If a new CAIO, CHRO, or Head of L&D funds attendance instead of hours returned, they bought theatre with a new job title.

Collective Campus AI Training exists so a team takes hours back from work they already own, with guardrails a risk lead will sign. Steal the sequence. Rename it if legal needs a local noun.

Hours returned, not completions

Slideware wins because it is easy to govern. A 40-page "AI strategy" can be approved without changing a Tuesday. It names pillars, risks, and a centre of excellence. It does not name the team whose week will change.

Slideware loses the moment a CFO, a regulator, or a journalist asks "what did we stop doing?" If the answer is "we stood up a guild," you bought theatre with a new acronym.

Logins are not hours. Completions are not hours. Licences are not hours. Champion-network membership is not hours.

Hours come back when a domain lead and a builder sit on the same workflow and leave a play the rest of the desk will copy. Baseline the clock. Re-measure at day 14 and day 30. Publish the number, including when it did not move. Kill the play that did not pay.

CHRO-shaped owners should care because job design will change. If the hours come back and you fill them with more of the old work, you did not buy productivity. You bought capacity for more theatre. Put a rule on the returned hours: maker time, customer time, or a deleted room. Atlassian-style and Canva-style cultures already argue this. Banks and telcos can steal the argument.

What to report (and what to stop reporting)

Report:

  • Workflow name and owner
  • Baseline minutes (ugly is fine)
  • Day 14 minutes and day 30 minutes
  • Human-check minutes the guardrail requires (do not hide them)
  • Plays adopted by people who were not in the room
  • Plays killed

Stop reporting as success:

  • LMS completions
  • Licence counts
  • Slack channel joins
  • Town-hall attendance
  • Models deployed
  • "Awareness" scores

Canva does not measure success as "people opened the app." Neither should a bank. DBS-style programmes that work treat enablement as operations: a use case, an owner, a control, a metric. Copy the physics. Do not invent their numbers.

Pick the hated workflow

Start with a calendar, not a capability map. Where do senior people still paste, summarise, reformat, and chase? That is the value chain.

Workflow chooser (score in the room, then pick one)

Write three candidates. Score each 1 (weak) to 5 (strong):

  • There is a clock (a recurring block someone can point at)
  • A domain lead will sit for two days
  • A builder (analyst, engineer, ops designer, sharp BA) exists
  • Risk can imagine a guardrail (data that never goes in, human check, audit)
  • Failure is recoverable (not a production decision with no review)
  • The output is something a Tuesday standup can reuse

Pick the highest score that a stopper will sign. If all three fail the stopper, you do not have an AI training problem. You have a permission problem.

Good seams (examples, not a menu you must copy)

  • The weekly regulatory or performance pack that takes a manager an afternoon to collapse
  • The second-line customer email that is mostly identical and still written from scratch
  • A credit memo or decision explanation at a DBS-scale, NAB-scale, or HSBC Hong Kong-scale desk
  • A Telstra care or field explanation rebuilt from a blank page
  • Test boilerplate one squad recopies every sprint (Atlassian-shaped)
  • An internal brief that is reformatted every week (Canva-shaped ops)
  • A letter or notice a public servant rewrites (GDS, Services Australia, IRD)
  • Intake notes a legal operator re-keys (Allens-shaped)
  • An offer page that waits on an agency cycle (Infinitev-shaped, claims in the room)

Bad seams (do not start here)

  • "Knowledge management"
  • "Customer service" as a theme
  • "Coding" as a theme
  • Anything that needs a production integration on day one
  • Anything a model is not allowed to draft even with a human check
  • A bake-off between vendors

If you cannot point at a calendar block, you do not have a use case. You have a theme.

Guardrail page (one page, in the folder with the play)

Write this before the prompts. HSBC Hong Kong and DBS both live inside regimes where "we tried a prompt" is not a control. Telstra and any AU Authorised Deposit-taking Institution rhyme. MAS and HKMA conversations get easier when this page exists. Canva has customer-content risk of a different flavour. Public bodies have IG. Allens has matter risk. AGL has claims and hardship. Write the page.

Guardrail canvas

Approved tools. Name them. Shadow tools are a risk incident, not a culture win.

Data that never goes in. Customer identifiers, secrets, privileged matter content, unpublished market information, health data, credentials. Be boring and specific.

What the model may draft. Summaries, first-pass letters, checklists, reformats, options. Not a decision.

What a human must check. Facts, numbers, promises, legal or credit language, anything a customer or a regulator will read. Name the role that checks.

What the model is not allowed to decide. Credit. Hardship outcome. Clinical advice. Legal advice. Firing. Pricing exceptions. Anything your licence treats as a regulated decision.

Audit. What you keep (prompt pattern, output, checker, time). Where it lives. Who can see it.

Incident. What to do if data went somewhere it should not. Named owner.

Kill or constrain triggers. Hallucinated customer-facing facts. Skipped checks. Minutes that do not move by day 30. Stopper withdraws the mark.

Named stopper. Operational risk, IG, data, or conduct. They mark this page before day one.

Put this page in the same folder as the play. Not in a PDF six clicks away. HSBC Hong Kong-style programmes that work put the control next to the prompt. Steal that.

Domain lead plus builder

A centre of excellence can coach. It cannot own. If the CoE presents the demo and the desk was not in the room, you have a roadshow.

Domain lead. Owns the work and the risk. Can change how the desk runs next Tuesday. Sits both days. Examples: credit manager, care lead, field supervisor, partner's operations lead, service owner, product manager.

Builder. Owns the tooling path. Analyst, engineer, ops designer, knowledge lead, or a BA who can make a repeatable play. Sits both days. Does not vanish into a CoE after the photo.

Facilitator / L&D. Owns the sequence, the time-box, and the residue note. Does not own the workflow.

Stopper. Marks the guardrail. Sits the first morning and the last afternoon. Available on a channel in between.

CAIO or CDO (if they exist). Funds the two days and the approved tool. Does not turn the room into a keynote.

If you cannot name the domain lead and the builder, do not book the room. Hire or free those two people first.

Two-day run of show (live work only)

Real documents. Real tickets. Real messy folders. The first afternoon will be ugly. Good. Ugly is how you find the step the slide skipped.

Day 0 (sixty to ninety minutes)

  • Domain lead names the workflow and the clock (baseline minutes, even if guessed then timed once).
  • Stopper marks a draft guardrail.
  • Builder confirms the approved tool works on a sample that is allowed.
  • CAIO or L&D confirms aftercare: who measures day 14 and day 30.
  • If the sample is a toy, reschedule.

Day 1 morning (frame and guardrail)

  • 15 min: job of the two days. Hours, not completions. No icebreaker.
  • 30 min: walk the current workflow with times. Paste, summarise, reformat, chase.
  • 30 min: finish the guardrail page. Stopper marks it.
  • 30 min: write the first play skeleton (trigger, inputs, tool, checks, output, owner).
  • 15 min: pick today's live artefacts (three).

Day 1 afternoon (first passes)

  • 90 min: run the play on live work. Builder drives. Domain lead judges truth.
  • 30 min: capture failures (skipped step, bad source, check too weak, check too heavy).
  • 30 min: rewrite the play. Kill clever prompts that the desk will not repeat.
  • 20 min: time one cycle. Write the ugly number.

Day 2 morning (tighten)

  • 30 min: teach-back. Someone who was quiet yesterday runs the play.
  • 60 min: second set of live work. Improve the check, not the poetry.
  • 30 min: write the Tuesday version (one page). If it needs a manual, it will not run.
  • 30 min: agree the hour-measurement method (below).

Day 2 afternoon (decide and hand over)

  • 20 min: keep, kill, or constrain. Decision-maker or domain lead says it out loud.
  • 20 min: rule for returned hours (maker time, customer time, deleted room).
  • 20 min: day 14 and day 30 calendar holds. Named measurer.
  • 20 min: residue note sent. Play and guardrail in the same folder.
  • 20 min: who teaches the next two people. Not a video for 400.

Two days is enough if the workflow was already hated. It is not enough if you are still choosing a model. Choosing a model is a different meeting. Do not mix them.

CAIO first 120 days

Watch the job, not the white paper. When a bank in Singapore or Hong Kong, a telco in Australia, or a product company in Sydney or Singapore stands up a CAIO, Head of AI Enablement, or an L&D lead with AI literacy in the first-year objectives, that is a programme. The first 120 days decide whether they bought an operating rhythm or a town-hall video.

We do not invent a quote from those people. The org chart is the quote.

Days 1 to 30

  • Decline the vendor parade as the programme.
  • Name three workflows with clocks. Pick one.
  • Name the domain lead and the builder. Book two days.
  • Write the guardrail with operational risk and data.
  • Baseline the clock once, in person.

Days 31 to 60

  • Run the two-day sequence on live work.
  • Measure hours at day 14 (this window will catch it).
  • Publish keep / kill / constrain internally.
  • Agree with CHRO what happens to returned hours.
  • Do not fund a second capability map.

Days 61 to 90

  • Measure hours at day 30.
  • Teach-back: L&D or the builder facilitates a second desk, supplier in the chair not at the whiteboard.
  • Kill plays that did not pay. Scale the one that did (same guardrail family, not a new religion).

Days 91 to 120

  • Third workflow, or the first team runs without the supplier.
  • Report to the executive: hours, plays adopted, plays killed, not completions.
  • Only then talk about a programme name the board can remember.

What each role owes

CAIO (or CDO / Head of AI). Guardrails, approved tools, the "what we will not automate" list. Funds the first two sprints. Does not fund a second map.

CHRO. Incentive and job design. Time to practice. Protects maker hours. Refuses a bonus line that punishes the experiment.

L&D / capability. Sequence and cohort design. A repeating two-day format with residue. Reports hours and plays adopted.

When these three appoint into a bank in HK or SG, a telco in AU, or a product company in Sydney or Singapore, the first 120 days are the product. Miss them and the next person inherits a portal and a shrug.

HK / SG bank variant (DBS, OCBC, HSBC Hong Kong)

Singapore's MAS-shaped environment and Hong Kong's HKMA-shaped environment both push banks toward controlled adoption. That produces a particular hire: someone who can translate a model into a play a control function will sign. You are not buying a Silicon Valley "move fast" course.

DBS has made digital and AI enablement a multi-year operating conversation in Singapore. OCBC has run its own digital and AI literacy pushes and hiring. HSBC Hong Kong sits inside a global bank and a local licence. The local enablement staffing is the tell that matters for buyers in HK. Three names. One physics: volume, regulation, and a workforce that will not be replaced by a chatbot demo in a town hall. We will not treat them as interchangeable. We will not invent their scores.

Variant rules

  • Guardrail page is a day-0 gate, not a day-2 appendix.
  • Domain lead is a desk that already owns a clock (memo, pack, letter, exception).
  • Builder may sit in enablement, but the desk owns the play.
  • Human check is timed and reported. A play that "saves" two hours and then skips the check is an incident.
  • Shadow tools are out of scope. Approved tool only.
  • Publish the kill. DBS-style and GDS-style cultures already know a killed play is a result. Copy that honesty.
  • Do not start with a model bake-off. The model is downstream of the workflow and the control.

If you just walked into that job

  1. Name three workflows with clocks. Pick one.
  2. Pair the domain lead and the builder. Book two days.
  3. Write the one-page guardrail with operational risk and data.
  4. Leave a play. Measure hours at day 14 and day 30.
  5. Publish the kill or the scale. Then pick workflow two.

If you do that, the appointment becomes a residue. If you do not, the appointment becomes a LinkedIn post and a quiet backfill later.

AU and NZ banks will rhyme. UK and Canadian banks will rhyme. Use HK and SG as a leading indicator of the job, not of a model brand.

Telstra / NAB rhyme

Telstra-scale telco and NAB-scale banking are staffing the same cluster in Australia: enablement, not only engineering. The tell is not another model announcement. The tell is whether care, field, and corporate teams (Telstra) or credit, ops, and product teams (NAB) get sequenced enablement: guardrails, use cases, practice.

Telstra-shaped sequence

Pick a care or field explanation with a clock. Guardrail: customer data and brand sentence. Domain lead: care or field manager. Builder: knowledge or ops designer. Measure minutes per explanation, then hours across a roster at day 14 and 30. Kill if the check is skipped or the minutes do not move. Do not start with a new knowledge platform.

NAB-shaped sequence

Pick a decision explanation, a small-business document chase, or a pack. Guardrail: conduct, operational risk, data. Domain lead: product plus ops. Builder: BA or engineer on an approved tool. Test as a shadow or a first-pass draft with a named checker. Kill if it requires a core change on day one or the checker is fictional. APRA and ASIC did not ban first-pass drafts. They ban uncontrolled change to a customer in production. Write the difference on the guardrail.

Shared rhyme with HK / SG

Same two days. Tighter or looser data rules, same artefacts: guardrail, play, hours. Capital One-style US product orgs and a Canadian bank desk will recognise the rhyme. Canva-scale product ops will recognise the hours argument and still need a content-risk line on the guardrail. Infinitev-shaped offer tests are a cousin (AI Design Sprint, claims in the room), not a substitute for desk literacy.

If Telstra-shaped and NAB-shaped programmes land agile training in the same quarter, read it as cadence plus hours, not two murals. AI plays should demo in a heartbeat to a sceptic. That is transformation language a listed company can live with.

Measuring hours at day 14 and day 30

If you do not baseline, you will hallucinate savings. If you only baseline, you will narrate a hope.

Measurement sheet (one row per play)

  • Play name
  • Owner
  • Roster size that should run it
  • Baseline minutes per cycle (time one real cycle on day 0 or day 1)
  • Cycles per week
  • Human-check minutes required by the guardrail
  • Day 14: minutes per cycle, cycles run, checks skipped (must be zero)
  • Day 30: same
  • Hours returned per week = ((baseline - new - check) * cycles * people who actually run it) / 60
  • Decision: keep / kill / constrain
  • Rule applied to returned hours (maker / customer / deleted room)

Publish the sheet. Include zeros. A play that saves six minutes will not survive contact with a CRO. Kill it. The literacy is in the killing as much as in the prompting.

How to time without theatre

  • Time a real artefact, not a demo file.
  • Include the check. A "saving" that deletes the check is a risk event.
  • Day 14 is a leading indicator (adoption and friction). Day 30 is the decision.
  • If only the two people from the room can run it at day 30, you do not have a play. You have a pair of heroes. Constrain or kill.
  • If minutes move and checks are skipped, constrain first (tool access, checklist). Then kill if it repeats.

90-day academy that is not an LMS (optional wrapper)

If the board needs a programme name, make it three sprints, not a content library.

  • Sprint 1: one workflow, one team, hours at day 30.
  • Sprint 2: a second workflow, teach-back so L&D can facilitate.
  • Sprint 3: the first team runs without the supplier. You measure whether the play still runs on a wet Tuesday.

The LMS can hold the guardrail and the play template. It cannot hold the residue.

Don'ts

Don't start with a model bake-off. Start with a workflow. The model is downstream.

Don't train 400 people and change zero desktop standards. You will get 400 shadow workflows and a headache for risk.

Don't let a vendor run "AI 101" as the programme. Awareness is a Tuesday lunch. Enablement is a sprint.

Don't measure success as licences or completions. Measure hours at day 14 and day 30.

Don't hide the failures. A killed play is a result.

Don't run the room without a domain lead and a builder. A CoE roadshow is not literacy.

Don't write the guardrail after the prompts. HSBC Hong Kong, DBS, OCBC, NAB, Telstra: the control sits next to the play.

Don't skip the human check to make the hours look good. That is how you buy an incident.

Don't fill returned hours with more of the old work. Put a rule on them.

Don't fund a second capability map in the CAIO's first 120 days. Fund two sprints.

Don't invent quotes or metrics for DBS, OCBC, HSBC Hong Kong, Telstra, NAB, Canva, Atlassian, or anyone else named here. Steal the sequence.

Don't confuse this playbook with an AI Design Sprint. Literacy returns hours on a desk play. A sprint ships a decision and a throwaway artefact (Infinitev-shaped pages live there). Buy the one you mean.

If you are 40 days into the CAIO job at a Telstra-scale telco, a NAB-scale bank, a DBS-scale or HSBC Hong Kong-scale or OCBC-scale bank, or a Canva-scale product org, send this to your CHRO and your L&D lead. Name the workflow. Book the two days. Write the guardrail. Measure the hours. Kill or scale.

That is AI training without the slideware. The rest is a keynote.