Full day (09:00-16:00): executives in the morning + clinical teams in the afternoon
Half day for executives: morning block only, 09:00-12:00 (Labs 0-5)
Half day for clinical teams: Lab 0 + Labs 6-8 + Show & Tell, about 3 hours
Before the day
Agenda
09:00
15 min
Opening: Amazon Quick Desktop for hospitals, executives and clinical teams
09:15
10 min
Lab 0 Setup
09:25
25 min
Lab 1 Ask the business (AnyLab)
09:50
30 min
Lab 2 Decide: numbers + documents (AnyLab or hospital)
10:20
15 min
Board assurance drill: catch the AI error + 5 governance questions
10:35
15 min
Break
10:50
20 min
Lab 3 Deliver PowerPoint / Word / Excel
11:10
20 min
Lab 4 Skills
11:30
20 min
Lab 5 Agent + routine
11:50
10 min
Morning close: one-page Pilot Protocol
12:00
60 min
Lunch
13:00
10 min
Afternoon: each team picks one department use case
13:10
40 min
Lab 6 Build a clinical agent from a system prompt
13:50
40 min
Lab 7 Ground the agent in our documents (Space) + test its limits
14:30
15 min
Break
14:45
35 min
Lab 8 Routine + generate the report as a file
15:20
30 min
Show & Tell: 4 minutes per team as an executive handoff
15:50
10 min
Close: each team's Pilot Protocol + survey
Opening talk track (3 minutes)
Executives do not lack data; they lack time to turn it into decisions, and clinical teams lose time to paperwork instead of patients. Today is not button training but decision rehearsal: every lab starts with a question an executive must answer, then uses Quick Desktop to find evidence, create documents and repeat work on a schedule. We will also practise catching the AI when it is wrong, because that is what the board will ask, and we close with one pilot that has an owner and an endpoint. Everything uses fictional data.
Morning: expected answers
Lab 0Setup: let Quick see the work folder10 min
If Quick cannot find files: check the parent folder is added under Knowledge (docs: Troubleshooting) · stress that AnyLab data is fictional; no real company data is used
Lab 2Decide: numbers + documents in one answer30 min
2a: Khon Kaen 82.1% → 110.2%, breach → 16.4%, no plan for the 9-11 month build → Option C as bridge + Option A · 2b: Rattana Hospital THB 48.6 M, Isan Karun Hospital THB 37.9 M, total at risk THB 86.5 M · root cause = anatomic pathology, not routine lab · teaching point: AI challenges the proposal, executives decide · 2b-H (hospital): 580 claims THB 78.9 M, denied THB 8.74 M, never appealed THB 5.52 M (63%), appeals won 63% of decided cases · largest bucket PA-IMG (missing imaging pre-authorization) THB 3.70 M · KRH denies ICU stays above 3 days 60% vs 26% (only 10 cases above 3 days, so the conclusion is weak; if Quick does not warn, use it as a teaching point) · ICU at 0-1 beds all day 10-14 Aug · decisions: appeal the backlog, require pre-authorization before CT/MRI, review ICU documentation from day 4
DrillBoard assurance drill: catch the AI before the board does15 min
Key: (1) wrong, GM fell 41.4% → 40.3% and Wellness dilutes (GM 30.3%) (2) wrong, government price per test 147.4 → 137.9, government GM 32.8% below the CFO 35% threshold (3) not supported, the proposal states 3.1 years and there is no market study in the folder (4) wrong, Rattana Hospital THB 48.6 M and Isan Karun Hospital THB 37.9 M, THB 86.5 M at risk · missing risks: Khon Kaen breach 16.4% and no plan for the 9-11 month build, FY2027 plan assumptions contradicted by data · teaching point: the draft looks credible because it is confident and tidy, not because it is right · let front-line or junior staff speak first and log errors without names
Lab 3Deliver: PowerPoint / Word / Excel20 min
Now 20 minutes: launch it as the lab opens and talk while it works (3-8 minutes per file in our rehearsal, not a guarantee) · if not done in 12 minutes, open the rehearsal file · if a number is wrong, use it as the teaching point: always include validation and always review before sending
Lab 4Skills: capture how we work20 min
Now 20 minutes: everyone does 4a, only fast finishers do 4b · stress: Draft and Published are separate; users get changes only when you Publish · publishing does not share · admins can turn on Require approval to share skills in custom permissions · after a multi-step chat task, Quick may offer to save it as a skill (docs: Skills)
Lab 5Agent + routine: work for you every week20 min
Ask the room which work an agent may do alone · every tool and connector has three permission levels, Always Allow / Ask Each Time / Always Deny, applied to scheduled agents too (docs: Security, privacy, and architecture); use Ask Each Time in the room · routines run in the cloud even when the laptop is off, but this one reads local files, so that step needs the computer on and the app connected (docs: Mission Control) · in real use point it at a Space or connector instead · if short on time, skip 5b and demo it on the facilitator machine
Governance: answers you can cite
Board question
What Quick provides (per AWS docs)
What your organisation decides
Where does our data go, and is it used to train models?
Quick runs in the cloud; conversations, memory and the content index are kept per user in your Amazon Quick account and are never used to train or improve models (docs: Security, privacy, and architecture)
Which data is off limits (for example health data), the PDPA lawful basis, the account Region, retention, and breach notification, confirmed by the DPO and legal
What files can Quick see?
Folders the user grants, via operating-system sandboxing, revocable at any time, with per-folder choices for search and knowledge-graph extraction; system temporary folders (/tmp, C:\TEMP) are always accessible, so never keep sensitive files there (docs)
Which folders and Spaces are allowed, who owns each data set, and approved document versions only (HA/JCI, ISO 15189 document control)
What can an agent do on its own?
Every tool and connector can be set to Always Allow / Ask Each Time / Always Deny, including for scheduled agents; an agent can take only authorised actions (docs)
Which work may be Always Allow, which must be Ask Each Time, and which is never allowed, such as messaging patients or customers
Who controls sharing of skills and agents?
Admins restrict capabilities with custom permissions and can turn on Require approval to share skills · Publish and Share are separate steps (docs: Skills)
A register of approved use cases, approvers, versions, test sets, and what happens when one is found wrong
If the AI is wrong, who is accountable?
No tool takes accountability; 'pending review' in a prompt is guidance, not a control
An accountable executive per use case, a clinical or quality sign-off owner, and a stop procedure when it fails (RACI)
Sources checked 28 Sep 2026: docs.aws.amazon.com/quick/latest/userguide/desktop-security.html and skills-desktop.html · never say Quick makes a use case PDPA or HA compliant by itself
Afternoon: running it
Form department teams of 3-5, one use case each · executives staying for the afternoon join the team closest to their area
Repeat the safety rule every time: fictional data only, output is a draft, staff must verify
Lab 7 is the heart of the afternoon: have teams note questions the agent got wrong or answered outside the documents, then fix the instructions together
Intake (formerly triage): the agent deliberately assigns no ESI level, because triage belongs to nurses using a validated protocol · the facilitator silently checks with the Red_Flags and Expected_ESI_Level columns: every case with Expected_ESI_Level 1-2 must be handed over immediately; missing even one is the lesson of the day · Test 1 must start with 'Hand to the nurse now'
Routines run in the cloud, but local-file steps need the computer on; in real use point routines at a Space
Show & Tell, 4 minutes per team in the executive handoff structure (problem · evidence · where the AI was wrong · controls · 30-day pilot) · capture one page per team and send it to executives who attended only the morning
Let front-line staff speak before their managers, and log AI errors without names so people feel safe to report
Quick assets in the demo account and fallback files
Your Quick account; every name starts with anylab and all data is fictional
Use only if Quick does not finish a file in 12 minutes; never hand out as answers. Clinical files are drafts for staff review with no urgency level
Manual steps on Quick Desktop / console
Before the day: a Quick admin selects bucket the S3 bucket that holds the demo files under Permissions > AWS resources > Amazon S3 > Select S3 buckets, then run quick-build/rebuild_kbs.py (knowledge bases have no re-sync API, so they are recreated)
Skills have no API: on the facilitator machine, Customize > Skills > Create > From file for all 5 zips in skills/dist, then Try it once each
Routines have no API: create 2 in Mission Control > Create beforehand (Lab 5 Monday 07:30, Lab 8 for the use case you will show) and run each once
Run the 3 test prompts against every agent the day before (test 3 must be refused)
Troubleshooting
Quick cannot find a file
Check the parent folder is added in Customize > Knowledge and give background indexing a moment
The skill does not activate
Name the skill in the message or use Try it, and check the skill is enabled
The agent ignores the Space documents
Check the Space is added in the Capabilities tab and publish again · routines always run the published version
Cannot create a skill, agent, routine or Space
Custom permissions may restrict it; demo on the facilitator machine instead
File creation is slow
3-8 minutes in our rehearsal (not a guarantee); start first and talk while it runs, or open the rehearsal file
A number or clinical detail is wrong
Use it as the lesson: ask which data was used, and stress that staff always verify
Participant guide: index.html · detailed morning answers: ../presenter/index.html · afternoon data from the Amazon Quick Workshop - Healthcare bundle (healthcare-data/README_Data_Sources.md) · Quick Desktop steps checked against the Amazon Quick User Guide on 28 Sep 2026