Chapter 1 is the foundation. Everything in Chapters 2, 3, and 4 assumes you can do the things below. Take your time — the depth here pays for itself in every later exercise.
AI is software that learns from data. Not magic, not consciousness, not a coding course.
Four families (ML/NLP/CV/GenAI) plus three business outcomes (HBR's automate/insight/engage).
Map every AI conversation back to one of MPV's four pillars and Vietnam's 2026 legal framework.
An honest, working definition for a non-technical employee: Artificial Intelligence is software that learns patterns from data, rather than following hand-written rules. That's it. No magic, no consciousness, no Hollywood. The difference shows up clearly when you compare it to MPV's existing systems.
| System | How it works | MPV example | When to use |
|---|---|---|---|
| Automation | Hand-coded "if-this-then-that" rules | If oven temperature > 180°C, send SMS to maintenance supervisor | Stable, predictable, repeated steps |
| Analytics | Summarises historical data into reports | Power BI dashboard shows monthly defect rate by line | Reporting, visibility, lagging KPIs |
| AI / ML | Learns patterns from examples | Predicts which injection-press will need maintenance based on 6 months of sensor data | Complex patterns, flexible inputs, predictions |
| Generative AI | Creates new content from instructions | Drafts a CAPA report in Vietnamese from inspection notes; QA edits and approves | Language-heavy knowledge work, drafting, summarising |
If any of these three is missing or weak, the AI project will struggle. Use this as a quick screening test before committing any MPV money.
Volume and quality. Predictive maintenance needs months of machine sensor logs. Defect detection needs thousands of labelled syringe images. A pilot with 50 images will fail; with 5,000 it can work.
The maths that learns from the data. You will almost never write these. You will pick from FPT.AI, Claude, ChatGPT, MISA AMIS, or a vendor — and the algorithm is already built.
The hardware to run it. Cloud (Azure, AWS, FPT AI Factory) for most cases; on-premise GPUs only when data sensitivity demands. MPV starts cloud-first.
AI as an idea is from the 1950s. AI as a workable tool is from about 2012. What changed: GPUs got 1,000× faster, the internet generated petabytes of training data, and the Transformer architecture (2017) made language models possible at scale. ChatGPT (Nov 2022) made it consumer-grade. By 2026, every Vietnamese smartphone has access to GPT-class intelligence for free. The window MPV is operating in is unprecedented.
Most AI conversations get lost because people mix these four families. Learn the names and one MPV example for each — you will sound informed in any vendor meeting.
Finds patterns in numerical data. Used for forecasting, classification, anomaly detection.
MPV: forecast monthly hospital orders by SKU; detect anomalous defect rate by shift; score supplier reliability.
Understands and produces human language. The basis of email triage, complaint classification, translation.
MPV: classify incoming customer emails (delivery / quality / billing / spec query); translate hospital tenders Vietnamese ↔ English.
Analyses images and video. The core of MPV's Quality Inspection pillar.
MPV: detect syringe barrel cracks; verify needle-hub alignment; check label print quality on packaging line.
Creates new text, images, code, or audio from instructions. ChatGPT / Claude / Gemini live here.
MPV: draft CAPA reports; write hospital reply emails; produce Vietnamese training materials from English source docs.
Knowing the types of AI tells you how a system works. To spot opportunities at MPV, you need a second lens: what AI is FOR in business terms. Harvard Business Review's January–February 2018 landmark article "Artificial Intelligence for the Real World" by Thomas Davenport and Rajeev Ronanki studied 152 cognitive-AI projects across 250 companies. They found that every project — without exception — fell into one of three categories. This framework is now 8 years old and has aged exceptionally well. It is the single most useful filter you will take from this course.
"Robotic-style" AI that performs repetitive admin and back-office tasks across systems. Think: reading invoices, transferring data between IT systems, classifying emails, updating customer records. These projects have the fastest ROI and the highest success rate in the HBR study.
"Analytics on steroids." AI detects patterns across vast data and interprets meaning — predicting outcomes, spotting anomalies, segmenting customers. This is where Predictive Maintenance and Demand Forecasting live.
Chatbots, intelligent agents, recommendation engines — natural-language interfaces that handle questions and conversations. Highest visibility, highest failure rate (technical and behavioural). Roll out carefully.
The same HBR study compared MD Anderson Cancer Center's $62 million moon-shot AI cancer-diagnosis project — put on hold without ever being used on patients — with the same hospital's small AI projects: hotel recommendations for patients' families, billing-help identification, IT support automation. The small projects worked. Patient satisfaction rose, finances improved, nurse-manager time was freed up. The moon shot did not.
In Chapter 4 you will build a personal AI Opportunity Map. Every entry must be labelled with the HBR category. Mr. Giang from the program review insisted on this — and he is right. Labelling forces clarity: leadership instantly knows what kind of tool, budget, and risk profile each idea represents. A "Category 1 quick automation" gets a different conversation than a "Category 3 customer chatbot".
You will use LLMs every day. You should understand what they actually do — because knowing how they fail prevents the failures from hurting MPV.
Hundreds of billions of words from books, websites, code. The model "reads" all of it and learns the statistical patterns of language.
"Write me a CAPA report for a syringe barrel-crack defect on Line 3, in Vietnamese."
Given your prompt and everything it has learned, it computes: "what word probably comes next?" Then the next. Then the next.
The result is grammatically perfect, contextually plausible — and may or may not be factually correct. The model does not know facts; it knows patterns.
| Reliable | Faked (Hallucinated) |
|---|---|
| Drafting an email in a specific tone | Specific clause numbers in regulations |
| Summarising a document you paste in | Specific dates of events (especially recent) |
| Translating between languages | Statistics not present in the prompt |
| Restructuring information you provide | Names of people, contracts, court cases |
| Brainstorming, ideation, alternatives | Numerical "facts" that sound precise |
Every workshop surfaces these. Address them now — they're the most common reasons good ideas die in committee.
No. Today's AI has no understanding, no goals, no consciousness. It is statistical pattern-matching at enormous scale. This is why it can write a perfect Vietnamese email and also confidently invent an ISO clause that does not exist. Treat it like a brilliant intern who is fluent in everything but sometimes makes things up — useful, but never unsupervised.
Not in Phase 1, not in Phase 2. AI replaces tasks, not jobs. A QA inspector still inspects; AI drafts her summary report so she spends less time on paperwork. A finance officer still owns the books; AI drafts the variance commentary so she focuses on the analysis. People using AI will outperform people not using AI — but that's an upskilling story, not a layoff story. MPV's positioning is to grow capability, not shed people.
The opposite. AI learns from data — and data carries the biases of whoever collected it. A hiring AI trained on 10 years of MPV hires will reproduce 10 years of MPV's hiring patterns, including any unconscious bias. A defect-detection AI trained only on Line 1's syringes will misclassify Line 3's variations. Bias is the responsibility of the deployer — Chapter 3 covers this in depth.
Not in 2026. ChatGPT, Claude, Gemini, FPT.AI, MISA AMIS, Power Automate — these are all no-code or low-code. The skill you need is prompting (Chapter 2's CRAFT framework), output evaluation (Chapter 3's rule-based scoring), and opportunity-spotting (Chapter 4). Coding is for Phase 3 system builders, not for Phase 1 users.
Not anymore. A ChatGPT Plus or Claude Pro subscription is ~500,000 VND/month per user. A factory-floor pilot using FPT.AI or VinAI for defect detection is in the tens of millions of VND, not hundreds. The expensive AI is the badly-scoped AI: a moon-shot project that consumes resources without delivering. Chapter 4's scoring grid prevents that. Done well, Phase 1 pays for Phase 2.
The market timing for MPV is unusually favourable. Understanding why explains every Phase-2 and Phase-3 investment decision the leadership team will make.
Vietnamese hospitals are buying ~$1.74B of medical devices a year and roughly 90% of that flows to imported brands — American, European, Japanese, Korean, increasingly Chinese. MPV competes in disposables (syringes, infusion sets, surgical consumables) where price sensitivity is highest and the local manufacturer's natural advantages — proximity, faster lead times, Vietnamese-language support, lower logistics cost — are real but easily eroded by larger competitors that already use AI for quality and supply chain.
Three pieces of legislation directly shape MPV's AI deployment window. You will see all three again in Chapter 3 in compliance depth; here, see the strategic framing.
| Instrument | Effective | What it means for MPV |
|---|---|---|
| Law on Digital Technology Industry | 1 Jan 2026 | Frames digital industry as strategic. Up to 10% preferential corporate tax for qualifying digital activities. R&D deductions. Sandbox provisions for new technology. |
| Law on Artificial Intelligence (No. 134/2025/QH15) | 1 Mar 2026 (in force now) | Vietnam's first standalone AI law. 35 articles. 4-tier risk classification. Human-centric principle. Healthcare gets 18-month grace period (until 1 Sep 2027). Penalties up to 2% annual revenue. |
| Decree 13/2023 on Personal Data Protection | 1 Jul 2023 (already in force) | Defines consent, lawful basis, cross-border transfer. 72-hour breach notification. Applies to any AI processing of employee/customer personal data. |
Every AI opportunity at MPV will fall under one of four operational pillars. These are not abstract — they map directly to existing budget owners and ISO 13485 process areas.
Computer Vision on the syringe / infusion-set line. Defect detection, label verification, batch-record image archival.
Owner: QA Manager · HBR Cat: ② Cognitive Insight
ML on injection-press sensors (vibration, temperature, current draw). Predicts failures 5–14 days early; reduces unplanned downtime.
Owner: Maintenance Engineer · HBR Cat: ② Cognitive Insight
ML for hospital demand forecasting by SKU; NLP for tender analysis; GenAI for supplier-spec translation and PO drafting.
Owner: Supply Chain Manager · HBR Cat: ① + ②
GenAI drafts CAPA reports, deviation investigations, MOH DMEC submissions, ISO 13485 audit responses. QA verifies and approves.
Owner: Regulatory Affairs · HBR Cat: ① Process Automation
| Department | Primary Pillar | First AI Touch |
|---|---|---|
| Quality Assurance | Pillar 1 (Inspection) + Pillar 4 (Compliance) | CAPA drafts in GenAI; CV vendor scoping |
| Maintenance Engineering | Pillar 2 (Predictive Maintenance) | Sensor-data audit and Phase 2 pilot scoping |
| Supply Chain & Planning | Pillar 3 (Forecast & Tender) | Demand-forecast prototype in Excel + GenAI |
| Regulatory Affairs | Pillar 4 (Documentation) | DMEC submission drafts; ISO 13485 audit prep |
| Finance & HR | All four (cross-cutting) | Vendor RFP drafts; policy translation; KPI dashboards |
MPV does not need to build AI from scratch. Vietnam has a mature stack of domestic AI providers, often better-suited to Vietnamese-language workflows and local data residency than US tools. Know these four names.
Vietnam's largest AI provider. Vision APIs, Vietnamese NLP, OCR, voice. FPT AI Factory (launched 2024) provides NVIDIA H100 compute for Vietnamese enterprises. Sales support in Vietnamese, contracts in VND, data residency in Vietnam.
VinGroup's AI lab. World-class Vietnamese-language models (PhoGPT, PhoBERT). Strong in Computer Vision for industrial inspection — relevant for Pillar 1.
Vietnam's leading SME ERP. Built-in AI features for invoice OCR, expense classification, contract review. Already in use across many Vietnamese SMEs in MPV's bracket.
The military-owned telecoms giant's AI arm. Strong in Vietnamese voice, OCR, and government-sector solutions. Useful for Decree 13 / AI Law-compliant deployments.
✅ AI literacy — the goal is safe, practical use; no coding required
✅ AI = software that learns from data, not hand-coded rules
✅ 4 types: ML, NLP, CV, GenAI — most real systems combine 2–3
✅ 3 HBR business categories: ① Process Automation 47% · ② Cognitive Insight 38% · ③ Cognitive Engagement 16%
✅ LLMs hallucinate — "AI drafts, human verifies" is the lifelong rule
✅ 5 myths debunked — AI doesn't think, doesn't replace whole jobs, isn't unbiased, doesn't need coding, isn't unaffordable
✅ Market context: $1.74B → $2.45B (2025–2030), 90% imported — MPV's window is open now
✅ Regulatory tailwind: AI Law in force 1 Mar 2026 (18-month healthcare grace until Sep 2027) + DTI Law (Jan 2026) + Decree 13/2023
✅ 4 MPV pillars: Quality Inspection · Predictive Maintenance · Supply Chain · Compliance
✅ Start with low-hanging fruit, not moon shots — MD Anderson's $62M lesson
✅ VN AI ecosystem: FPT.AI, VinAI, MISA AMIS, Viettel AI