💡 Click any ring to explore that layer
How The Layers Build On Each Other
Each layer adds a new capability on top of the one below. AnyCompany Support uses every level — pick the right layer for the job.
⚡ What Can Generative AI Do?
A practical menu of GenAI capabilities — illustrated through work GS Support teams already do today, every shift, across every market.
Text generation
Drafts that read like a teammate wrote them — first-response emails, escalation Slack templates, DSAT root-cause write-ups, post-call notes.
Summarization
40-turn THA chat transcript → 1-page case context. 20-min IRT call → 5-bullet timeline. ARC chatbot handover → structured brief for the agent.
Q&A over your SOPs
"How do I handle a refund for a cancelled GrabFood order with partial delivery in MY?" — answered from the SOP / KB corpus, with citation. Anchored on your SOP Lookup Top Idea.
Classification with reasoning
"This is a true Safety case (severity P1) because…" — auto-fill PAC tags + DSAT root cause + Safety severity, with rationale visible for QA.
Translation & localisation
Read a Vietnamese chat transcript, write the case summary in English. Localise an SOP step for KH/MM agents working in a 2nd language.
Code & SQL generation
"Write the SQL to pull this week's IRT cases by symptom_l1 from the datalake" — useful for the WFM / Reporting specialists in this room (4 of you).
Image & multimodal
Read a screenshot a PAX sent, extract the booking ID, parse a receipt image. Multimodal models handle screenshots, photos, and chat captures the way agents actually receive them.
Voice & transcription
Transcribe an outbound IRT call, extract action items, generate the post-call note (Project Steno / Project Echo territory). The agent reviews and approves.
🏗️ Foundation Models — The Building Blocks
Behind every GenAI application is a foundation model — a large pre-trained AI that's been fine-tuned for the task. Knowing the basics helps you choose wisely.
What is a foundation model?
A massive AI trained on huge volumes of text, code, images, and audio — that can be adapted to many downstream tasks without re-training from scratch.
Trained on broad data
Books, websites, papers, code repositories, transcripts. The model learns language patterns, world knowledge, and reasoning shortcuts from this corpus.
One model, many tasks
Same Claude can write a poem, summarize a 10-K, classify a vendor, and explain an audit finding. Generality is the breakthrough.
Scale matters
Bigger models (Opus, Nova Premier) capture more nuance but cost more. Smaller models (Haiku, Nova Micro) are fast and cheap. Match model to task.
The four families you'll meet in this workshop
| Family | Maker | Strengths | Where it shines for finance |
|---|---|---|---|
| Claude (Opus / Sonnet / Haiku) | Anthropic | Reasoning, instruction-following, safety, long context | The default for Claude Cowork. Case summaries, SOP Q&A, DSAT write-ups, edge-case Safety reasoning. |
| Amazon Nova (Premier / Pro / Lite / Micro) | Amazon | Cost-efficient at every tier; wide multimodal range | Cost-sensitive bulk workloads. PAC auto-tagging across 400k MIWI cases / month, ticket classification. |
| Llama (Meta) | Meta | Open-weight, customizable, strong general-purpose | Self-hosted scenarios. Specialist-tier note: useful when you need to fine-tune on Grab-specific terminology. |
| Mistral / SEA-LION / others | Various | Specialised strengths — multilingual, including SEA languages | SEA-language transcripts (THA / BHS / VNM / FIL) where general models struggle with native phrasing or short forms. |
☁️ AWS Generative AI Services — The Toolkit
AWS provides a stack of services from the foundational model layer to ready-to-use applications. Here's the part that matters when you're scoping a Cyborg build with GTS / GSTF.
Amazon Bedrock
The unified API to access foundation models — Claude, Nova, Llama, Mistral, and others — without managing infrastructure. Where your Cyborg AI workloads run when GTS / GSTF productionise them.
Bedrock Guardrails
The safety layer — redacts PAX / DAX / MEX PII, blocks prompt injection, filters off-topic responses, enforces grounding. Module 7 has the live demo.
Bedrock Knowledge Bases
Managed RAG — point Bedrock at your SOP / KB corpus, and it handles embeddings, vector store, retrieval. Module 9 covers the pattern (your SOP Lookup Top Idea sits here).
Bedrock Agents & AgentCore
Build autonomous agents that orchestrate Skills, tools, and data. Day 2 territory — your Cowork agent productionised inside the Grab AWS account.
Amazon SageMaker
End-to-end ML platform — used by data-science teams building custom models (e.g., the AI Forecasting Module from your Top Ideas list). GTS / Eternals territory.
Amazon Quick
The new agentic AI workspace from AWS — connects to enterprise data (docs, dashboards, warehouses), answers questions, and takes action. Successor to Amazon Q Business; analogous to Glean + Valet for the AWS-native side.
Specialised AI services
Textract (receipt / screenshot extraction), Transcribe (call audio → text — Project Echo / Steno territory), Translate (SEA languages), Comprehend (NLP). Useful building blocks inside larger agents.
The integration layer (GTS / GSTF)
API Gateway, Lambda, Step Functions — what plumbs an agent into D365, the datalake, Slack, NICE IEX. Day 2's Plugins / MCP topic touches this. You don't build it — but you scope it.
What you actually need to know
| Layer | Who works at this layer | What you do here |
|---|---|---|
| Application layer (Cowork, GrabGPT, Valet, Glean) | You — GS Ops, Specialists, TQM, WFM | Use the AI directly. Write prompts. Build Skills. This workshop's primary level. Where Cyborg apps live. |
| Model layer (Bedrock, Bedrock Guardrails) | GSTF / GTS / Cyborg Specialist tier | Choose models, configure guardrails, monitor cost & quality. The Cyborg builders in this room will dip into this. |
| Infrastructure layer (SageMaker, Lambda, Step Functions) | GTS, Eternals, platform engineering | Customise, deploy, integrate with D365 and the datalake. Not your job — but knowing it exists helps you scope and hand-off cleanly. |
⚠️ Three Challenges Every Cyborg Builder Should Know
Generative AI is powerful — and unevenly reliable. These are the three failure modes you'll see most often in GS workflows, and where the rest of the workshop addresses each.
Hallucination
The model invents plausible-sounding facts — a SOP step that doesn't exist, a refund threshold that's wrong for your market, a confidently invented booking ID. Always verify policy answers, amounts, and named entities against D365 / KB.
Privacy & data leakage
Real PAX / DAX / MEX names, phone numbers, booking IDs, and case content must never leak into public AI tools unprotected. Bedrock Guardrails enforces redaction; steering rules in Cowork enforce the rule per project.
Cost runaway
The same case-summary task can cost 1¢ or $1 depending on model choice and conversation length. At 400k MIWI cases / month, that's the difference between a fundable Cyborg pilot and a budget incident.
🎓 Where this workshop takes you next
You've now seen the full GenAI landscape — the hierarchy, what it can do, the foundation models, the AWS toolkit, ML vs GenAI, and the three biggest risks. From here, the rest of Day 1 makes you fluent and Day 2 makes you operational.
Traditional ML vs Generative AI
Both are valuable — the right choice depends on your task.
| Aspect | Traditional ML | Generative AI |
|---|---|---|
| Architecture | Task-specific models (XGBoost, Random Forest) | Foundation models (Claude, GPT, Nova) |
| Training | One model per task, your data | Pre-trained on massive data, adapted to your task |
| Best for | Structured data, predictions, scoring | Text generation, summarization, reasoning |
| Speed | Fast inference (milliseconds) | Slower (seconds), more compute |
| Cost | Low per prediction | Higher (token-based pricing) |
| GS example | DUI / Safety scoring: flag a true Safety case in milliseconds when it lands | Case narrative: generate the structured case context summary explaining the score for IRT |