Want to create interactive content? It’s easy in Genially!

Get started free

Chap 1 day 4

Talent Growth Team

Created on October 13, 2025

Start designing with a free template

Discover more than 1500 professional designs like these:

Akihabara Agenda

Akihabara Content Repository

Internal Guidelines for Artificial Intelligence Use

Correct Concepts

Sorting Cards

Interactive Scoreboard

Semicircle Mind Map

Transcript

Chapter 1 AI Fundamentals Day 4

Generative AI Basics

start

Understand the

technology eco-sytem

You'll use interactive elements throughout this module. To get you started click on the button here to view the learning objectives

Learning objectives

Now click here to see what you covered previously

Coming up...

good to know

Next

🛡️The Three-Layer Defence

Effective governance requires controls at three levels. Most organisations implement only one or two layers, leaving critical gaps. Complete frameworks address all three layers.

These are passive protections that don't require human intervention—like seatbelts in a car. Input validation: Automatically detects and blocks sensitive data (credit card numbers, PII, confidential information) before it reaches the AI Output filtering: Scans AI responses for toxic content, profanity, policy violations, or known hallucination patterns before showing them to users Rate limiting: Prevents runaway costs by capping how many requests can be made per hour/day/user (connects to Module 3's cost management) Audit logging: Automatically records who prompted what, when, and what the output was—creating an accountability trail for investigations

Use this side of the card to provide more information about a topic. Focus on one concept. Make learning and communication more efficient.

Layer 1: Technical Controls

Title

Write a brief description here

What the system enforces automatically. Flip over this card to learn more

Why does this layer matter?It protects you even when users make mistakes. Technical controls catch issues at machine speed, preventing problems before they reach humans.

Next

Back

1/8

🛡️The Three-Layer Defence

Complete frameworks address all three layers.

Technical controls catch obvious problems, but judgement requires people. Process controls define who reviews what, and when. Review gates: Identify which outputs need approval before reaching clients (e.g., all contract terms reviewed by legal, all financial projections reviewed by finance, all public statements reviewed by comms) Escalation paths: Define what happens when something goes wrong—who gets alerted, who has authority to shut down a capability, how quickly must they respond, what's the investigation procedure Update protocols: Establish how often knowledge bases refresh, who approves changes, how you test after updates to ensure quality hasn't degraded Ownership assignment: Every AI capability has one named person accountable for its governance and performance

Use this side of the card to provide more information about a topic. Focus on one concept. Make learning and communication more efficient.

Layer 2: Process Controls

Title

Write a brief description here

What humans do. Flip over this card to learn more

Why does this layer matter?It embeds accountability into workflows. When issues arise, everyone knows their role—no diffusion of responsibility.

Next

Back

2/8

🛡️The Three-Layer Defence

Complete frameworks address all three layers.

The first two layers manage systems and workflows. This layer manages people—ensuring they understand how to use AI responsibly and when to escalate decisions to humans. Training: Everyone using AI understands capabilities, limitations, and verification requirements (that's what this programme does—you're building Layer 3) Champions: Identify who leads responsible AI adoption in each team, who others can ask for guidance, who shares lessons learned Principles: Create shared understanding of when to use AI and when not to (e.g., "We never use AI for final decisions affecting people's livelihoods without human review", "We prioritise accuracy over speed for regulatory content") Feedback culture: Teams feel safe reporting AI failures or near-misses without blame, so issues surface quickly rather than being hidden

Use this side of the card to provide more information about a topic. Focus on one concept. Make learning and communication more efficient.

Layer 3: Cultural Controls

Title

Write a brief description here

What the organisation values. Flip over this card to learn more

Why does this layer matter?It prevents governance becoming a checklist people work around. When teams understand the "why" behind controls, they apply judgement appropriately—especially in ambiguous situations the rules didn't anticipate.

Next

Back

3/8

Why All Three Layers Are Necessary Technical controls provide the safety net, process controls embed accountability, cultural controls ensure appropriate judgement. This combination scales reliably.

Click on the 3 buttons above to read more into each layer

Next

Back

4/8

The pre-launch go / no go checklist

Before you switch on any AI capability—whether you're using an off-the-shelf tool or building a custom solution—complete this checklist. If you cannot tick all five boxes, do not launch.

Data residency confirmed

Do you know where your data is processed and stored? Does this meet client contractual requirements and regulatory obligations? (Links to vendor governance questions in Day 6)

PII handling documented

*Personal Identifiable Information (PII) Handling

How will personal data be detected, handled, or excluded? Have you completed a data protection impact assessment if required? Do you have contractual guarantees from your AI vendor?

Human review gate identified

Who is accountable for checking outputs before they reach clients or production systems? Is this person aware of their responsibility and trained to perform it?

Rollback procedure documented

If the AI produces harmful output or quality degrades, how do you quickly disable it whilst you investigate? Who has authority to make that decision?

One person accountable

Who owns this AI capability and is responsible for its governance, performance monitoring, and continuous improvement?

This isn't an exhaustive list of every possible governance concern. This is your Minimum Viable Governance, the absolute essentials. Start here and build upon this foundation as your use of AI matures.

Next

Back

5/8

Connecting guardrails to the model lifecycle

Day 1 & 2 introduced the model lifecycle: train, validate, deploy, monitor, and improve. Guardrails must exist at each stage:

TrainEnsure training data is representative, balanced, and doesn't contain inappropriate content. Document data sources and provenance. Validate Test across demographic groups for bias. Run adversarial tests to find edge cases. Validate against held-out data the model never saw. Deploy Complete Day One Checklist. Implement technical controls (filtering, logging, rate limiting). Define process controls (review gates, escalation paths). Monitor Track accuracy, bias metrics, cost, and user feedback. Set alerts for degradation. Review audit logs periodically. Improve Use monitoring insights to refine prompts, update knowledge bases, or retrain models. Document changes and revalidate.

Next

Back

6/8

⚡ Try It — Identify governance gaps (90 seconds)

ScenarioReview this partially complete governance framework for an AI-powered customer service chatbot:

What's in place: Technical Layer: Output filtering blocks profanity and offensive content Technical Layer: Audit logs record all conversations Process Layer: Customer service team reviews flagged conversations weekly

Next

Back

7/8

⚡ Can you answer this?

Next

Back

8/8

Click the top right corner and press the X button to exit Day 3 and save your progress

AI Fundamentals

Day 4

Understanding the technology eco- system completed

📄 Tomorrow's ApplicationList the repetitive work in your current project. For each item, ask: "Are we repeatedly creating similar content or repeatedly analysing similar data?" If it's content creation, you've found a generative AI opportunity. Share your findings in #ai-updates for peer feedback.

Back

The kitchen sink

What it looks like: Including every possible piece of background information, much of it irrelevant to the task Why it fails: Irrelevant context dilutes focus. Remember from day 5 & 6: you pay for every token, and excess information increases cost whilst reducing quality. Fix: Include only context that directly affects the output. If removing a detail wouldn't change the result, leave it out.

What's present: Technical Layer (partial): Output filtering and audit logs Process Layer (partial): Weekly review of flagged conversations What's missing: Technical Layer: No input validation (customers could submit PII or malicious prompts), no rate limiting (runaway costs possible) Process Layer: No escalation path for harmful outputs, no human review gate before responses reach customers, no update protocol for knowledge base Cultural Layer: Entirely absent — no training for customer service team, no identified champion, no principles guiding when AI hands off to humans Risk: This framework protects against obvious offensive content but leaves major gaps in data handling, accountability, and human judgement. It needs all three layers to be complete.

Process only: Creates bureaucracy without automatic protection. Relies entirely on humans remembering to check—doesn't scale.

Minor variations in phrasing, same substance:

Normal behaviour. The predictability setting (mentioned in Day 7) isn't set to maximum consistency, but the core content remains stable.

This outputs multiple verification failures: Project numbers: "Over 500 AI projects since 2018" requires verification—likely inflated Client names: NHS England, GCHQ, European Commission—these may breach confidentiality even if accurate Proprietary framework: "AIcelerate™" doesn't exist—AI invented it Research citation: Stanford University (2023) study is fabricated Certification claim: "ISO 27001 level 5" is nonsensical—ISO 27001 has no levels Unsubstantiated superiority: "UK's leading" requires evidence Gate assessment: This passes Gate 1 (looks professional) but catastrophically fails Gate 2 (factual validation). Every specific claim requires verification before this reaches a client.

Cultural only: Well-intentioned but inconsistent. Different teams interpret principles differently, leading to fragmented governance.

Amanda Jackson

Data Analyst

Shares interesting facts in meetings and organizes photographic exhibitions in the office.

Interpret large volumes of information with precision and clarity. Enjoy photography and reading historical novels. Has carried out several photographic exhibitions.

close

Recently at AND

For Gousto, we built a map of their food taxonomy — linking “quick midweek meals” to prep time, ingredient count, and equipment. Recommendations became far more accurate than with a generic model that lacked this context.

NAME PROJECT

CONTEXTUALIZE YOUR TOPIC

We don't like to bore. We don't want to be repetitive. Communicating as always is boring and doesn't engage. We do it differently. We sabotage boredom. We create what the brain likes to consume because it stimulates.

The one-shot

What it looks like: Expecting perfection on the first attempt, giving up when initial output isn't perfect Why it fails: Professional use of AI is iterative. First outputs are drafts—you refine the prompt based on what works and what doesn't. Fix: Treat prompting as a conversation. Run the prompt, review the output, adjust your instructions, run again. Effective prompts often take 2-3 iterations to perfect.

🧠By the end of day 4 you'll be able to,

  • Map out the technology layers and identify where AND creates value
  • Apply our decision framework to choose between build, buy, or compose
  • Recognize and mitigate the real business risks of AI implementation
  • Speak confidently about technology choices with both technical teams and business stakeholders

The Assumption

What it looks like: Using internal acronyms, assuming knowledge of your house style, expecting AI to know your brand voice Why it fails: AI doesn't know your organisation's norms unless you tell it. What's obvious to you is invisible to the model. Fix: Define acronyms, specify style preferences, provide examples of your brand voice.

The vague brief

What it looks like: "Make it better" or "Write something professional" Why it fails: AI cannot read your mind! "Better" and "professional" mean different things to different audiences. Without specific criteria, AI guesses—and often guesses wrong. Fix: Define "better" with measurable criteria (shorter, more formal, fewer technical terms, focused on outcomes).

Valeria Johnson

Project Manager

She always arrives by bike and organizes group walks on weekends.

Expert in organization and team management, always focused on improving efficiency. She loves hiking and nature photography. She has completed the Camino de Santiago three times.

close

Matthew Davis

Human Resources Responsible

Brings his dog on Fridays and organizes volunteering events.

Empathetic and skilled interpersonally, creating a positive work environment. Practices climbing and volunteers. Has climbed emblematic mountains and works in animal shelters.

close

Technical only: Catches obvious violations but can't handle context-dependent decisions. Systems don't understand nuance.

🧠 From understanding to action

Now that you understand what generative AI can do, let's explore how it actually works—and more importantly, how to architect solutions that deliver real business value. This isn't about becoming an AI engineer; it's about knowing enough to make smart decisions, have informed conversations with technical teams, and identify the right approach for each client challenge.

Martha Moore

Executive Assistant

Organizes writing contests and always has salsa music playing.

Efficient and decisive, always one step ahead in the organization. She loves creative writing and dancing. She has published short stories and dances salsa in an amateur company.

close

Emily Taylor

Graphic Designer

Decorates desk with exotic plants and draws caricatures of colleagues.

Stands out for creativity and artistic skills, bringing freshness to each project. Loves sculpture and gardening. Has won several awards for sculptures.

close

Wildly different outputs

Your prompt is under-constrained. AI is making assumptions or interpreting ambiguity differently each time. This is a prompt quality issue, not an AI failure.

NAME PROJECT

CONTEXUALIZE YOUR TOPIC

We don't like to bore. We don't want to be repetitive. Communicating as always is boring and doesn't engage. We do it differently. We sabotage boredom. We create what the brain likes to consume because it stimulates it.

📚Learning Objectives

  1. Map the six layers of the generative AI technology stack
  2. Identify where AND Digital creates value in the AI ecosystem
  3. Apply a decision framework for choosing between buy, compose, or build approaches
  4. Recognize and mitigate five critical business risks with generative AI

With Network Rail, we split large technical documents into smaller, retrievable sections. This kept answers relevant while cutting analysis costs by more than 90% compared with a naïve “ingest everything” approach.

Consistent outputs

minimal variation

Your prompt has sufficient constraints. AI understands the boundaries and produces reliable results within them.

Lucía Pérez

Marketing Manager

Start the day with yoga and bring homemade desserts to the team.

Create innovative and effective campaigns, standing out for creativity. Enjoys painting and yoga. Has exhibited in local galleries and is a certified yoga instructor.

close

Sophia White

Product Development Manager

Brings experimental culinary creations and organizes cycling outings on weekends.

Innovative and efficient, always improving processes and products. Enjoys mountain biking and experimental cooking. Participates in mountain biking competitions.

close

James Smith

Senior Software Engineer

Organizes chess games and shares chapters of his novel for feedback.

Creative and resourceful, always seeking effective solutions. Enjoys chess and science fiction. Has won chess tournaments and is writing a science fiction novel.

close