AI in Tourism — Day 1

AI in Tourism

Day 1 — Technical Foundations, Strategy and Prompting

March 2, 2026

Opening

Learning Objectives

By the end of the day you will be able to:

Explain how an LLM works (tokenisation, transformer, generation)
Distinguish fine-tuning, RAG and prompting and choose the right approach
Understand what AI agents are and their current limitations
Write structured and effective prompts for tourism use cases
Identify concrete opportunities in your business
Apply the basic rules of security and GDPR compliance

Opening

Course Context

About me

PhD in Physics, AI Engineer
AI Consultant and Trainer
Focus on AI applications for SMEs

How we work today

Theory + concrete examples
Questions at any time
Practical exercises in the afternoon

Goal: understand how AI works and make informed choices.

Opening

Day Schedule

Time	Part	Session	Topics
10:30 – 11:00	Morning	Course Introduction	AI in Tourism: Overview Definitions & Common Myths The Paradigm Shift
11:00 – 11:30	Morning	First Part	Large Language Models Tokenisation & Transformer Context Window
11:30 – 11:45	☕ Break
11:45 – 13:00	Morning	Second Part	Fine-tuning RAG — Retrieval Augmented Generation AI Agents Tools & Function Calling
13:00 – 14:00	🍽 Lunch Break
14:00 – 14:55	Afternoon	First Part	Identifying Opportunities Framework & Impact/Effort Matrix Use Cases by Sector
14:55 – 15:05	☕ Break
15:05 – 15:50	Afternoon	Second Part	Effective Prompting Templates & Practical Examples Security & Privacy (GDPR)
15:50 – 16:00	Closing	Wrap-up	Summary Q&A

Opening

The Day's Thread

🔍

Understand

AI Overview in Tourism
How an LLM works
Context Window

⚙️

Customise

Fine-tuning
RAG

🤖

Automate

AI Agents
Tools & Functions
Practical Tools

🎯

Apply

Opportunities in Tourism
Effective Prompting
Security & GDPR

Section 1

AI in Tourism: Overview

Definitions · State of the Art · Common Myths · The Paradigm Shift

Sec. 1 · Overview

AI, ML, Deep Learning, Generative AI

AI — systems that perform tasks typically done by humans
ML — learns from data, without hand-written rules
Deep Learning — ML with deep neural networks
Generative AI — produces text, images, audio, code

Sec. 1 · Overview

AI in Tourism: where we stand today

Hotels & OTAs

Predictive revenue management
Booking chatbots
Automated review responses

Destinations

Personalised itineraries
Tourist sentiment analysis
Flow forecasting

Travel Agencies

Automated search and comparison
Generated travel proposals
After-sales support

Adoption is still uneven: large OTAs are years ahead of SMEs

Sec. 1 · Overview

What AI is not

Myth	Reality
"AI is always accurate"	It produces convincing errors (hallucinations)
"It understands everything like a human"	It predicts the next token — it doesn't reason
"It is autonomous and unpredictable"	It follows statistical distributions over data
"It will replace everyone"	It automates repetitive tasks, not entire roles
"It is too complex for SMEs"	APIs and no-code tools lower the barrier to entry

Sec. 1 · Overview

The Paradigm Shift

Research AI (pre-2020)

Specialised models for a single task
Required large labelled datasets
Accessible only to large companies
Output: classifications, numerical predictions

Generative AI (today)

Foundation models: one model for hundreds of tasks
Instructed via natural language
APIs accessible to anyone
Output: text, images, code, audio

The barrier to entry has dropped dramatically: a well-written prompt replaces months of ML development.

Sec. 1 · Overview

Challenge

In your business, which process takes up the most time each week and could benefit from generative AI? Think of tasks that repeat, involve text or communication, and have a definable output.

forms.gle/quizDay1

Section 2

Large Language Models

How large language models work

Sec. 2 · LLM

What is an LLM

A Large Language Model is a neural network trained to predict the next token in a sequence of text.

Input: "The best hotel in Rome for families is" ↓ [LLM: billions of parameters] ↓ Output: "the" → "Grand" → "Hotel" → "Palatino" → ...

"Large" = billions of parameters (GPT-4: ~1.8T estimated)
"Language" = operates on text (and now also images, audio)
"Model" = a mathematical function learned from data

There is no understanding, no intentionality: it is statistical prediction at enormous scale.

Sec. 2 · LLM

Tokenisation: text becomes numbers

Before processing text, the LLM converts it into tokens (word fragments).

"Reservation" → ["Res", "erv", "ation"] → [4921, 1203, 3847] "ChatGPT" → ["Chat", "G", "PT"] → [10314, 38, 2898] "€" → ["€"] → [27975]

~1 token ≈ 0.75 words in English / ~0.6 words in Italian
API cost is measured in tokens (input + output)
Rare or technical words → more tokens → more expensive

Interactive tool: platform.openai.com/tokenizer

Sec. 2 · LLM

The Transformer architecture (simplified)

Input tokens ↓ [Embedding: token → numeric vector] ↓ [Attention: each token "looks at" all others] ↓ [Feed-forward: non-linear transformation] ↓ × N blocks ↓ [Output: probability distribution over vocabulary] ↓ Token with highest prob → generated text

The self-attention mechanism allows the model to understand context: "bank" in "river bank" vs "bank account".

Sec. 2 · LLM

Pre-training: learning to predict

What happens

Input: trillions of words from the internet, books, code
Objective: predict the next token
Duration: weeks on thousands of GPUs
Cost: tens to hundreds of millions of $

What it learns

Grammar and syntax
Facts about the world
Basic logical reasoning
Different writing styles

The model "knows" only what was in the training data. It has no knowledge of recent news or proprietary company data.

Sec. 2 · LLM

Post-training: from predictor to assistant

Base model (prediction only) ↓ [Supervised Fine-Tuning on instructions] ↓ Instruction-tuned model (follows commands) ↓ [RLHF: human feedback + reward model] ↓ Aligned model (helpful, safe, honest)

SFT — (instruction → ideal response) pairs written by humans
RLHF — human raters compare responses, the model learns what they prefer
Result: the model "converses" instead of completing sentences

Sec. 2 · LLM

Generation parameters

Parameter	Effect	Example use
`temperature`	0 = deterministic, 1 = creative, >1 = chaotic	0.2 for FAQs, 0.9 for copywriting
`top_p`	Truncates to top P% of cumulative probability	0.9 safe default
`max_tokens`	Limit of generated tokens	150 for tweets, 2000 for articles
`stop`	String that stops generation	"\n\n" for short responses

For deterministic tasks (classification, data extraction): temperature = 0.

Sec. 2 · LLM

The leading models today

The up-to-date model rankings are maintained by a third-party site — artificialanalysis.ai

Sec. 2 · LLM

Closed vs open source models

Proprietary (third-party APIs)

No infrastructure to manage
Automatic updates
Variable cost per token
Data sent to the provider
Example: OpenAI, Anthropic, Google

Open-source / open-weight

Full data control
Fixed cost (hardware)
Requires technical expertise
Full customisation
Example: LLaMA, Mistral, Phi

For a tourism SME: start with APIs, consider open-source only if you have privacy requirements or very high volumes.

Sec. 2 · LLM

Consumer-grade interfaces vs API

Consumer-grade interfaces (ChatGPT, Claude.ai, Gemini)

Manual use, session by session
No integration with business systems
Suitable for experimentation and individual use
Data not used for training (paid plan)

API (OpenAI API, Claude API, etc.)

Integrable into apps, CRM, PMS, website
Automatable and scalable
Requires technical development (or no-code)
Pay-per-use billing

Sec. 2 · LLM

How to choose a model: criteria for SMEs

Task quality — use top models only if the task is critical (e.g. responses to VIP guests)
Cost per call — check the pricing: input and output tokens have different rates
Latency — real-time chatbots require fast models (e.g. GPT-4o mini, Claude Haiku)
Language support — test explicitly: not all models perform equally in non-English languages
Data compliance — verify DPA, data residency (EU), terms of use

Practical rule: use the smallest/cheapest model that solves your problem.

Sec. 2 · LLM

Challenge

If you were to introduce an AI assistant to your team starting this week, for which specific activity would you use it? Would you choose a consumer interface (ChatGPT, Claude.ai) or an API integration — and why?

forms.gle/quizDay1

Section 3

Context Window

The model's working memory: limits and practical implications

Sec. 3 · Context Window

What is the context window

The context window is the maximum amount of text the model can "see" in a single processing moment.

┌─────────────────────────── Context Window ──────────────────────────────┐ │ System prompt │ Previous conversation │ Documents │ Latest question │ └─────────────────────────────────────────────────────────────────────────┘ ↑ everything must fit in here ↑

Model	Context window	≈ English words
GPT-4o	128,000 tokens	~96,000
Claude 3.5 Sonnet	200,000 tokens	~150,000
Gemini 2 Pro	2,000,000 tokens	~1,500,000
GPT-4o mini	128,000 tokens	~96,000

Sec. 3 · Context Window

What happens when the context is full

The model cannot process more text: the API call returns an error
Some implementations "trim" the oldest parts of the conversation
The model "forgets" the trimmed information: no native persistent memory
Response quality degrades if the context is very long (lost-in-the-middle)

A 100-page PDF can easily saturate 128k token contexts. Always check before sending long documents.

Sec. 3 · Context Window

Managing context in real applications

Software strategies

Progressive summarization — summarize the chat every N turns
Sliding window — keep only the last K exchanges
RAG — retrieve only the relevant pieces (→ section 5)
Entity extraction — save important facts to a DB

Cost implications

Input tokens are charged on every call
Long conversation = growing cost
Keeping the system prompt short matters
Models with caching (Claude) reduce costs

🏨

Hotel Bellavista Assistant

● online

I booked a double room, March 14–17.

Found it! Garden-view double room. How can I help you?

Can I have an extra pillow?

Of course, I've noted the request.

What time is breakfast?

Buffet 7:00–10:30, included in your rate.

↕ context window

What about parking? I have a car.

Open-air parking is free. Covered garage €15/night — shall I book it?

Yes please. Can I do a late check-out on the 17th?

Garage added. I'll check late check-out availability and contact you this evening.

Perfect. Do you have a spa?

Yes, spa on the 2nd floor open 9:00–21:00. Treatments by appointment, complimentary for guests.

Type a message…

➤

Sec. 3 · Context Window

Context window in tourism

Scenario	Estimated tokens	Strategy
Booking chatbot (short chat)	1,000–5,000	No problem, mini model
Analysis of 50 reviews	~20,000	Batch, one analysis at a time
Q&A on full catalogue (500 products)	>100,000	RAG required
Single reply to a guest	200–800	Direct API call, inexpensive

Sec. 3 · Context Window

The trend: growing context windows

2023: GPT-4 with 8k → 32k tokens
2024: Claude with 200k, Gemini with 1M tokens
2025: Gemini 2.0 with 2M tokens (~1,500 pages)
Direction: "unlimited" context window in the long term

Larger contexts do not eliminate the need for RAG: attention quality on long texts remains an open problem, and cost is proportional.

Do not confuse "can process 1M tokens" with "processes them accurately": always validate on your specific use case with appropriate benchmarks.

Sec. 3 · Context Window

Challenge

Which business documents (price list, cancellation policy, FAQ, operations manual) would you want an AI assistant to consult when answering guests? Are they already in digital format and kept up to date?

forms.gle/quizDay1

Section 4

Fine-tuning

Training a model on your own data

Sec. 4 · Fine-tuning

The adaptation continuum

Zero-shot Few-shot Fine-tuning │ │ │ ▼ ▼ ▼ "Classify "Here are 3 examples, Train the model this now classify" with thousands of review" my own examples │ │ │ No data A few examples Dataset required Immediate in the prompt Time + cost

For most tourism use cases, prompting and few-shot are sufficient. Fine-tuning only when you have a concrete problem that cannot be solved otherwise.

Sec. 4 · Fine-tuning

Training types compared

Phase	Data	Who does it	Cost
Pre-training	Trillions of tokens from the internet	Labs only (OpenAI, Anthropic…)	$10M–$100M+
Supervised FT	A few hundred to thousands of examples	Companies, developers	$100–$10,000
RLHF / RLAIF	Human or AI preferences	Mainly the labs	High
Prompting	Examples in the text	Anyone	API call cost

Sec. 4 · Fine-tuning

Full fine-tuning vs LoRA

Full fine-tuning

Updates all model parameters
Maximum adaptation capacity
Requires high-end GPUs
Risk of "forgetting" base knowledge (catastrophic forgetting)

LoRA / PEFT

Adds small adapters to the base model
Trains only a fraction of parameters (<1%)
Consumer-grade GPUs are sufficient
The base model remains intact
The dominant technique today for local fine-tuning

Sec. 4 · Fine-tuning

When fine-tuning is worth it

The model must adopt a very specific communication style (brand tone, register)
Repetitive task with a rigid output format (e.g. structured JSON, specific forms)
Long prompt that repeats every time: fine-tuning "bakes in" the instructions
You have hundreds of high-quality examples already validated

Do not fine-tune to teach the model new information: use RAG for that. Fine-tuning teaches how to respond, not what to know.

Sec. 4 · Fine-tuning

Requirements: data, costs, infrastructure

Data

Min. 50–100 examples (quality > quantity)
Format: JSONL with prompt/completion
Must be clean and representative

Cost (API)

OpenAI FT: ~$8 / 1M training tokens
One-time training + ongoing inference
Model hosting may have a fixed cost

Infrastructure

Via API: no infrastructure needed
Self-hosted: GPU with 16–80 GB VRAM
Cloud ML: Vertex AI, SageMaker, RunPod

Sec. 4 · Fine-tuning

Example: specialised chatbot for a hotel

Without fine-tuning

Prompt: "You are the Hotel Bellavista assistant.
Reply in a warm and professional manner.
Here are our policies: [500 tokens]
Here are our services: [800 tokens]
Guest question: late check-out?"

~1,400 tokens of context on every call

With fine-tuning

Prompt: "Guest question:
late check-out?"

Tone, policies and style already baked into the model

~30 tokens of context on every call

Savings on input tokens = direct long-term savings on API costs.

Sec. 4 · Fine-tuning

Fine-tuning vs prompting: the choice

Criterion	Use prompting	Consider fine-tuning
Time available	Available immediately	Weeks for data + training
Request volume	Low / medium	Very high (context cost)
Output style	Generic or semi-specific	Very rigid or branded
Available data	Few or none	Tens to hundreds of examples
Technical skills	None required	Developer or ML engineer

Sec. 4 · Fine-tuning

Challenge

Does your company have a recognisable communication style towards guests? Have you already collected examples of emails or replies that represent your standard — and how many could you put together in a month?

forms.gle/quizDay1

Section 5

RAG — Retrieval Augmented Generation

Connecting an LLM to your data without training it

Sec. 5 · RAG

The problem: the LLM doesn't know your data

What the LLM doesn't know

Your updated price list
Your hotel's specific policies
Real-time availability
Internal company documents
News from the past few months

The (bad) alternatives

Put everything in the prompt → context explodes, high costs
Fine-tuning → doesn't learn updatable facts, only style
Ignore the problem → the model hallucinates

RAG = retrieve the relevant documents at question time, insert them into the context, generate the answer.

Sec. 5 · RAG

RAG Architecture — Core concept

Sec. 5 · RAG

RAG Architecture — Application diagram

[Business documents] ──→ [Chunking] ──→ [Embedding] ──→ [Vector DB] │ │ (indexing, once) │ User question ──→ [Question embedding] ──→ [Similarity search] │ Top-K relevant chunks │ ▼ [Prompt = question + relevant chunks] ──→ [LLM] ──→ Answer

Sec. 5 · RAG

Embeddings: text as coordinates

An embedding model converts text into a vector of numbers that represents its meaning.

"room with sea view" → [0.12, -0.87, 0.45, 0.33, ...] "ocean view bedroom" → [0.11, -0.85, 0.46, 0.31, ...] ← similar! "breakfast included" → [0.78, 0.23, -0.91, 0.55, ...] ← different

Texts with similar meaning → nearby vectors in the space
Cosine similarity measures how semantically related two texts are
Embedding models: OpenAI text-embedding-3, Cohere, sentence-transformers

Sec. 5 · RAG

Embeddings: the vector space

Sec. 5 · RAG

Embeddings: word vectors

Word vector representation and dimensionality reduction

Sec. 5 · RAG

Vector database: what it is and how it works

Traditional database

Searches for exact words

SELECT * FROM faq
WHERE text LIKE '%check-out%'

Does not find "departure time"

Vector database

Searches by semantic meaning

query = embed("what time do I need to leave the room?")
results = db.search(query, top_k=3)

Finds check-out, departure, late check-out…

Common vector database solutions: Pinecone, Weaviate, Chroma, pgvector (PostgreSQL), Qdrant

Sec. 5 · RAG

RAG pipeline: step by step

Document loading — PDF, Word, database, website
Chunking — split into pieces of 200–500 tokens with overlap
Embedding — convert each chunk into a vector
Indexing — save the vectors to the vector DB
— runtime —
Question embedding — convert the user question into a vector
Search — find the top-3/5 most similar chunks
Augmentation — insert the chunks into the prompt
Generation — the LLM responds based on the chunks

Sec. 5 · RAG

RAG in tourism: practical examples

Dynamic hotel FAQ

Operations manual → chunked
Chatbot answers on policies, hours, services
Update: upload new PDF

Tour catalogue

500 tour sheets → indexed
"Tour suitable for children with snorkeling" → top-3 tours
Response with precise details

Prices and availability

Nightly update from the PMS
Chatbot responds with fresh data
No risk of outdated information

Sec. 5 · RAG

RAG vs fine-tuning: comparison table

Aspect	RAG	Fine-tuning
Data updates	Easy (reload documents)	Expensive (retrain)
Transparency	High (source is visible)	Low (data baked in)
Factual hallucinations	Reduced (chunks as anchor)	Present
Output style	Does not change model behaviour	Modifies style and behaviour
Technical setup	Medium (vector DB, pipeline)	Medium-high (dataset, GPU)

In practice the two are often combined: fine-tuning for style + RAG for up-to-date data.

Sec. 5 · RAG

RAG limitations

Chunking quality — poorly split chunks = fragmented context = wrong answer
Multi-hop questions — "compare the price of tour X with tour Y" requires multiple retrievals
Non-textual data — complex tables, images, PDFs with complex layout are difficult
Added latency — vector search adds 50–300ms per query
Hallucination not eliminated — the model can still ignore the provided context

RAG is not a magic wand: it requires careful engineering and continuous evaluation of response quality.

Sec. 5 · RAG

Challenge

If you could ask an AI assistant trained on your business documents a question and receive a precise, up-to-date answer, what would that question be? Which documents would you give it access to?

forms.gle/quizDay1

Section 6

AI Agents

Autonomous systems that plan and act in the real world

Sec. 6 · AI Agents

The difference: LLM, Assistant, Agent

🧠

Pure LLM

Predicts the next token
No memory
No tools

          Input → Output
        

💬

Assistant

Responds in natural language
Maintains context
Follows system instructions

          Input → Output
        

🤖

Agent

Plans complex tasks
Uses tools and executes actions
Observes results and iterates

          Goal → Plan → Actions → Goal
        

An agent uses an LLM as its reasoning engine, but has operational autonomy.

Sec. 6 · AI Agents

The ReAct cycle: Reason → Act → Observe

Goal: "Find the cheapest flight Rome-Tokyo for July and add it to the calendar" [Reason] → I need to search for flights in July [Act] → search_flights("Rome", "Tokyo", "July") [Observe] → {results: [{price:650, date:"Jul 5"}, ...]} [Reason] → The cheapest is €650 on July 5th [Act] → create_calendar_event("July 5", "Tokyo flight") [Observe] → {status: "created", id: "evt_123"} [Reason] → Task completed [Output] → "Found flight €650 (Jul 5) and created the event"

Sec. 6 · AI Agents

Planning and memory in agents

Memory types

In-context — everything in the current conversation
External — databases, files, persistent vectors
Procedural — instructions in the system prompt
Episodic — logs of previous sessions

Planning types

Chain-of-thought — step by step in the text
Tree-of-thought — explores multiple paths
Plan-and-execute — full plan then execution
ReAct — interleaves reasoning and actions

Sec. 6 · AI Agents

Multi-agent systems

👤 "Plan a weekend in Florence for 2 people, budget €400"

↓

🎯 Orchestrator

↓

✈️ Flights Agent

API Skyscanner

↓

🏨 Hotel Agent

PMS + Booking API

↓

🗺️ Itinerary Agent

LLM + local RAG

↓

✅ Full plan: flights + hotel + itinerary within budget

Key concepts

Specialisation — each agent has a precise domain and dedicated tools
Orchestration — the orchestrator assigns tasks, collects results and resolves conflicts
Parallelism — multiple agents work simultaneously, reducing latency
Composability — sub-agents are reusable in other workflows

Frameworks

LangGraph — agent graphs with explicit state
CrewAI — agents with roles and goals in natural language
AutoGen — multi-agent conversations (Microsoft)
Claude MCP — standard protocol for tools and agents

Sec. 6 · AI Agents

AI Agents in tourism

🏨 Automated booking

Receives request, checks PMS availability, applies rate policies, generates confirmation and sends email to the guest

💰 Revenue optimisation

Analyses history, local events and competitors → automatically proposes the weekly rate plan to the revenue manager

🔍 Competitor monitoring

Scans OTAs every hour, compares rates with your own, notifies the team if the gap exceeds a defined threshold

⭐ Review management

Monitors TripAdvisor, Google, Booking in real time, drafts personalised replies and sends them after human approval

🗺️ Traveller assistant

Plans tailor-made itineraries, books restaurants and activities, answers pre- and post-stay questions in natural language

📊 Operations reports

Aggregates data from PMS, CRM and OTA channels every morning → generates KPI report with highlighted anomalies, ready for the briefing

Agents in production require human oversight on high-impact decisions (e.g. rate changes, public replies, sending communications).

Sec. 6 · AI Agents

Risks and limitations of agents

Cascading errors — a mistake in one step is amplified in subsequent steps
Infinite loops — the agent does not terminate if it cannot find the stop condition
Prompt injection — malicious external data crafted to hijack the agent's behaviour
Unpredictable costs — every iteration consumes tokens: complex tasks cost a lot
Irreversible actions — sending emails, modifying databases, spending money

Rule: every irreversible action must have an explicit human approval checkpoint.

Sec. 6 · AI Agents

When are agents reliable today?

Scenario	Reliability	Notes
Simple tasks with clear tools	High	1-3 steps, well-defined tools
Information search and synthesis	High	No irreversible actions
Multi-step workflow (5-10 steps)	Medium	Requires supervision
Open-ended tasks, ambiguous goal	Low	Not recommended in production
Autonomous financial actions	Not recommended	Human supervision always

Sec. 6 · AI Agents

Challenge

For which operations would you be willing to authorise an AI agent to act autonomously — changing a rate, sending a notification, updating a booking — and where would you always want to keep human approval?

forms.gle/quizDay1

Section 7

Tools and Function Calling

How an LLM executes actions in the real world

Sec. 7 · Tools

From the LLM to the real world

An LLM on its own only produces text. To do something, it needs tools.

User: "What is the availability of the suite on August 15th?" LLM recognises: I need to call the "check_availability" tool ↓ Generates structured JSON: {"tool": "check_availability", "params": {"room": "suite", "date": "2026-08-15"}} ↓ The application code executes the call to the PMS ↓ The result comes back to the LLM as text ↓ LLM formulates the final response to the user

Sec. 7 · Tools

Function calling: defining tools for the LLM

Each tool is defined with a schema that the model uses to generate correct calls.

{
  "name": "check_availability",
  "description": "Checks room availability in the PMS",
  "parameters": {
    "room_type": {
      "type": "string",
      "enum": ["standard", "deluxe", "suite"],
      "description": "Requested room type"
    },
    "arrival_date": {
      "type": "string",
      "format": "date",
      "description": "Arrival date (YYYY-MM-DD)"
    },
    "nights": {
      "type": "integer",
      "description": "Number of nights"
    }
  },
  "required": ["room_type", "arrival_date", "nights"]
}

Sec. 7 · Tools

Tool examples in tourism

Tool	Action	Integrated system
`check_availability`	Check available rooms	PMS (Opera, Mews, Cloudbeds)
`create_booking`	Book and confirm	PMS + Channel Manager
`get_weather`	Weather forecast for destination	Weather API
`send_notification`	Email/SMS to the guest	CRM, Mailchimp, Twilio
`update_rate`	Update prices	Revenue management system

Sec. 7 · Tools

Multi-agent architecture

👤 "Book a double room for Saturday, dinner included, gluten-free guest — send them a confirmation"

▼

🧠 Orchestrator

▼

🛏️ Availability

check_rooms

create_booking

update_calendar

calculate_rate

▼

PMS · Booking.com · Airbnb

🍽️ Restaurant

check_tables

book_table

log_allergies

alert_kitchen

▼

Back-office · TheFork

🗺️ Hospitality

get_weather

local_events

suggest_activities

RAG_local_guide

▼

Maps · OpenWeather · FAQ

📬 Communication

send_confirmation

send_reminder

update_CRM

request_review

▼

Email · WhatsApp · CRM

Sec. 7 · Tools

MCP: Model Context Protocol

MCP is an open standard (Anthropic, 2024) for connecting LLMs to tools and data sources in an interoperable way.

Before MCP

Every integrator builds their own format
Tools not reusable across different models
Multiplied maintenance

With MCP

Standard MCP server for each system
Works with Claude, GPT-4, Gemini…
Ecosystem of ready-made servers (Google Drive, GitHub, databases…)

MCP is still being adopted but represents the direction the industry is moving towards.

Sec. 7 · Tools

Classic function calling: context overhead

Example: web search + article generation (3 sources) [LLM] → {"tool":"web_search", "query":"AI news"} ← round-trip 1 [LLM] ← {urls: [url_1, url_2, url_3]} → ctx +tok_0 [LLM] → {"tool":"web_fetch", "url": url_1} ← round-trip 2 [LLM] ← {content: "…full article…"} → ctx +tok_1 [LLM] → {"tool":"web_fetch", "url": url_2} ← round-trip 3 [LLM] ← {content: "…full article…"} → ctx +tok_2 [LLM] → {"tool":"web_fetch", "url": url_3} ← round-trip 4 [LLM] ← {content: "…full article…"} → ctx +tok_3 [LLM] → {"tool":"write_blog", "sources": [c_1, c_2, c_3]} ← round-trip 5

Every intermediate result enters the context: the model must process each response before proceeding
With N sources: 2N + 1 round-trips, context usage O(N)
Saturated context → performance degradation (lost-in-the-middle); latency and cost grow linearly

Sec. 7 · Tools

Programmatic Tool Calling

Classic function calling

→ web_search("AI news") [ctx +tok_0] ← {urls: [...]} → web_fetch(url_1) [ctx +tok_1] ← {content: "..."} → web_fetch(url_2) [ctx +tok_2] ← {content: "..."} → write_blog(c_1, c_2, c_3) 5 round-trips · all content in context

Programmatic Tool Calling

# The model generates this code:
results = web_search("AI news")

content_list = []
for r in results:
    content = web_fetch(r.url)
    content_list.append(content)

write_blog(content_list)

1 round-trip · only the final output re-enters the LLM context

Instead of emitting a JSON for each tool call, the model generates code (Python / TypeScript) in which tools are callable functions. The code is executed in the execution environment; intermediate results reside in local variables, not in the model's context. Available from Claude Opus 4.5 and Sonnet 4.5 onwards.

Sec. 7 · Tools

Code Execution: enabling and impact

How to enable it (Anthropic API)

tools = [
  # 1. Add the execution environment
  {"type": "code_execution_20250522",
   "name": "code_execution"},

  # 2. Mark tools callable from code
  {
    "name": "query_database",
    "description": "...",
    "input_schema": {...},
    "allowed": ["code_execution"]
  }
]

The code_execution tool provides the sandbox; allowed exposes the function as callable in the environment.

Unlocked capabilities

Batch processing — for loops over N elements, executed in parallel
Conditional tool selection — if/else on the result of a tool to decide the next tool
Local data filtering — transformation and aggregation of results before inserting them into context
Round-trip reduction — from 2N+1 calls to 1 (request) + 1 (final response)

Measured reduction: 37–50% of tokens for equivalent tasks.
Anthropic benchmark: context at equivalent task approximately 4× lower compared to classic function calling.

Sec. 7 · Tools

Advanced optimisations for high-efficiency agents

Dynamic Web Filtering

The system automatically filters the content of web pages retrieved during web search
Only semantically relevant sections for the query enter the context
Input token reduction: 24% on average (range 33–62% depending on content)
Activation: add the web_search tool — filtering is applied automatically

Tool Search

With many MCP servers, loading all definitions into context is prohibitive
A special search_tool searches for relevant definitions on demand, via regex or BM25 query
Savings: ~95% of context occupied by tool definitions
Deferred loading: the definition is loaded only when needed and removed afterwards, freeing space

Tool Use Examples

Add input examples to the tool definition (few-shot approach at schema level)
Five examples are sufficient in most cases
Accuracy in correct tool invocation: 70% → 90%
Particularly useful for tools with many parameters or complex structured inputs

The three optimisations are orthogonal and cumulative: combining them simultaneously reduces latency, token consumption and error rate in agents with a high number of tools.

Sec. 7 · Tools

Challenge

Which systems do you use today in your business (PMS, CRM, channel manager, email)? Which integration with an AI assistant would generate the highest immediate value — and why?

forms.gle/quizDay1

Break

Lunch

Resuming in the afternoon

Section 7b

AI Tools: practical guide

ChatGPT, Claude, Copilot and Gemini — access, costs and first hands-on activities

Sec. 7b · AI Tools

The consumer tools landscape

Tool	Company	Key strength	Free?
ChatGPT	OpenAI	General purpose, the best-known; images (DALL-E 3)	Yes (limited)
Claude.ai	Anthropic	Long-form writing, PDF document analysis	Yes (limited)
Microsoft Copilot	Microsoft	Integrated in Outlook, Teams, Excel, Word	Yes (web)
Google Gemini	Google	Integrated in Gmail, Docs, Drive, Sheets	Yes (basic)

All four work via browser — no installation required. The free tier is sufficient to start experimenting.

Sec. 7b · AI Tools

ChatGPT — chatgpt.com

For tourism

Responding to positive and negative reviews
Writing social posts (Instagram, Facebook)
Creating personalised itineraries on request
Translating communications into multiple languages
Generating room and package descriptions

Plans

Free — GPT-5.2 limited
Plus — $20/month · GPT-5.2 unlimited, files, images

First activity:

Go to chatgpt.com → paste a recent negative review → write:

"Reply in a professional and empathetic tone, in 100 words or fewer."

Do not enter the guest's name, email or personal data in the prompt.

Sec. 7b · AI Tools

Claude.ai — claude.ai

For tourism

Analysing long documents (contracts, policies, PDFs)
Answering questions about the content of a document
Writing articulate texts in the brand's tone of voice
Summarising feedback and identifying patterns
Reasoning through complex scenarios (pricing decisions, offers)

Plans

Free — Claude Sonnet, Haiku (limited daily use)
Pro — $18/month · Claude Opus, extended use, files, Projects

First activity:

Go to claude.ai → upload your cancellation policy PDF → write:

"If a guest cancels 8 days in advance, are they entitled to a refund? Answer clearly and concisely."

The Projects feature (Pro plan) saves a fixed context — policy, tone, instructions — without having to rewrite it for every conversation.

Sec. 7b · AI Tools

Microsoft Copilot — copilot.microsoft.com

For tourism

Outlook — draft welcome, pre-arrival and follow-up emails
Teams — meeting summary with action items
Excel — booking data analysis, occupancy charts
Word — standard contracts, commercial proposals
Web (free) — search with sources, real-time answers

Plans

Free — copilot.microsoft.com (GPT-5.1 + web search)
M365 Copilot — $30/user/month · integrated in Office apps

First activity (free):

Go to copilot.microsoft.com → write:

"Write a welcome email for a guest arriving tomorrow at my boutique hotel in Florence. Warm tone, max 80 words."

If you already use Microsoft 365, check with your IT provider whether the M365 Copilot plan can be activated on your tenant.

Sec. 7b · AI Tools

Google Gemini — gemini.google.com

For tourism

Gmail — draft replies to booking enquiries
Google Docs — generating proposals, local guides, press releases
Google Sheets — data analysis with natural language
Real-time web search (competitor prices, local events)
Image analysis — property photos, menus, decor

Plans

Free — Gemini 2.5 Flash on gemini.google.com
Google One AI Premium — €21.99/month · Gemini 3.1 Pro + Workspace

First activity:

Go to gemini.google.com → paste 5–10 recent reviews → write:

"Analyse these reviews: list the 3 most-mentioned strengths and the 2 most-mentioned areas for improvement."

Those already using Google Workspace get the most value from the paid plan, thanks to native integrations with Gmail and Drive.

Sec. 7b · AI Tools

Which tool to choose

If…	Use	Plan
You already use Microsoft 365 (Outlook, Teams, Excel)	Microsoft Copilot M365	$30/user/month
You already use Google Workspace (Gmail, Drive)	Google Gemini	€21.99/month
You want to analyse documents and write articulate texts	Claude.ai	Free to start, Pro ($18/month) for daily use
You want the most versatile option, with image generation	ChatGPT	Free to start, Plus ($20/month) for daily use

You don't have to choose just one. Start free, move to a paid plan only if you use it every day.

Sec. 7b · AI Tools

Challenge

Among the four tools presented (ChatGPT, Claude, Copilot, Gemini), which would you try first in your business — and for which specific task?

forms.gle/quizDay1

Section 8

Identifying Opportunities

Where AI creates value in your tourism business

Sec. 8 · Opportunities

Framework: mapping repetitive processes

The starting point for identifying AI opportunities is a map of repetitive business processes.

Signals that a task is suitable for AI

It repeats more than 10 times a day
It requires text / language processing
It has a definable and verifiable output
A new team member would learn it in < 1 hour
It does not require deep empathy or a personal relationship

Signals that a task is NOT suitable

Each case is unique and complex
It requires ethical or legal judgement
It involves a human relationship as the core value
Errors have serious consequences
Pure structured data (use a traditional database instead)

Sec. 8 · Opportunities

Impact / effort matrix for AI

high

IMPACT

low

⚡ Quick wins

Do it now

🎯 Strategic

Plan carefully

⏳ Defer

Low priority

✗ Avoid now

High cost, low return

low

EFFORT →

high

Quick wins: review responses, social post generation, FAQ chatbot, guest email translation
Strategic: AI revenue management, automated bookings, RAG on catalogue, multichannel agent

Sec. 8 · Opportunities

Suitable vs not suitable for generative AI

Suitable	Not suitable (or not yet)
Responding to template emails with personalisation	Managing complex disputes with guests
Generating room / package descriptions	Legal decisions (refunds, damages)
Summarising feedback and reviews	Physical operational supervision
Proposing personalised itineraries	Relationships with high-value VIP clients
Translating multilingual communications	HR performance evaluations of staff

Sec. 8 · Opportunities

Use cases by sector

Hotel / B&B

Booking and FAQ chatbot
Automated check-in/out via chat
Automated review responses
Personalised pre-arrival upsell
Monthly feedback summary

Restaurants

Generating menu descriptions
Responding to bookings and dietary requests
Automated social posts
TripAdvisor review analysis

Travel agencies / DMC

Personalised itinerary proposals
Supplier research and comparison
Automated follow-up emails
Client briefing processing

Sec. 8 · Opportunities

Exercise: identify 3 processes in your business

10 minutes — individual work

Think about the tasks you (or your team) carry out repeatedly every week
For each one, assess: High/Medium/Low frequency × High/Medium/Low impact if automated
Position it on the impact/effort matrix

Task	Frequency	Impact	Estimated effort
1.
2.
3.

Sec. 8 · Opportunities

Challenge

You have completed the impact/effort matrix. What is your "quick win" — the high-impact, low-effort process you could automate within the next four weeks? What is the first concrete step?

forms.gle/quizDay1

Section 9

Effective Prompting

Communicating with AI models precisely and systematically

Sec. 9 · Prompting

Anatomy of a prompt: 5 components

┌──────────────────────────────────────────────────────────┐ │ ROLE "You are a tourism marketing expert..." │ ├──────────────────────────────────────────────────────────┤ │ CONTEXT "The hotel is called Bellavista, 4-star, │ │ family-oriented, Amalfi Coast..." │ ├──────────────────────────────────────────────────────────┤ │ TASK "Write a reply to the review below" │ ├──────────────────────────────────────────────────────────┤ │ FORMAT "Max 120 words, warm tone, no jargon" │ ├──────────────────────────────────────────────────────────┤ │ CONSTRAINTS "Do not promise refunds. Do not mention │ │ competitors." │ └──────────────────────────────────────────────────────────┘

You don't always need all 5: add only what is needed. A longer prompt is not always better.

Sec. 9 · Prompting

Zero-shot, few-shot, chain-of-thought

Zero-shot

No examples, instruction only

"Classify this review
as positive or negative:
'The breakfast was cold'"

Fast, cheap, sufficient for simple tasks

Few-shot

2–5 examples before the task

"Positive: 'Breathtaking view'
Negative: 'Small room'
Classify: 'Cold breakfast'"

Improves precision and output format

Chain-of-thought

Ask for explicit reasoning

"Reason step by step:
1) Identify the problem
2) Assess the tone
3) Give the classification"

Complex tasks, reasoning, calculations

Sec. 9 · Prompting

Prompt templating: reusable structures

A template is a prompt with variables to be filled in at runtime.

You are the guest relations manager of {{property_name}}.

You have received this review on {{platform}}:
---
{{review_text}}
---

Write a professional response that:
- Thanks the guest for their feedback
- Addresses the main point specifically
- {{specific_instruction}}
- Stays under {{max_words}} words

Templates are saved in a shared document and passed to every team member: prompt engineering as a business asset.

Sec. 9 · Prompting

Practical example: social post

Prompt

"You are the social media manager
of Agriturismo Le Colline.

Write 3 Instagram post variants
to promote the wine-tasting weekend
on 15 March.

Format:
- Emoji at the start
- Max 150 characters
- 3–5 relevant hashtags
- Warm, authentic tone"

Expected output

🍷 Saturday, 15 March, join us
among our vineyards for an
unforgettable wine tasting.
#agriturismo #wine #tuscany

🌿 A weekend of flavour and nature:
tasting with a view over the hills.
Limited spots!
#wine #tuscany #weekend

✨ Every sip tells the story of our
land. Join us on 15 March.
#winetasting #wine #italianwine

Sec. 9 · Prompting

Practical example: replying to reviews

Review (negative)

"Disappointing room, different from the photos. The bathroom was small and the WiFi didn't work. I won't be back." — 2 stars

Prompt

"Reply to this negative review.
Tone: professional and understanding.
Acknowledge the specific issues.
Offer a direct point of contact.
Max 100 words. Do not be defensive."

Output

"Dear guest, thank you for your feedback. We are sorry that your experience did not meet your expectations, particularly regarding the room and connectivity. These are aspects we are actively working to improve. We invite you to contact us directly at guests@hotel.it so we can discuss further and, if possible, make it right. We hope to welcome you back for an experience that truly meets your expectations."

Sec. 9 · Prompting

Practical example: personalised itineraries

You are an expert in eastern Sicily tourism.

Create a 5-day itinerary for:
- Couple, aged 35–40
- Interests: Greco-Roman history, local cuisine, moderate walks
- Budget: mid-to-high
- Period: April
- Base: Syracuse
- Self-guided: independent travel by car

For each day indicate:
- Morning / afternoon / evening
- Estimated driving time
- 1 recommended restaurant with a typical speciality
- 1 practical note (booking recommended, opening hours, etc.)

Sec. 9 · Prompting

Practical examples: FAQ and descriptions

Automated FAQ

"Given this cancellation policy:
[policy text]

Answer this guest question:
'If I cancel 10 days in advance
do I get a full refund?'

Be precise, brief, and friendly in tone."

Package description

"Write a commercial description
for this package:
- 3 nights sea-view suite
- Breakfast included
- Unlimited SPA access
- Romantic dinner (1 evening)
- Price: €480 per night

Target: couples, special occasions.
Max 80 words. Evocative tone."

Sec. 9 · Prompting

Common mistakes in prompting

Error	Effect	Solution
Prompt too vague	Generic, unusable output	Specify format, length, tone
No business context	Generic, non-branded response	Add name, sector, values
Only saying what NOT to do	The model doesn't know what to do	Say WHAT you want, not just what to avoid
Multiple tasks in one prompt	Mediocre result on everything	One prompt = one task
No output structure	Unpredictable format	Specify: "reply as a list", "use JSON"

Sec. 9 · Prompting

Systematic prompt debugging

1. Write the initial prompt ↓ 2. Test with 3–5 different inputs ↓ 3. Identify where it fails ├─ Output too long? → add "max X words" ├─ Wrong tone? → specify the role more clearly ├─ Wrong format? → add an example (few-shot) └─ Missing information? → add context ↓ 4. Change ONE element at a time ↓ 5. Retest, compare ↓ 6. Save the working prompt in the template library

Sec. 9 · Prompting

Prompt for review analysis

Analyse the following hotel reviews and produce a structured report.

Reviews:
---
{{reviews_list}}
---

For each review, identify:
1. Overall sentiment: Positive / Neutral / Negative
2. Main themes mentioned (max 3 per review)
3. Implicit score 1–5

Then produce:
- Top 3 recurring strengths
- Top 3 areas for improvement
- 1 concrete recommended action

Output format: structured JSON

With 50 reviews per month, this automated analysis is worth hours of manual work.

Sec. 9 · Prompting

Prompt Engineering vs. Context Engineering

Source: anthropic.com/engineering/…context-engineering

Prompt Engineering

Optimising the text of a single prompt: role, instruction, format, constraints, few-shot examples.

Context Engineering

Designing the entire context passed to the model:

Tools — which tools are available and with what definitions
Retrieved documents — RAG chunks, web results, project files
Conversation history — which turns to keep or compress
Structured outputs — previous tool call results in the context

The conceptual leap: output quality depends more on what is in the context than on the wording of the prompt. Especially relevant for agentic systems.

Sec. 9 · Prompting

Challenge

Have you already used ChatGPT, Claude or similar tools for a work task? Describe a case where the result surprised you — positively or negatively — and what you would have done differently with the prompt.

forms.gle/quizDay1

Section 10

Security and Privacy

GDPR, practical risks and responsible AI use

Sec. 10 · Security

GDPR and AI: obligations and responsibilities

Relevant GDPR principles

Data minimisation — use only necessary data
Purpose limitation — do not reuse data for other purposes
Accountability — you are the data controller
Privacy by design — design with privacy in mind

What changes with AI

Data sent to the API may be used for training (check the terms)
Automated decisions require transparency (Art. 22)
Guest profiling → legal basis required
Data breach includes leaks via prompts

Sec. 10 · Security

Specific risks: what can go wrong

Risk	Description	Impact
Prompt injection	A malicious input alters the model's behaviour	Exposed data, unauthorised actions
Data leakage	The model "remembers" sensitive data inserted in the context	Exposure of client/business data
Model inversion	Attacks that extract data from model parameters	Relevant only for fine-tuning on sensitive data
Hallucinations in factual data	The model invents prices, policies, legal information	Reputational damage, disputes

Sec. 10 · Security

Practical rules: what to never include in a prompt

Never include:

Personally identifiable data (name + email + phone together)
Credit card numbers, banking details, identity documents
Access credentials (passwords, API keys, tokens)
Health data or sensitive guest preferences
Trade secrets or confidential contracts

If in doubt: anonymise first. "Guest X booked…" → "A guest booked…"

For batch processing of client data: always use pseudonymisation and keep the mapping locally.

Sec. 10 · Security

AI services with enterprise guarantees

Enterprise options

OpenAI Enterprise / API — data not used for training, SOC 2
Azure OpenAI — EU data residency, tenant isolation, Microsoft DPA
Anthropic Claude API — data not used for training, robust privacy policy
Google Vertex AI — data residency, GDPR compliance

What to verify

Is data used for training? (no for paid plans)
Where is data processed? (data residency)
Is a DPA (Data Processing Agreement) available?
Certifications: SOC 2, ISO 27001, GDPR?

Free, consumer-grade versions (ChatGPT free) do not offer the same guarantees as paid API plans.

Sec. 10 · Security

Checklist: safe AI use in a tourism SME

I have signed a DPA with the AI provider
Shadow AI policy: staff must not include personal client data in prompts
I use paid plans only (no training on data)
I have updated the data processing register
Privacy policy updated (mention of AI tools)
AI outputs reviewed by a human before sending
No API keys in website / app code
Access to AI tools restricted to necessary users
Auditing: logs of critical AI operations retained
Data breach response plan updated

Sec. 10 · Security

Challenge

In your company, who has access to AI tools and what guest data flows through prompts? Are measures already in place — such as a signed DPA or an internal policy — to ensure safe and GDPR-compliant use?

forms.gle/quizDay1

Section 11

Day 1 Closing

Recap · Resources · Next steps

Sec. 11 · Closing

Recap — Technical foundations

Key concepts

LLM = statistical prediction at enormous scale
Token = basic unit of cost and context
Context window = the model's working memory
Fine-tuning = style and format, not new facts
RAG = your data + LLM without training
Agents = LLM + tools + limited autonomy
Function calling = interface between LLM and real systems

LLM ──→ Context Window │ ├── Prompting (immediate, low-cost) ├── Fine-tuning (style, volume) ├── RAG (up-to-date data) └── Agents (autonomy, risk) Choose the right technique for your specific problem

Sec. 11 · Closing

Recap — Strategy and prompting

Identifying opportunities

Repetitive processes with definable output
Impact/effort matrix
Quick wins first

Effective prompting

5 components: Role, Context, Task, Format, Constraints
Reusable templates
Iterate one element at a time

Security

No sensitive data in prompts
DPA with the provider
Human review of critical outputs

AI is not a substitute for human judgement: it is a productivity amplifier.

Sec. 11 · Closing

Resources for further reading

Reading

The Alignment Problem — Brian Christian
Co-Intelligence — Ethan Mollick
Anthropic Research Blog
OpenAI Research
Claude API Docs — technical guides and in-depth concepts

Online courses

deeplearning.ai — practical courses, free
fast.ai — applied ML
Prompt Engineering Guide

Tools to explore

Claude.ai — Anthropic interface
ChatGPT — OpenAI interface
Perplexity — AI-powered search
Make / n8n — no-code AI automations
Dify / Flowise — no-code RAG and chatbots

Community

r/LocalLLaMA (Reddit)
Hugging Face community

Today's slides available at this link.

Questions?

Thank you for today.

See you on Day 2.

Section 12

Glossary

Abbreviations and terms used during the day

Sec. 12 · Glossary

Key abbreviations and terms

Tourism & Business

OTA	Online Travel Agency — online booking platforms (Booking.com, Expedia, Airbnb)
PMS	Property Management System — hotel management software for bookings, rooms and billing
CRM	Customer Relationship Management — system for managing client data and relationships
KPI	Key Performance Indicator — key metric for measuring business performance
FAQ	Frequently Asked Questions — common questions with pre-written answers
DMC	Destination Management Company — local operator organising on-site tourism services
B2B / B2C	Business-to-Business / Business-to-Consumer — type of commercial relationship

AI & Technology

LLM	Large Language Model — AI model trained on large amounts of text
ML	Machine Learning — AI discipline that learns from data without explicit rules
RAG	Retrieval Augmented Generation — technique that connects an LLM to external documents
API	Application Programming Interface — interface for communication between software systems
MCP	Model Context Protocol — Anthropic standard for connecting AI agents to external tools
GPU	Graphics Processing Unit — specialised processor used for training AI models
GDPR	General Data Protection Regulation — EU regulation on personal data protection
DPA	Data Processing Agreement — contractual agreement on data processing with third-party providers
SOC 2	Service Organization Control 2 — security certification for cloud service providers