A Practical AI Session for Odoo Teams

AI Is Not Magic

What it does, what it fakes, and how to tell the difference
Format60 min presentation · 30 min workshop
AudienceSales · Consultants · PMs · Admin · Support
Foundations

AI is statistical prediction machinery

Modern AI systems are statistical models trained on large datasets to predict, classify, rank, or generate outputs. Prediction does not mean random guessing — it means using learned patterns to estimate the most likely useful output.

Different systems predict different things

  • The next word
  • The next pixel
  • The next customer likely to churn
  • The next best move in a game
  • The shape of a protein
  • Whether a face matches another face
  • Whether an invoice looks suspicious
  • Whether code is likely to break
"Likely output" and "true answer" are not the same thing.
Foundations

Two very different things get called AI

Logic-based — humans wrote the rules

A human explicitly programmed every condition. The system executes it exactly. No learning, no statistics, no surprise.

  • Odoo automated actions
  • If invoice > €1,000 → flag for approval
  • If email contains "unsubscribe" → filter it
Deterministic. Predictable. Does exactly what it was told — nothing more.

Statistical AI — learned from data

Nobody wrote the rules. The system was trained on millions of examples and learned to recognise patterns on its own.

  • Netflix learns what you'll watch next
  • A fraud model flags suspicious transactions
  • A churn model predicts who is about to leave
Probabilistic. Improves with more data. Can handle inputs nobody anticipated.
When people say "AI," they usually mean the second kind — even if it has been running quietly inside products for years.
Foundations

AI is a family of technologies, not one thing

Perception

  • Vision: cameras that detect tumors in scans before radiologists do; warehouse robots that navigate live environments
  • Documents: extract structured data from any invoice, receipt, or contract — handwritten or not
  • Speech: real-time transcription and spoken language translation across 100+ languages

Strategy & Discovery

  • Games: AlphaGo beat the world champion at Go — a game with more positions than atoms in the universe
  • Science: AlphaFold predicted the 3D structure of nearly every known protein — a problem biologists had worked on for 50 years

Generation

  • Images: photorealistic visuals from a text description in seconds
  • Video: synthesize footage of scenes that never existed
  • Voice: clone a voice from a few seconds of audio

Prediction

  • Recommendation: YouTube predicts what 2.5 billion people each want to watch next
  • Risk: fraud detection, churn prediction, demand forecasting
Each of these does one thing very well. None of them can hold a conversation or explain their reasoning. That is what makes language models different.
Core Concepts

How language models actually work

Predict the next token

Given everything written so far, what comes next? Text is broken into tokens — fragments smaller than words (un · believ · able). Train a model on that one task across all of human writing and it is forced to learn grammar, facts, style, and reasoning just to do it well.

What it accidentally learned

Nobody programmed translation — it emerged. Nobody programmed coding, reasoning, or summarisation — those emerged too. All from one objective, applied at massive scale.

Why it matters — and where it breaks

Most work is text in, text out
meeting → notes → requirements → tasks → emails → documents → reports
The fundamental limitIt predicts likely text — not verified truth. Fluent and confident is not the same as correct. That gap is what all the failure modes are about.
Core Concepts

Google lit the fuse

In 2017, Google researchers published "Attention Is All You Need." It solved the context problem directly.

The key idea — attentionInstead of reading left to right and compressing context into a fading memory, attention lets every token look at every other token in the sequence simultaneously — and decide which ones matter most. "The bank by the river" — the model can directly connect those two words regardless of how far apart they are. No sequential bottleneck. No forgetting.

Three forces then converged — and each amplified the others

Internet-scale dataMore text than any previous training set — diverse enough for a model to encounter nearly every concept, language pattern, and reasoning style humans had written down.
GPUsAttention requires comparing every token to every other token — a task that maps naturally onto GPU hardware. More compute made training at scale feasible.
Bigger modelsLarger networks captured more complex patterns. Researchers found that scaling up kept improving results — with no obvious ceiling in sight.
History

ChatGPT didn't build better AI - it made it accessible

For five years after the Transformer paper, powerful AI existed. But to use it, you needed API keys, engineering knowledge, and technical infrastructure. On November 30, 2022, OpenAI launched ChatGPT replacing all that with a text box.

What actually changed

The model did not suddenly get smarter overnight. The friction disappeared. No setup, no code, no expertise — just type what you need. The same shift happened with the PC when the GUI arrived, and the internet when the browser did. Each time, the interface was the missing piece.

The result

100 million users in two months — the fastest consumer adoption in tech history at the time. Not because AI improved, but because the barrier to entry dropped to zero.

What this means for you

The technology that required an engineering team in 2021 now runs in a chat window on your phone. The barrier between you and powerful AI is a sentence.

Capabilities

What LLMs are good and bad at

Strong — language-shaped work

  • Summarizing meetings and calls
  • Drafting and rewriting text
  • Explaining technical topics plainly
  • Turning notes into requirements
  • Creating test cases
  • Reading and explaining unfamiliar code

Weak — reality must be verified

  • Confirming facts without sources
  • Making commitments to customers
  • Legal or contractual judgment
  • Security and pricing decisions
  • Anything requiring live data or system access
  • Decisions with real accountability attached
Why this split existsThe model is strongest at transforming text — that is exactly what it was trained to do. It breaks down when the task requires live facts, private context, deterministic calculation, or someone to actually own the outcome.
AI can draft the answer. It cannot take responsibility for the answer.
Failure Mode 1

Confidence is not evidence

The model generates high-probability tokens that fit the prompt. A biased prompt shifts the output toward biased answers.

Bad Prompt
This is just adding one approval step in Odoo,
so it should be simple, right?
Can you explain why this should be
a small customization?
AI Response
Yes, adding one approval step is typically
a small and straightforward customization.
You can position it as a limited change
with low implementation risk.
No Odoo version · No module list · No access rules · No test scope · No integration details
Better Prompt
Assess whether this Odoo customization
is likely to be simple or risky.
Challenge my assumption that "one approval
step" means low effort.
List missing information, dependencies,
testing needs, and questions for the developer.
AI Response
I can't assess complexity without knowing:
Odoo version · which approval module · access
rules affected · integration points · whether
this is a new field or workflow change.
"One step" can mean 2 hours or 2 weeks.
The AI did not prove the task is easy. It converted a biased question into a confident explanation.
Failure Mode 2

AI moves fast. That is exactly the risk.

AI produces polished output in seconds. The speed feels like productivity. So the text gets forwarded, the spec gets sent, the update gets approved — without the person who sent it fully processing what it says.

What the failure looks like

  • A consultant sends a spec the AI wrote — without noticing the AI filled three gaps with assumptions
  • A PM forwards a status update — without checking whether "on track" is actually true
  • A sales rep sends a capability confirmation — without validating the claim
The output looked complete and professional. That is what made it dangerous. Nobody paused to read it as a reviewer would.

The test

Before sending anything AI-generated, ask: could I defend every sentence in this without re-reading it? If not, you moved too fast.

What slowing down looks like

  • Read the output as if a colleague sent it to you for review
  • Flag every claim or commitment and verify it independently
  • Ask the AI: "what did you have to assume to produce this?"
AI makes the first draft effortless. That is when the work of thinking actually starts.
Failure Mode 3

It can make a bad idea sound good

LLMs are excellent style transfer machines. They convert rough text into polished text without improving the underlying logic.

Bad Prompt
Make this sound professional:
Testing is probably unnecessary because
this change only affects one field.
AI Response
Due to the limited scope of the change,
extensive testing may not be required.
The writing improved. The judgment did not. Bad reasoning used to look rough; with AI it can look structured and senior.
Better Prompt
Improve this message, but flag anything
professionally risky or unsupported:
Testing is probably unnecessary because
this change only affects one field.
Better Direction
I would not recommend saying this.
A safer message: Even a small field change
can affect views, access rights, reports,
imports, automations, and integrations.
A focused validation is recommended.
AI can turn weak thinking into polished weak thinking.
Failure Mode 4

Missing context becomes invented context

When project-specific information is missing, the model backs off to statistical patterns from training data.

Bad Prompt
The project failed because the customer
kept changing their mind. Help me explain
the root cause professionally.
AI Response
The project was impacted by repeated
requirement changes from the customer,
which created scope instability and
delivery delays.
Maybe true. Maybe politically convenient. Maybe incomplete. Your specific project gets replaced by the average project.
Better Prompt
I want to analyze why this Odoo project
struggled. Do not accept my initial
explanation as true. List possible causes
across sales, scoping, solution design,
data migration, development, testing,
customer decisions, and governance.
Separate evidence from assumptions.
AI Response
Possible causes beyond requirements drift:
· Sales: scope underspecified at sign-off
· Scoping: no UAT criteria defined
· Data: migration complexity underestimated
· Governance: no decision log, no escalation path
Which of these do you have evidence for?
If the context is missing, AI fills the gap with vibes.
Failure Mode 5

Bigger prompt does not always mean better answer

Long context is not the same as reliable attention. The information may be inside the context window, but still not weighted correctly.

Bad Prompt
Here are 80 pages of meeting notes,
emails, specs, and random comments.
Tell me what to do.

Why it fails

  • Important details get buried
  • Contradictory notes confuse the model
  • Irrelevant documents distract it
  • Small constraints can be missed
Better Prompt
I will give you project context in sections.
First, summarize only the confirmed facts.
Second, list open questions.
Third, identify contradictions.
Do not recommend actions until I say "analyze".

Why it works

  • Forces structure before synthesis
  • Surfaces contradictions explicitly
  • Separates facts from assumptions
  • Keeps you in control of the pace
AI does not read like a careful lawyer. It reads like an extremely fast intern under pressure.
Failure Mode 6

Most AI usage is input → output

Most AI interactions are stateless inference calls. A prompt goes in, tokens come out.

It does not naturally

  • Keep thinking after the answer
  • Monitor your project
  • Know that a customer replied
  • Remember every decision forever
  • Notice that requirements changed
  • Check if code deployed successfully

Unless

  • The product has memory
  • It is connected to tools
  • It is connected to your systems
  • You provide the latest context
AI does not wake up tomorrow remembering your project unless the system gives it memory.
Failure Mode 7

A model alone cannot verify reality

A language model is not automatically connected to your systems. Text generation is not system access.

By default, an LLM cannot

  • Check Odoo production
  • Read your Git repository
  • Inspect customer tickets
  • Verify current pricing
  • Confirm an email was sent
  • Know whether a test passed
  • See the latest law or tax rule

When tools are added, it can

  • Query a live database
  • Read a file or repository
  • Call an API for current data
  • Send an action and get a result
  • Search a knowledge base at runtime
  • Confirm a workflow step completed
The model gives language. Tools give contact with reality.
The Rule

AI accelerates. Humans are accountable.

The model has no consequence loop. It does not lose the customer, fix the bug, attend the steering committee, or handle the escalation if the answer is wrong.

AI can help

  • Think faster
  • Write faster
  • Understand faster
  • Prepare better

AI should not own

  • Customer promises
  • Contractual commitments
  • Security decisions
  • Production changes
  • Final scope & pricing
AI can reduce effort. It cannot absorb accountability.
Habits

Ask AI to challenge you before it helps you

The model cannot reliably tell the difference between a fact you gave it and a biased assumption you smuggled into the prompt — unless you force it to check.

Use this before important work
Before answering, challenge my assumptions.
What am I missing?
What could go wrong?
What would need to be true for this to be safe?
What should I verify with a human?
What should I avoid promising?
Turn AI from a yes-assistant into a thinking partner.
Practical testIf the answer does not separate facts, assumptions, and things to verify — ask again before using it.
Habits

The prompt decides what the model is allowed to assume

In normal software, hidden assumptions become bugs. In AI usage, hidden assumptions become confident prose. Two habits fix this.

State your assumptions explicitly

Leave no room for the model to fill gaps on its own.

In your prompt
Assume the project is in discovery phase.
Assume no customization has been agreed.
Assume the customer has not seen a demo.
Do not assume anything about timeline.

Ask it to flag what it assumes

Surface inferences before they become facts.

Add to any prompt
Flag every assumption you make.
Separate facts I gave you from inferences
you drew. Mark anything I should verify
before using this output.
Weak
Can you write this?
Can you justify this?
Can you make this sound better?
Strong
What are you assuming?
What is fact vs inference?
What should I verify before sending this?
The quality of AI output is mostly the quality of the assumptions you allow it to make.
Tools

Do not pick by brand loyalty. Pick by job.

Closed assistants

ChatGPT, Claude, Gemini — best for general work, writing, analysis, multimodal use, and polished product experience.

Open-weight & self-hosted

DeepSeek, Mistral, Llama-style models — best when cost, data control, sovereignty, or private deployment matter.

Coding & agent tools

Codex, Claude Code, Cursor — best when the AI needs to inspect files, understand code, make changes, and run checks.

Automation tools

n8n, Make, Zapier, Odoo automations — best when the output should trigger a repeatable business process.

Practical architecture
model + context + tools + permissions + human review = useful business AI
Practical Use

AI as a thinking partner before you act

Sales — Brainstorming

  • Surface blind spots before a discovery call
  • Stress-test assumptions before a proposal
  • Anticipate objections before a demo
Sales Prompt
I have a discovery call with a logistics company
considering Odoo. Brainstorm every angle I should
explore: risks they may not mention, common
integration assumptions, change management concerns,
and questions that reveal real decision criteria.

PM — Challenge and structure

  • Turn rough notes into a structured status update
  • Surface blockers hidden in vague language
  • Prepare a steering committee update
PM Prompt
Challenge my framing before writing this update.
List every claim that implies certainty, on-track
status, or commitment I cannot fully verify.
Then draft a transparent customer update with
blockers, open decisions, owners, and dates.
ReminderBefore sending AI-assisted text, check every sentence that implies scope, price, timeline, responsibility, or certainty.
Practical Use

AI as a structuring and drafting tool

Consultant — Extract and analyze

  • Turn meeting notes into structured requirements
  • Identify missing decisions before spec writing
  • Generate UAT scenarios from a description
Consultant Prompt
Read these meeting notes. Do not write
requirements yet. First list confirmed facts,
open decisions, contradictions, and missing
information. Then ask me the three most
important questions before proceeding.

Admin — Draft with guardrails

  • Draft internal policies and process documents
  • Create onboarding checklists and FAQs
  • Summarize or restructure long documents
Admin Prompt
Draft this internal process document.
Use explicit placeholders where rules are
not confirmed. After drafting, list every
section that needs validation from HR,
finance, legal, or management.
ReminderIf the output looks complete, ask what the AI had to invent to make it complete.
Now — 30 minutes

Workshop

Two exercises. Real tools.
No coding required.

Workshop — Setup

Getting started

Steps

  1. Download and install Antigravity
  2. Download the repository .zip from Google Drive
  3. Extract / unzip the repository
  4. Open a terminal
  5. Navigate to the extracted folder
  6. Open the folder in Antigravity
Antigravity is much more useful when it can inspect a full folder instead of one pasted file.
Mac / Linux
pwd          # show current folder
ls           # list files
cd ~/Downloads
cd folder-name
cd ..        # go back one folder
Windows PowerShell
pwd
dir
cd Downloads
cd folder-name
cd ..
Example
cd ~/Downloads
cd workshop-repository-main
Workshop — Exercises

Two exercises

Exercise 1 — Read a codebase without coding
Start here
Explain this repository to a non-technical
Odoo consultant. Return:
1. What this module appears to do
2. The main Odoo models involved
3. The business features implemented
4. Where I should look first to understand the flow
Follow-up
Explain this Odoo code for a functional consultant.
What business behavior does it change?
Which users would notice this?
What should I test in UAT?
What questions should I ask the developer?
Exercise 2 — Build a presentation from scratch
Prompt
Help me create a single-file HTML slide deck.

Topic: [your topic]
Audience: [who will read it]
Slides:
1. [title + key points]
2. [title + key points]
3. [title + key points]

Make it clean, add next/previous navigation,
keep everything in one file, and tell me
how to open it in my browser.
Ask Antigravity to cite files when explaining code. Do not accept vague explanations.
Takeaways

Use AI aggressively, but not blindly

  • AI predicts — it does not know. Confidence is a style, not evidence.
  • LLMs are powerful because most professional work is language-shaped.
  • The prompt decides what the model is allowed to assume. State yours explicitly.
  • Ask it to challenge you before it helps you.
  • AI accelerates. Humans are accountable.
AI makes polish cheap. Judgment becomes more valuable.

Thank you