In het Nederlands
Interactive presentation · click anything for an explanation
From API wrapper
to an agent
you can trust
How every company builds its own AI, which walls it runs into, and which element of the periodic table solves each wall.
Speaker notes
Everyone wants their own AI. Most companies start in the same place and run into the same walls, in the same order.
We do this in two parts. First the building blocks: the periodic table of AI. Then a real-world example, step by step, and at every step we point to the element in the table.
Four words to hold on to, the four layers of the table: building blocks, modules, system, evolution.
What makes your own AI reliable is what you build around it. The model is one tile in a table of nineteen. Almost everyone starts with that one and discovers the rest by running into something. You can also know them in advance.
Speaker notes
If you remember one sentence: the model is the easy part. ChatGPT, Claude, Gemini: you rent those. What makes it your AI, and what makes it safe, is what you build around it.
The example coming up shows exactly that: eight steps, and not one of them is about a smarter model.
The periodic table of AI
Columns are families: what an element does. Rows are layers: how composed it is. Click anything for an explanation.
Read a column from top to bottom: in Reactive it goes from control (Pr) via action (Fc) to autonomy (Ag). And just like with Mendeleev, there is a gap.
Speaker notes
Mendeleev did not order chemistry by who discovered an element, but by what it is and how it behaves. That is what we do here with AI.
Five families as columns, like the groups in chemistry. Reactive responds to a question. Retrieval fetches. Orchestration decides who does what and when. Validation checks. Models is the engine.
Four layers as rows, like the periods. Layer 1 is the building blocks, the DNA. Layer 2 is working modules, the organs. Layer 3 is the operational system, the living organism. Layer 4 is what is coming next.
Click through layer by layer. At the end, point to the empty box: Validation does not have a building block yet. We come back to that at step 8.
Combine elements into compounds
Basic interaction
You ask, the model tells. No knowledge of you, no hands.
PrPrompt+LgLLMRAG
First retrieve from your sources, then answer. The smart intern gets access to the internal files.
EmEmbeddings+VxVector DB+LgLLMDeployed solution
RAG with hands, limits and oversight, built into a real process.
RgRAG+FcFunction call+GrGuardrails+InOversightThis is exactly the route of the example coming up: from a single building block to a system you can trust.
Speaker notes
Just like in chemistry, elements only get interesting when you combine them. Prompt plus model is the basic interaction. Embeddings plus vector database plus model is RAG. RAG plus function calls plus guardrails plus monitoring is a solution that really runs in your process.
Keep these three in mind. The example coming up walks through these three layers.
How every company builds its own AI: eight walls, eight solutions
Every wall forces a new element. Click a step to go there.
Speaker notes
This is the whole story on one screen. Each card is a moment where things go wrong, and below it what you add. Top right you see the elements from the table, in the colour of their family.
Look at the colours: four of the eight steps are yellow, the Validation family. Half the work is control.
"Our own AI"
You take ChatGPT, wrap an API call around it and call it the company's AI assistant.
A rented model with your logo on it. No knowledge of its own, no behaviour of its own, no control.
Click any block in the diagram for an explanation.
Speaker notes
Everyone starts here. An employee asks a question, it goes through a thin layer of code to a vendor's model, and the answer comes back.
It feels like your own AI. But there is nothing of you in it. Of the nineteen elements you use one, and even that one is rented.
After months of work
Grounded in your policy, safe, affordable and fully visible.
"Who are you?" And the agent says: "I'm ChatGPT." Someone had changed the system prompt and skipped the evals.
You don't build Validation once. It is an agreement you keep with every change.
Speaker notes
The joke of the story is that it ends where it began. All the layers are in place. And still the agent says "I'm ChatGPT" again, because one person made one change without running the test.
That is the real lesson. The technology can be solved. The discipline around it is the work.
One question, five families
- Models: who thinksThe frontier model for hard questions, a small model for the rest.
- Reactive: who it is and what it doesThe system prompt gives it a role, function calls via MCP give it hands.
- Retrieval: what it knowsEmbeddings and the vector database fetch the right passages from your policy.
- Orchestration: who does what, whenThe RAG pipeline, the cache and the router. Usually just code in a framework.
- Validation: what is not allowed, and how you see itPermissions on every action, filters on input and output, a log of everything and a test for every change.
Speaker notes
This is the diagram you take home. Follow one question from top to bottom: the employee asks it, it passes the input check, the cache checks whether it was asked before, the router picks a model, the agent fetches knowledge or does something in a system, and the output is filtered. Around everything sits the log, and at the bottom the gate that stops every change until the test is green.
The colours are the families from the table. Click any block when a question comes up.
Which elements did the example touch?
Speaker notes
Put the example back on the table. Every tile that was touched shows the step it belongs to. The rest is dimmed.
Three things stand out. The yellow column carries half the work. A whole corner stays empty: skills, goal loops, multi-agent. And there is the empty box at the top of Validation. Just like with Mendeleev, the gaps are the most interesting part of the table: they tell you what is coming.
- True: you buy many layersA good vendor supplies permissions, filters and logging. You don't have to build them yourself.
- But you need to know which ones you buyAs soon as AI is connected to your own documents and systems (steps 2 and 3), steps 4 and 5 come with it right away. You can't postpone those.
- Start with the biggest riskNot with the nicest feature. Which wall will you hit first? That is where you start.
Speaker notes
To be fair: not every company builds this itself. You buy many layers, and that is fine.
But you still need the table, because you have to be able to ask which elements you are buying and which ones you have to fill in yourself. And one thing you don't postpone: as soon as the AI can reach your own data or do something in your systems, permissions and filters come with it. Start with the biggest risk, not with the nicest demo.
The same elements, in our own AI layer
| Element | In the example | At Newway |
|---|---|---|
| PrPrompt | System prompt with name and tone | A fixed rules file that runs with every session: how we write, what goes where, what is never allowed |
| RgRAG | Policy documents in a vector database | Search our own knowledge base first, only then elsewhere |
| FcFunction call | HR, Jira and Slack via MCP | Odoo, mail, Slack and the knowledge base via MCP |
| GrAccess | RBAC and OAuth | Separate connections for reading and for writing; permissions per folder in the knowledge base |
| GrFilters | Input classifier, PII redaction | No client health or personal data in the knowledge base, with a scan beforehand; never an unrequested message to a client |
| SkSkill | Not in the example | Our own skills for the whole team, managed centrally |
| GlMemory | Only a cache | Memory on disk: memory files, daily notes and decisions that stay in place |
| InSeeing and testing | Observability and evals | Every change in Odoo is in the chatter. We don't have fixed evals yet: our own empty box. |
Speaker notes
We work with these elements ourselves every day. They are just filled in for our own work.
And we are honest about our own gap: we partly have observability, through the chatter in Odoo. A fixed test set that runs with every change is something we don't have yet. Exactly the empty box in the table, and exactly the wall from the plot twist.
What makes it reliable is what you build around it.
Do this today: take the AI tool you use and tick every tile it uses. Every tile you are not sure about is a wall still to come.
Speaker notes
Back to the sentence from the start. The model is the easy part. And the four layers: building blocks, modules, system, evolution.
Fifteen minutes of homework: take your own AI tool and walk through the table. Wherever you don't know, that is where the next wall is.