CLSS · T&I · October 2026
From one model
to a portfolio
Open weights, small models and fine-tuning
Every client has dozens of tasks where AI could help
Today: one model for everything, or no AI at all
The future is a portfolio: the right model for each task
One frontier model provider via API
Data too sensitive
Not good enough, or too costly
Not implemented yet
Frontier model
Complex reasoning
Model A
When cost per token matters
Model B
Sensitive data
Model C
High-volume, narrow tasks
Model D
Fast classification
The why, the proof and the how
What changed, and why it matters
What we learned in our hackathon
How we bring it to our clients
Two terms to know
Open weights
Download the model, run it anywhere
Small language models
Small enough for a single GPU
The capability race
Closed models keep raising the bar
Open models follow close behind
~4 months
Two reasons to diversify beyond one frontier model
Sovereignty
- Your data.
- Your model.
Business case
With the right model for each use case
Sovereignty is a scale, not a binary
1Chinese API
Data to China
2US API
Data to the US
3US cloud, EU region
EU data, US law
4EU provider API
EU data, EU law
5Your EU cloud
Your model, rented GPUs
6Your own GPUs
Everything yours
Closer to you, more sovereign
Many of our clients todayEven without sovereignty, a portfolio pays off
Quality
Latency
Cost
We tested it on insurance tickets
Fine-tuned: trained further on our own labelled tickets
Out of the box, a model is a generalist
Fine-tuned, it becomes a specialist
Out of the box, Claude still led
Fine-tuning made the difference
0.76–0.86
for open chat models. Claude Opus 5 led with 0.88.
0.82 → 0.97
Decision models are a new class of fast, cheap classifiers
“I reported water damage three weeks ago (claim CLM-20931) and still haven’t heard back. The repair company needs an answer by Friday.”
Chat model
0 msDecision model
0 msWrites its answer as text, word by word
Picks from fixed answers, with a confidence score
Back to our test, ranked by speed
Jev answers in a third of a second
0.66 s
Our fine-tuned gpt-oss 20B.
0.35 s
Jev, with 0.91 accuracy: second only to our fine-tuned model.
Remember today: one model, or no AI at all
Now we know how to split a portfolio
Tasks Today Where we are headingOne frontier model provider via API
For every task that uses AI today
Frontier model via API
Complex reasoning
No AI at all
Too sensitive, or too costly at scale
One potential use case for four practices
AML alert triage
Sovereign AI strategy
Self-hosted coding agents
Competitor product matching
Five potential propositions
A natural add-on to our AI projects
Use case scan
“Where would this pay off?”
Evaluation framework
“Is a smaller model good enough?”
Fine-tuning
“Can we close the quality gap?”
Hosting PoC
“Where can it run, on our terms?”
Strategic support
“How do we set this up well?”
Listen for these, and bring us in
Your client
We turned off Opus for everyone after spending €100k in one week.
Our developers aren’t allowed to use Copilot.
With this latency, our customer experience is terrible.
A whole team sorts tickets all day, but the data is too sensitive for AI.
We want to be less dependent on the US.
We want to purchase GPUs to start testing open-source models.
Three things to remember
There’s more out there than frontier models, and they’re worth it
Good enough today, with more sovereignty and a better business case
It’s a choice per use case, not a default
Model, hosting and sovereignty depend on the task, data and budget
We’re ready to bring this to clients now
Five propositions, on their own or combined
Bonus: a real way to make impact
For sovereignty, the climate and the client itself
Let’s bring sovereign, efficient AI to our clients
Let’s build it together.