Why You Don’t Need the Biggest AI Model

Smaller models. Specialized models. > Any trendy model
…even the latest one that supposedly can replace your entire finance team in 15 minutes. 😅

We have a joke at Capto recently:
“Guess why they are slowing down the AI race?”

My bet?
They made too many unrealistic promises.
📉 Their KPIs increasingly show that AI by itself won't justify the valuations currently attached to it.
So maybe they slow things down…
Maybe they postpone the promised impact…
Maybe they just need more time for reality to catch up with the hype.

Why do we think that?
Because one thing we have been saying, and proving, for a long time is:
You don't need the biggest model.
You need the right model.

And when it comes to finance, some of the smaller, specialized models we use can absolutely beat the latest general-purpose models at the specific task they were designed for.

A few examples 👇
📄 #LayoutLMv3
Very good at understanding document layouts: invoice headers, line items, totals, etc.
It's open-weight, so we can fine-tune it directly into an AP/AR intake agent without adjusting to it.

📊 #SECBERT
Built around financial and regulatory language.
Useful for mapping financial statement language to reporting taxonomies, while the actual reporting workflow remains yours.

🕸️ #GraphSAGE
Designed for graph-based problems.
Think accounts as nodes and transactions as edges.
A great building block for scoring engines behind compliance and AML workflows.

💰 #FinBERT
Fine-tuned for financial text.
Useful for risk classification, analysing financial language, flagging red flags in correspondence or disclosures, and supporting compliance workflows.

And what do they have in common?
⚡ Small enough to self-host cheaply.
🎯 Specialized enough to perform extremely well at their specific job.
🔧 Easy to integrate into a broader automation workflow.

And very often, the best solution is a combination of AI and pure code.
That's what we've been building at Capto for years:
Using the right tool for the right part of the process, instead of forcing the latest trendy model into everything.

Hopefully you enjoyed this second technical post where I try to help you separate the trend from the opportunity of Automation & AI... our little garden. 🌱

Have a great Tuesday!
Alexis

—

Originally posted on LinkedIn:

https://lnkd.in/p/dUPEwAC7