Foundation Data In Action: Closing the “Last 10%” Gap in LLM Performance
Generic AI is remarkably good at creating content that looks good at first glance. It can process your documents and create output that seems reasonable.
When “seems reasonable” isn't good enough, Foundation Data can help us close the “last 10%” gap in LLM performance.
Real World Example: Foundation Data Workflow
Let's look at a real example of how SLI used Foundation Data in its own business:
SLI’s Foundation Data wasn't created for the FAQ specifically.
It’s a persistent knowledge base “baked in” to SLI’s internal AI assistants (LIAs) so they maintain awareness of our unique business context.
They each get a “slice” of institutional knowledge that makes sense for their role, just as people do.
LIAs save the work of attaching Foundation Data and explaining the AI’s role each time you start a new chat.
Output Comparison: Before and After Foundation Data
The side-by-side comparison below shows the impact Foundation Data had on the SLI FAQ workflow.
An out-of-the-box LLM and a Foundation Data-trained LLM were each asked to create an FAQ outline for SLI’s website using nearly identical prompts (e.g., I didn’t say “Hey Albus” to the generic model).
The generic version:
Inoffensive and extremely medium
Seems reasonable as a starting point
The Foundation Data-enhanced version:
Shows conceptual understanding of SLI’s business model, service delivery, and strategic positioning
Saved me several hours of additional cognitive work and cleanup.
This has compounding effects over time, in both directions. I’ve been plus or minus an hour on a task many many times, and quality Foundation Data is usually the difference. You can see the final FAQ here, if you’re curious.
Is the “last 10%” the same as AI’s “Legal Blind Spot”?
Nope. The last 10% is about facts, the legal blind spot is about law.
The “last 10%” is about the critical business context gap (as seen in our FAQ example). This is a natural result of an “out of the box” LLM not knowing who you are.
AI also has a legal “blind spot”, which is a natural result of an “out of the box” LLM not knowing how to think like a lawyer. You can read more about this in Vining Legal’s “AI's Legal Blind Spot” post.
Implementing AI for contract review or other use cases requiring legal judgment need specialized legal reasoning frameworks to address the legal blind spot, in addition to Foundation Data.
These are real issues, but in my view they sound a lot scarier than they actually are.
For contract redlining, SLI and VL have flat-fee joint offerings to tackle both at the same time (Foundation Data provides business context, VL provides the “Legal Frame”).
