Hi David, This was very interesting/relevant to a piece of research I am doing. I have been reviewing your Latent Genius site but can’t find more detail on this particular project. Are you looking to develop this further into a full Contract Lifecycle Management project/agent? Have you seen anything else out there that is heading that way? Thanks.
Thanks for reaching out, and glad this resonated with your research!
Honest answer: this started as a "scratch my own itch" project. I wanted to solve a real world problem for a user-base that required them to understand what was actually in a stack of contracts without reading them line by line. Classic document intelligence problem.
But the deeper I got, multi-model OCR, entity extraction, GraphRAG for relationship mapping, the more I kept bumping into CLM-shaped problems. Version tracking, obligation extraction, renewal alerting, counterparty risk analysis... it's all right there once you have the foundation.
So I'm genuinely at a decision point:
Keep it focused — a sharp tool that does contract analysis really well, plugs into existing workflows via MCP, and stays lean
Build toward full CLM — which is a much bigger surface area but arguably where the real enterprise value lives
I lean toward thinking there's a gap in the market for an AI-native CLM, most incumbents bolted on AI after the fact. But I also know scope creep kills projects.
Curious what angle your research is taking? If you're exploring the CLM space, I'd be interested to hear what you're seeing. Always open to comparing notes.
Thanks for replying and the extra detail. This is very interesting perspective and consistent with what we are seeing.
I'm interested in your point 2 - the full CLM. We are researching the market to understand (1) How close are we to this (full CLM). (2) Who is likely to get there first and why. E.g. one of the specific CLM disruptors like Sirion and Icertis. An incumbent like Docusign, an 'adjacent' like a SalesForce, or a more start-up like a HarveyAI. We also looked at a hyperscaler (specifically MSFT) and the AI Labs themselves, but both are more focused on eco-system enablement and consumption, rather than owning domain specific AI apps.
Or, none of these and something new and/or still under the radar.
Would be interested in your thoughts.
The rise of the legal GPTs have added another interesting angle.
With all of these dynamics, there is definitely growing interest in this space and investor appetite. We suspect some convergence plays and consolidation / M&A in 2026.
Thanks again and looking forward to continuing the conversation.
Really appreciate this, and glad the dimension-limit note was useful. You’re exactly right that retrieval optimization often gets overlooked relative to embedding generation cost, even though it can dominate user-perceived latency and total run cost at scale. On schema-as-instructions, I’ve found the big payoff is consistency over time: extraction quality becomes less dependent on prompt wording and more tied to explicit structure you can version, test, and evolve. If you’re building a similar pipeline, I’d also recommend instrumenting cache hit rate by query type, because that quickly shows where your retrieval stack is actually earning its keep.
Hi David, This was very interesting/relevant to a piece of research I am doing. I have been reviewing your Latent Genius site but can’t find more detail on this particular project. Are you looking to develop this further into a full Contract Lifecycle Management project/agent? Have you seen anything else out there that is heading that way? Thanks.
Thanks for reaching out, and glad this resonated with your research!
Honest answer: this started as a "scratch my own itch" project. I wanted to solve a real world problem for a user-base that required them to understand what was actually in a stack of contracts without reading them line by line. Classic document intelligence problem.
But the deeper I got, multi-model OCR, entity extraction, GraphRAG for relationship mapping, the more I kept bumping into CLM-shaped problems. Version tracking, obligation extraction, renewal alerting, counterparty risk analysis... it's all right there once you have the foundation.
So I'm genuinely at a decision point:
Keep it focused — a sharp tool that does contract analysis really well, plugs into existing workflows via MCP, and stays lean
Build toward full CLM — which is a much bigger surface area but arguably where the real enterprise value lives
I lean toward thinking there's a gap in the market for an AI-native CLM, most incumbents bolted on AI after the fact. But I also know scope creep kills projects.
Curious what angle your research is taking? If you're exploring the CLM space, I'd be interested to hear what you're seeing. Always open to comparing notes.
Hi David,
Thanks for replying and the extra detail. This is very interesting perspective and consistent with what we are seeing.
I'm interested in your point 2 - the full CLM. We are researching the market to understand (1) How close are we to this (full CLM). (2) Who is likely to get there first and why. E.g. one of the specific CLM disruptors like Sirion and Icertis. An incumbent like Docusign, an 'adjacent' like a SalesForce, or a more start-up like a HarveyAI. We also looked at a hyperscaler (specifically MSFT) and the AI Labs themselves, but both are more focused on eco-system enablement and consumption, rather than owning domain specific AI apps.
Or, none of these and something new and/or still under the radar.
Would be interested in your thoughts.
The rise of the legal GPTs have added another interesting angle.
With all of these dynamics, there is definitely growing interest in this space and investor appetite. We suspect some convergence plays and consolidation / M&A in 2026.
Thanks again and looking forward to continuing the conversation.
All the best for now.
Really appreciate this, and glad the dimension-limit note was useful. You’re exactly right that retrieval optimization often gets overlooked relative to embedding generation cost, even though it can dominate user-perceived latency and total run cost at scale. On schema-as-instructions, I’ve found the big payoff is consistency over time: extraction quality becomes less dependent on prompt wording and more tied to explicit structure you can version, test, and evolve. If you’re building a similar pipeline, I’d also recommend instrumenting cache hit rate by query type, because that quickly shows where your retrieval stack is actually earning its keep.