In this episode of The Tech Trek, ModelOp CTO Jim Olsen walks through what actually happens when an organization decides to bring a foundation model inside rather than build its own. He describes the two dominant enterprise patterns — RAG-based chatbots, often reached through an existing Microsoft contract, and summarization, where he sees far more self-hosted smaller models because the task doesn't demand a frontier model — and notes that most buyers are evaluating the business outcome rather than analyzing the model itself. He frames governance on three pillars: a robust inventory detailing what you have and what makes up each model, a regimented process running from use-case risk assessment through review, deployment, monitoring and eventual retirement, and ongoing monitoring for drift and exceptions. His analogy is the pre-DevOps era of software: developers shipping straight from their desks, resisting process until the 2 a.m. call taught them otherwise. The most practical stretch of the conversation covers measuring performance against business use rather than model internals — establishing baselines early, tracking sentiment drift and top parts of speech, and using embedding cosine similarity between answers and source documents in a RAG system. He closes on when to stop doing this manually, and his answer is criticality to the business rather than volume.
- The two dominant enterprise patterns: RAG chatbots and summarization, and why summarization goes self-hosted.
- Why most buyers evaluate the business outcome rather than analyzing the foundation model itself.
- The three pillars of governance: inventory, regimented process, and ongoing monitoring.
- Why a robust inventory has to detail what makes up each model, not just which vendor you use.
- Starting with use-case risk assessment before deciding whether an LLM is even the right tool.
- All models have a shelf life — retirement and recertification are part of the life cycle.
- Governance is about business risk as much as regulatory compliance.
- Real-time guardrails catch the obvious failures; the hard question is whether it serves the business well.
- Why regulations increasingly demand a human in the loop for health-related decisions.
- Measuring against business use: baselines, sentiment drift, top parts of speech, and toxicity checks.
- Using cosine similarity between a RAG answer and its source documents to detect drift from intent.
- The pre-DevOps analogy, and why the 2 a.m. call is what changes minds about process.
- Minimum viable governance, and why criticality to the business — not volume — determines when to automate.
[00:04] – Introduction
[00:40] – What ModelOp does
[01:38] – The process of bringing a foundation model into a business
[02:20] – RAG chatbots and summarization as the two common use cases
[03:11] – Why buyers evaluate the outcome rather than the model
[04:00] – Defining AI governance
[04:45] – Pillar one: a robust inventory of what you have
[05:30] – Pillar two: the regimented process and SR 11-7
[06:18] – Retirement, recertification, and shelf life
[07:00] – Why governance is business risk, not just regulation
[07:50] – Why the guardrails are essential in unfamiliar territory
[08:40] – The pre-DevOps parallel
[09:23] – Why bad predictions are harder to spot than crashes
[10:56] – How changing regulations affect models in production
[11:40] – Vendor models and data you cannot audit
[12:30] – RAG references, data privacy, and proving due diligence
[13:20] – Reporting obligations and human-in-the-loop requirements
[14:04] – Performance against business outcome
[15:39] – Upfront metrics versus establishing a baseline
[16:20] – Sentiment drift and top parts of speech
[17:12] – Cosine similarity against source documents in RAG
[18:00] – Toxicity and gibberish checks
[18:46] – Why higher-level abstractions matter more than vector clusters
[19:30] – Lessons from skipping governance early
[20:19] – The 2 a.m. call that changes minds
[21:00] – Why regulated industries were the earliest customers
[21:51] – Looking five to ten years out
[22:40] – Why governance becomes as standard as CI/CD
[23:24] – Minimum viable governance and the continuum
[24:58] – When to stop doing this manually
[25:40] – Why criticality to the business is the deciding factor
[26:30] – Where to learn more
[27:10] – Closing remarks
ModelOp's CTO, Jim Olsen, joins the The Tech Trek Podcast to unpack the challenges and opportunities of integrating foundational models into enterprise operations. Jim, who leads ModelOp’s technical innovation from his off-grid cabin in Colorado, shares his insights on regulatory compliance, IT infrastructure, and real-world AI applications like chatbots and summarization tools. A must-listen for AI leaders!
