If your leadership is debating which AI tool to roll out, they are having the wrong discussion.

Barbara Lampl has spent more than 20 years developing and implementing data and AI strategies, including building and fixing the models themselves. On LinkedIn she goes by AI Babsi. Her core claim in this conversation: AI adoption fails long before anyone touches a model, because organizations refuse to redesign their processes and never captured the knowledge that actually keeps them running.

About this episode

The only thing you control with a large language model is the input. From there, Barbara works outward. Processes first: touch every single one, because even a clean, human-optimized workflow was built for deterministic, rule-based systems, and an LLM is neither. Then data. Meeting transcripts and ticket systems are step one, a data lake of unstructured material. Step two is extraction and pattern matching, because only 20 to 30 percent of a transcript is worth keeping. The real problem sits underneath: the implicit knowledge that never made it into Salesforce, Jira, or Confluence.

That problem gets urgent with the retirement wave. For companies with five years on the clock before a generation walks out, Barbara's method is simple and uncomfortable: identify who leaves first and put them on video. Structured exit interviews never capture what the veteran on the shop floor knows about the delivery that goes wrong every second Tuesday. Unstructured knowledge needs unstructured capture, plus radical transparency about why you are recording, plus top-down buy-in. People you ignored for 30 years will not spill their secrets into a form.

The second half gets technical. RAG as an umbrella term, not a synonym for vector database. The six-month trade-off when you move from vectors to a knowledge graph. Why fine-tuning your own model requires exactly the data work you were hoping to skip, since embeddings do not grow on trees. How LLM-led data migration cuts legacy monolith projects from three years to months, with expectations managed accordingly. The token economics of file formats. And the week's news: the Claude Fable 5 re-rollout, its guardrails and Opus 4.8 fallbacks, and why European calls for a homegrown Anthropic ring hollow without the compute infrastructure to run one.

"AI is never the reason. It's just the light showing on the cracks where we duct-taped it."

Key takeaways

Touch every process. None were designed for non-deterministic systems, so step zero is orchestration redesign, not tool selection. Kill what makes no sense while you are at it.

Exit interviews do not capture tribal knowledge. Video your retiring experts, transcribe, extract. Unstructured knowledge needs unstructured capture, and that culture shift usually needs someone from outside.

Slice by value, not by workflow. Start where good enough data already exists and the payoff is highest, then do the heavy preparation in the background.

File formats are a cost lever. PowerPoint runs a factor of 9 to 13 in token usage, PDFs around 5 to 6, and markdown is what LLMs love. Stop producing slides for internal use.

If your leadership is discussing tools, the first investment is C-level training. A tool rollout without strategy, knowledge, and clear business goals wastes money you are racing against time to spend well.

"There is no AI strategy. There's just a business strategy supported by data and AI."

About the guest

Barbara Lampl is a behavioral mathematician with a background in math, finance, and psychology (University of Konstanz and University of Zurich). She has developed and implemented data and AI strategies for over 20 years, including model building and fixing, and lectures on data science. She runs empathic business by Barbara Lampl and, together with her partners, the Data AI Institute. Find her on LinkedIn, where she posts as AI Babsi.

Resources mentioned

Barbara Lampl on LinkedIn

The Data AI Institute

Genie: Uber's Gen AI on-call copilot

Anthropic on Claude Fable 5 and Mythos 5

Listen & subscribe

Find the Stellar Work Podcast on Spotify, Apple Podcasts, YouTube, and more.

For weekly essays on transformation, flow, and AI in knowledge work, join the Stellar Work newsletter.

The Stellar Work Podcast is hosted by Ben, founder of Stellar Work. Conversations with the people shaping how work actually gets done.

‍