In partnership with
This week on SHIFT
We’ve all see the 95 percent of AI projects fail stat. It usually gets cited as proof the whole thing is overhyped. Philip Rathle reads it the other way around. To him, that number is evidence the checks and balances inside companies are alive and well, catching systems before they ship and tank a reputation, a balance sheet, or someone's safety.
Which leaves the more interesting question: what do the surviving 5 percent actually look like? Philip is the CTO of Neo4j, with 30-plus years in data and databases, and he has spent the last decade building graph technology that, as he puts it, accidentally became the left brain to the LLM's right brain. Jennifer digs into why the AI systems reaching production are composites rather than single models, what LLMs still cannot do no matter how large they get, and why "attackers think in graphs" even when defenders do not, in the latest installment of our oral history project.
Advertisement
AI made PMs faster. Multiplayer mode is still broken.
A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.
Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.
And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.
AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.
Three things to listen for
1. The failure rate is the guardrail working
Rathle argues the MIT figure is not a verdict on the technology but a sign that companies are axing the less than useful and ungated systems. The ones that do make it, he says, aren’t a single model doing all the AI. They are composites: large and small models plus databases, knowledge layers, and tools that cover what LLMs are bad at.
2. A lot of high-stakes AI is really connect-the-dots
Walmart runs a system that lets any of its 1.6 million employees look at any role in the company and trace the viable paths to get there. That is not a probability problem, it is a graph problem with an exact answer. Rathle makes the case that rules-based symbolic AI, widely written off, is quietly back in the stack alongside the probabilistic kind.
3. "Attackers think in graphs"
The line originated at Microsoft, and the point is that criminals know most systems cannot follow a chain through multiple intermediaries, whether that is a string of shell corporations or a string of hopped machines. If defenders cannot see the connections, they often find out years later, after the money or the data is gone. Intuit, he notes, uses the company's technology across more than 100 million endpoints.
On AI experiments: They only know what they know. They're going to be wrong, and that's okay. You have no obligation to let them decide your life for you.
Listen now
SHIFT publishes every week. If someone forwarded this to you and you'd like to get it directly, subscribe at shiftshow.ai. And if you have a question, a guest suggestion, or a story we should cover, reply to this email. Jennifer reads them all.