LinkedIn’s People You May Know algorithm was linked to what jobs people got a year later.
1. HBR On Strategy focuses on the right way to launch an AI initiative, with Harvard Business School assistant professor Yavor Bozhinov explaining why so many projects stall or fail.
2. Bozhinov matters because he is a former data scientist at LinkedIn and an HBS professor who wrote the HBR article "Keep Your AI Projects on Track."
3. The episode asks how leaders should choose, build, launch, evaluate, and monitor AI projects so that they actually get adopted.
4. Bozhinov says AI projects are not deterministic like IT projects, and the same prompt can produce different answers in ChatGPT.
5. He says the reported 80% failure rate can come from bad project selection, low model accuracy, or bias and unfairness after the build.
6. He gives a LinkedIn example where an AI tool reduced analysis work from weeks to a day or two, but employees still did not use it after launch.
7. He says the launch failed because users did not trust the product, even after a strong announcement and event.
8. He says "if you build it, they will not come," and he has seen the same pattern in multiple large-company cases.
9. He recommends choosing AI projects by impact first, because data science teams often pick technically interesting work that is misaligned with business strategy.
10. He says feasibility also includes data, infrastructure, privacy, fairness, and transparency, and those issues are expensive to bolt on later.
11. Bozhinov says trust has layers, including trust in the algorithm and trust in the developers who built it.
12. He says intended users should be brought in early so the team solves the right problem for employees or customers.
13. He distinguishes internal-facing tools, such as sales prioritization systems and customer-support chatbots, from external-facing products such as Netflix ranking, Google ranking, and ChatGPT.
14. He says internal products usually need more direct involvement from employees, while customer-facing products rely more on focus groups and experimentation.
15. He cites a LinkedIn study showing that experimentation improved a final product by about 20% on key business indicators.
16. He points to Etsy’s infinite-scroll experiment, where the team had to separate two hypotheses and discovered that more pages and small delays did not matter much.
17. He says evaluation is necessary because many AI products have neutral or negative effects on the same metrics they were built to improve.
18. He describes LinkedIn’s People You May Know audit, where a connection metric looked good but a Science study found effects on job applications and jobs a year later.
19. The conversation is practical and interview-driven, with Kurt Nikish pressing for examples, trade-offs, and concrete management steps.
20. Listeners who manage AI projects, product teams, or experimentation programs will get the most value, while people wanting a purely technical model discussion may skip it.
