Generative AI by sector: where it really creates value in 2025

“AI will transform every sector” — that’s true, but not at the same pace or in the same way. After working on projects across multiple industries, here’s what I actually observe in the field.

Finance and insurance

What works. Analysis of contractual and regulatory documents. LLMs are excellent at extracting precise information from long contracts, comparing clauses, identifying inconsistencies. Tasks that took junior lawyers hours are done in minutes.

What’s still limited. Anything touching regulated decisions. An LLM can’t approve a credit or validate a tax return — it can assist the decision, not replace it.

Healthcare

What works. Medical record synthesis for practitioners. Summarizing a 50-page patient history into 2 relevant pages is a mature use case with real time savings.

What’s still limited. AI-assisted diagnosis. LLM hallucinations in a medical context can have serious consequences. Serious projects in this domain invest massively in evaluation and human oversight.

Legal

What works. Jurisprudence research, first drafting of standard contracts, document version comparison. Firms are significantly reducing time spent on research and initial drafting tasks.

What’s still limited. Formal legal opinions. An LLM can produce something that looks like a legal opinion but is wrong on specific points. Liability risk is real for firms.

E-commerce and retail

What works really well. Product description generation at scale, description personalization, automated customer support on standard questions. Mature use cases with measurable ROI.

What’s limited. Pure product recommendation. LLMs are less effective than classic recommendation systems for predicting purchases from behavioral history.

HR and recruiting

What works. CV pre-screening on objective criteria, job posting writing, interview synthesis.

What poses problems. Bias. LLMs trained on historical data can reproduce and amplify existing biases in recruitment processes. This is a serious legal and ethical risk.

What I take away for an AI PO

The most advanced sectors are those where the task is well-defined, errors are detectable, and human supervision is natural in the workflow.

For each project, the real question isn’t “can AI do this?” but “under what conditions, with what supervision, and at what reliability level?”