Rapid democratization via falling inference costs — querying advanced models became orders of magnitude cheaper (cost to match GPT‑3.5 fell ~280× between Nov 2022 and Oct 2024).
Industry dominance in model releases — nearly 90% of notable models in 2024 came from industry; academia still leads in highly cited research.
Convergence at the frontier — performance gaps between top models narrowed substantially, increasing competition and lowering premium access barriers.
Financial impact is real but modest per function — most companies report cost savings <10% and revenue gains commonly <5% per function.
Domain foundation models accelerate science and medicine — clinical benchmarks and FDA approvals surged; medical foundation models proliferated in 2024.
Lower inference costs enable scalable in-app LLM features: conversational assistants, dispute resolution, merchant support.
Vendor concentration risk: design multi-provider fallbacks and evaluate open-weight models to reduce dependency.
Small gains compound: prioritize high-frequency automations (KYC parsing, reconciliation, fraud triage) to realize cumulative impact.
Trust and regulation: invest in explainability, consent flows, and responsible-AI practices to maintain user confidence and comply with rising regulation.
Leverage vertical models: plan for finance-tuned foundation models for better AML, risk scoring, and personalized offers; include validation and monitoring.
AI is becoming cheaper and more widely used, with big companies leading its development. Top AI systems are now similar in performance, and while the financial impact is small, it adds up over time. AI is also becoming specialized for different fields. In fintech, this means companies can build smarter features at low cost, but must avoid depending on one provider, focus on automating frequent tasks, ensure transparency to build trust, and use specialized AI models for better results.
For me as someone targeting AI PM roles in consumer apps and fintech, this means I should focus on products that automate high-frequency tasks and can explain their decisions clearly to users.