Focus on workflows, evals, reuse—and humans
One year into real-world agentic AI, the difference between hype and value comes down to workflow redesign, rigorous evaluation, and smart human–agent collaboration. Our fieldwork across more than 50 builds distills six lessons that move beyond demos to measurable outcomes. The mandate: pick the right tool for each task, make every step observable, and build for reuse so you can scale without the “AI slop.”
Points clés
- McKinsey examined more than 50 agentic AI builds and dozens more in the market to extract six deployment lessons.
- Value arises from reimagining end-to-end workflows; agents act as orchestrators within frameworks such as AutoGen, CrewAI, and LangGraph.
- An alternative-legal-services provider embedded learning loops by logging every user edit, enabling agents to codify evolving legal expertise.
- Agents aren’t always the answer: low-variance, tightly governed tasks (for example, investor onboarding or regulatory disclosures) may favor rules or predictive analytics, while high-variance tasks benefited a financial-services company by reducing human validation in complex data extraction.
- To prevent “AI slop,” teams should treat agent onboarding like hiring—define job descriptions, build granular evaluations, and iterate; a global bank refined know-your-customer and credit-risk agents by repeatedly probing “why” to close logic gaps.
- Build observability to verify each step: when accuracy dipped in a legal document workflow, instrumentation revealed poor upstream data quality; fixes to data collection and parsing restored performance.
- Reuse beats one-offs: centralizing validated services (for example, LLM observability, preapproved prompts) and reusable assets on a single platform can eliminate 30–50 percent of nonessential work.
- Human oversight remains essential for judgment, compliance, and edge cases; roles and team sizes will shift as workflows transform.
- Thoughtful UI design accelerates trust: a property and casualty insurer’s reviewers used highlights and auto-scrolling to validate AI summaries, reaching user acceptance near 95 percent.
- Authors Lareina Yee, Michael Chui, Roger Roberts, and Stephen Xu publish these lessons under McKinsey’s QuantumBlack Labs on September 12, 2025.
À retenir
Want agentic AI that actually helps? Start by mapping the workflow, not shopping for the shiniest agent. Pick the simplest tool that works, wire in evals and observability before your users invent new words for “AI slop,” build reusable components, and keep people in the loop with crisp UI. Do this and your agents will look less like overeager digital interns and more like teammates who actually get things done.
Sources
