---
title: "Unique Research | McKinsey: Why Agents Are GenAI’s Next Frontier"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2024-09-06T16:00:00+08:00"
canonical: "https://ffcap.cn/en/research/20240906a064dy00"
source: "https://ffcap.cn/en/research/20240906a064dy00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20240906A064DY00"
---

# Unique Research | McKinsey: Why Agents Are GenAI’s Next Frontier

_Unique Research · Insight / Long Read · September 6, 2024 · Adapted from McKinsey (authors: Lareina Yee, Michael Chui, Roger Roberts, Stephen Xu)_

From information to action — imagine virtual colleagues that finish complex workflows — and you get the next productivity wave. Foundation models already extract and generate across text, audio, image, and video. The next stage is sharper: from knowledge tools (Q&A chatbots) to generative AI **agents** that run multi-step digital workflows. In short: from thinking to doing.

“Agentic” systems interact autonomously in dynamic environments. Early forms existed; GenAI’s language stack lets them plan and act, use online tools, collaborate with other agents and humans, and keep learning. Examples: a virtual assistant that plans and books a complex personal trip across platforms; an engineering agent that writes, tests, iterates, and deploys a feature from a plain-language ask.

Rule-based or narrowly trained agents were hard. Foundation-model agents inherit LLM-like adaptability to untrained scenes. Natural language — not code — becomes the instruction layer; multi-agent systems decompose work, assign specialists, use tool ecosystems, and iterate with humans.

Still early for broad commercial readiness, but capital and libraries moved fast (Google, Microsoft, OpenAI). Copilot, Amazon Q, Google Project Astra are shifting from knowledge to action; Adept, crewAI, Imbue and peers build agent models and multi-agent stacks. Agents may become as common as today’s chatbots.

## Where the enterprise value is

Agents unlock automation of **long-tail, high-variance** workflows that were never cheap or fast to encode — e.g. business travel with airline/hotel/loyalty/restaurant/activity sprawl across sites. GenAI agents help in three ways:

1.  **Manage complexity** — linear rule systems break on unforeseen branches; foundation-model agents adapt in real time to rare cases.
2.  **Natural-language instruction** — non-engineers can encode complex workflows, widening access and easing tech/non-tech collaboration.
3.  **Work with existing tools** — draw, sheet, web search, gather feedback, call other models; tool use is how agents act in the real digital world without brittle ETL glue.

## How next-gen agents work (four steps)

1.  User instructs in natural language; system identifies the scenario and asks for missing detail.
2.  System plans, assigns, executes — manager sub-agents route specialist sub-agents with domain knowledge, tools, org data.
3.  Iterate with user input until output is accurate/relevant.
4.  Execute real-world actions that finish the request.

## Three illustrative applications

**1\. Loan underwriting** — credit memos need multi-party collect/analyze/review. Agent team: relationship-manager agent, document assembler, financial-ratio analyst, reviewer loop until memo ready.

**2\. Code documentation & modernization** — legacy risk vs costly manual rewrite. Legacy-expert agent documents/translates; QA agent critiques and generates tests; reusable flywheel across migrations.

**3\. Online marketing campaigns** — strategy → personalized copy/creative → multi-platform tests across many tools. Strategy agent (web + CRM + research + multimodal models) + content/copy/design agents + human brand review, iterating for impact with brand-risk controls.

## How enterprises prepare

Invest while the stack is young. Map core processes agents could accelerate; explore APIs/toolkits (Microsoft Autogen, Hugging Face, LangChain). Three readiness pillars:

1.  **Knowledge systems** — document processes agents can learn; capture expert know-how as natural-language guidance.
2.  **Strategic tech layout** — integrate data/IT so agents plug into infra; capture interaction feedback without breaking live ops.
3.  **Human–machine control** — verify accuracy, compliance, fairness; expert co-maintenance; continuous learning loops; decide where human-in-the-loop is mandatory.

McKinsey’s State of AI: **72%+** of surveyed firms already deploying AI — put agents on the roadmap. Treat new agents like new hires: test, train, coach before autonomy. Early-stage risk is real; so is the upside of virtual colleagues that change how work runs.

_Authors (McKinsey): Lareina Yee (senior partner, Bay Area); Michael Chui & Roger Roberts (partners, Bay Area); Stephen Xu (senior director of project management, Toronto). Thanks noted to Aneri Shah, Arun Mittal, Henry Zhang, Kimberly Te, Mara Pometti, Rickard Ström._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240906a064dy00
On-site reading page: https://ffcap.cn/en/research/20240906a064dy00