Introduction

Most AI failures aren’t due to a lack of intelligence in the models, but rather the absence of context engineering—information wasn’t properly “written, selected, compressed, or isolated.”
Ignoring context equates to real financial losses: from the Bard launch debacle to “260 chicken nuggets,” companies are paying the price for memory deficiencies.
Blindly extending context only amplifies noise and attack surfaces; small and precise context management is the key to performance and security.
Establish context first, then discuss large models: common gains include 80% reduction in input costs and 15-90% increase in accuracy, making it far more cost-effective than switching to larger models.

Business practices from 2023 to 2025 demonstrate that AI application failures stem not from a lack of intelligent models but from the absence of “context engineering.” Google lost $100 billion in market value, while companies mastering this technology experienced a 40-90% performance increase.

I. The $100 Billion Lesson: What Happens When AI “Forgets”

The Fatal Blow to Google’s Bard

In February 2023, Google confidently presented its AI chatbot, Bard, to the world. However, during this highly anticipated unveiling, Bard made a shocking mistake.

When asked about the achievements of the James Webb Space Telescope, Bard confidently replied, “It captured the first image of an exoplanet.” While this sounds impressive, there was a critical flaw—it was incorrect. In reality, the first image of an exoplanet was taken in 2004 by the European Southern Observatory, nearly 20 years before the Webb Telescope was launched.

This seemingly minor error triggered a snowball effect. Investors immediately realized that if Google’s AI couldn’t accurately grasp basic facts, how could it reliably operate in more complex business scenarios? On that day, Alphabet (Google’s parent company) saw its stock price plummet by 9%, resulting in a loss of over $100 billion in market value. [Source: CNN, NPR, Time reports]

Canada’s Costly “Misleading”

By the end of 2023, Canadian passenger Jake Moffatt urgently needed to buy a ticket due to his grandmother’s passing. He consulted AI customer service from Air Canada, which provided a seemingly thoughtful response: “You can buy a full-price ticket first, then request a bereavement discount refund within 90 days.”

Following the AI’s suggestion, Moffatt was told when he applied for the refund that bereavement discounts must be requested before purchasing the ticket and cannot be made retroactively. The AI customer service had provided completely incorrect policy information.

This case eventually went to court. The Canadian Civil Arbitration Tribunal made a landmark ruling: companies must be held legally accountable for erroneous recommendations made by their AI systems. Air Canada was ordered to pay $812.02 and update its AI systems. [Source: CIO’s coverage of AI disaster cases]

McDonald’s “260 Chicken Nuggets” Nightmare

In June 2024, McDonald’s ended a three-year AI ordering partnership with IBM. This decision was backed by a series of laughable failures.

The most infamous incident occurred at a McDonald’s drive-thru. A customer, initially wanting to order a few nuggets, found the AI system malfunctioning and repeatedly adding more nuggets to the order. Despite the customer frantically shouting “Stop! Stop!”, the AI continued to ignore them, ultimately resulting in an order of 260 McNuggets.

This video went viral on social media, becoming a classic example of AI failure. McDonald’s had to shut down the AI testing systems in over 100 stores, effectively rendering three years of research and development null. [Source: CIO’s analysis of corporate AI failures]

II. Unmasking the Truth: It’s Not That AI Isn’t Smart, But Its “Memory System” is the Problem

A Genius with Severe Alzheimer’s

Imagine this scenario: you hire a top expert with an IQ of 180 as your assistant, a genius in various fields with exceptional calculation abilities. However, there’s one issue—he suffers from severe short-term memory loss and forgets previous conversations every few minutes.

This is the reality for most corporate AI systems today. They do not lack “intelligence” (model capabilities) but rather effective “memory management” (context engineering).

What is “Context”? Understanding Through Meeting Minutes

In our daily work, “context” is ubiquitous. Imagine you attend an important project meeting:

Meeting Background: Why is this meeting being held? (akin to AI’s system prompts)
Historical Records: What was discussed in previous meetings? (akin to conversation history)
Relevant Documents: Reports, data, contracts that need to be referenced (akin to a knowledge base)
Attendees: Roles and permissions of each participant (akin to tools and permissions definitions)
Meeting Minutes: Key decisions and action items (akin to memory summarization)

Without this “context,” even the brightest experts cannot make correct decisions. This is the fundamental reason behind Google Bard’s error—it lacked precise historical data and a mechanism for fact verification when responding to questions.

Manufacturing’s Painful Lessons

According to research by Gartner, the manufacturing industry faces particularly severe challenges in AI application:

Only 20% of generative AI projects are deemed successful.
85% of AI projects fail to meet expectations.
42% of companies plan to abandon their AI initiatives by 2025 (this figure was only 17% in 2024).

[Source: Appinventiv, SupplyChainBrain’s manufacturing AI report]

Why is the failure rate in manufacturing so high? The answer lies again in the absence of context engineering:

Historical Data Gaps: New AI systems can’t access critical production data from legacy systems.
Lack of Real-Time Information: The AI cannot see the current equipment status or inventory levels when making decisions.
Knowledge Silos: Different departmental AI systems operate independently, unable to share critical information.

III. Context Engineering: The Solution to Equipping AI with “Complete Memory”

Equipping AI with a “Smart Secretary”

The essence of context engineering is akin to providing your AI system with a highly competent secretary. The secretary’s role involves:

Recording Important Information (Write)
- Save key decisions and conclusions
- Similar to how a secretary organizes meeting minutes
Selecting Relevant Materials (Select)
- Identify current needs from vast amounts of information
- Like a secretary preparing relevant documents for you
Summarizing Key Points (Compress)
- Condense lengthy reports into essential insights
- Just as a secretary drafts executive summaries
Coordinating Team Roles (Isolate)
- Allow different experts to handle their areas of expertise
- Similar to how a secretary arranges specialized meetings

Real Case Study: The Impressive Transformation of an Insurance Company

Five Sigma Insurance radically transformed its claims processing by implementing context engineering: [Source: MarkTechPost Case Study]

Pre-Transformation Challenges:

The AI system frequently provided claim advice that contradicted policy terms.
It failed to recognize fraud patterns because it could not see historical claims data.
It often made errors in handling complex cases.

After Implementing Context Engineering:

The system could simultaneously access: policy terms, claims history, regulatory requirements, and fraud databases.
Claim processing errors reduced by 80%.
Claims adjusters’ efficiency improved by 25%.
Underwriting accuracy exceeded 95%.

The key was that they did not change the AI model; they simply improved how information was organized and delivered.

Microsoft’s Developer Tool Revolution

Microsoft’s AI programming assistant showcases the power of context engineering: [Source: Microsoft Official Blog]

By integrating the following context information:

Developer project history
Team coding standards
Relevant technical documentation
Codebase dependency relationships

Results Achieved:

Software task completion rate increased by 26%.
Code errors decreased by 65%.
New hire ramp-up time reduced by 55%.
Code quality improved by 70%.

IV. The Pitfalls of Long Contexts: Why “Remembering More” Does Not Equal “Performing Better”

AWS Security Team’s Warning

In 2024, the AWS security research team uncovered a critical issue: when AI systems’ “memories” become overloaded, deadly vulnerabilities arise. [Source: Technical Analysis from Towards Data Science]

Imagine this scenario: your AI assistant needs to process a 1000-page report. Theoretically, the new AI model could “remember” everything. However, what actually happens is:

Important instructions at the beginning get “pushed out” of memory.
Malicious users can “pollute” the AI’s memory with irrelevant information.
The AI begins to hallucinate, making decisions based on incorrect information.

It’s like a person trying to remember an entire encyclopedia at once—too much information creates chaos instead of clarity.

Tesla’s Solution for Autonomous Driving

Tesla’s Full Self-Driving (FSD) system is one of the most complex implementations of context engineering: [Source: Tesla Official Website, Wikipedia]

48 neural networks work in tandem.
Each time step outputs 1000 different tensors.
Processes real-time video streams from 8 cameras.
Has cumulatively driven over 1 billion miles.

How does Tesla manage such a massive flow of information? The answer lies in “smart filtering”:

Not all information is equally important.
Urgent information (like a suddenly appearing pedestrian) takes priority.
Historical information is stored according to its importance.
Different neural networks are responsible for different types of information.

V. Major Breakthroughs from Industry Giants: What We Learned from a $100 Billion Lesson

OpenAI’s Model Context Protocol (MCP)

At the end of 2024, OpenAI introduced the revolutionary MCP protocol to tackle the “M×N problem”: [Source: Pluralsight, Microsoft Learn]

Challenges with Traditional Approaches:

10 AI models × 100 data sources = requires 1000 custom interfaces.
Each interface requires separate development and maintenance.

MCP’s Solution:

Create a unified “universal language.”
Any AI model can access any data source through standardized interfaces.
Reduces integration costs by over 90%.

Anthropic’s “Constitutional AI”

Anthropic, the developer of Claude, adopted a unique approach: [Source: Anthropic Official Research]

They invited 1000 U.S. citizens to help establish the AI’s “code of conduct,” ensuring the AI system:

Understands and adheres to human values.
Makes ethical decisions in complex situations.
Reduced the success rate of malicious exploitation from 86% to 4.4%.

Google’s Gemini with Million-Token Context

Learning from Bard’s failure, Google’s Gemini 1.5 Pro achieved: [Source: Google Official Blog]

Stable context of 1 million tokens (equivalent to 700,000 Chinese characters).
Simultaneously processes audio, video, text, and code.
Capable of analyzing entire movies or hundreds of pages of documents.

However, Google also admits that more context doesn’t equal better performance; the key lies in how to organize and utilize that information effectively.

Microsoft’s Intelligent Routing for Azure

Microsoft provides multiple model variants within Azure AI Foundry: [Source: Microsoft Azure Blog]

GPT-5: 272K context, suitable for complex reasoning.
GPT-5 mini: optimized for real-time experiences.
GPT-5 nano: ultra-low latency responses.
Intelligent routers automatically select the most suitable model, saving 60% in costs.

VI. Multi-Agent Collaboration: Practices from Amazon and Walmart

Amazon’s Army of 750,000 Robots

Amazon’s warehouse automation system showcases the power of large-scale context management: [Source: Amazon Official Coverage, LinkedIn Analysis]

750,000 mobile robots deployed in 2023.
The Sequoia system reduced order processing time by 25%.
Through route optimization, it saved 30 million miles in travel distance.
Reduced CO₂ emissions by 94 million pounds.
Package damage rates remained below 0.1%.

The secret to success lies in “layered context management”:

Each robot only needs to know its task.
Area controllers coordinate local groups of robots.
A central AI system oversees global optimization.

Walmart’s AI Inventory Revolution

Walmart’s AI system deployed in over 4,700 stores integrates: [Source: Walmart Official News, Walmart Tech Blog]

Multidimensional Context Information:

Historical sales data.
Weather forecasts (affecting purchase patterns).
Macroeconomic trends.
Local demographics.
Social media trends.

Unique Innovations:

“Anomaly Forgetting” patented technology: automatically excludes the impact of one-time events (e.g., pandemic stockpiling) on predictions.
Dynamic adjustment algorithms: real-time adjustments based on holidays and promotions.

Results:

24% growth in Q3 2023.
Route optimization avoided 30 million unnecessary driving miles.
Aiming for 65% store automation by FY 2026.

VII. GE’s “Humble AI”: Knowing What It Doesn’t Know

The Wisdom of 1.2 Million Digital Twins

General Electric (GE) created over 1.2 million digital twins between 2016 and 2017, generating $600 billion in value: [Source: Emerj, Microsoft Research]

Their “humble AI” framework is particularly noteworthy:

The AI system can recognize its limits.
When encountering situations beyond its understanding, it automatically switches to safe mode.
Actively requests human expert intervention.

Practical Outcomes:

Increased wind farm output by 20%.
Prevented 400 occurrences of unplanned maintenance each year (in the aviation sector).
Reduced unplanned maintenance by 30% through predictive maintenance.

This approach avoids catastrophic consequences stemming from AI “pretending to know.”

VIII. Four Core Technologies of Context Engineering

Based on research by experts like Phil Schmid and Lance Martin, along with practices from LangChain and LlamaIndex, context engineering encompasses four core operations: [Source: philschmid.de, rlancemartin.github.io, blog.langchain.com]

1. Write: Establishing AI’s “Long-Term Memory”

Just as humans keep diaries and notes, AI systems also need to capture important information:

In-Session Writing:

Temporary drafts (e.g., calculation processes).
Intermediate thought processes.
Planning for current tasks.

Persistent Writing:

Summaries of user preferences.
Key business rules.
Records of historical decisions.

Applications like ChatGPT and Cursor leverage this method, allowing AI to “learn” and “grow” through continuous interaction with users.

2. Select: Finding the Most Relevant Information “Right Now”

Imagine your assistant needs to prepare a report; they wouldn’t bring the entire library, but instead choose the necessary materials precisely:

Deterministic Selection:

Always loading certain essential documents (e.g., company policies).

Model-Driven Selection:

Letting AI assess what information it needs.

Retrieval-Based Selection:

Using similarity searches to find relevant content.

3. Compress: Turning “War and Peace” into a One-Page Summary

When information overwhelms, smart compression is necessary:

Automatic Summarization:

Condensing a 1000-word email into three sentences.

Priority Ordering:

Retaining the most critical 20% of information that covers 80% of the value.

Incremental Updates:

Recording only the parts that changed instead of making complete copies.

4. Isolate: Teamwork Among Expert Agents

Complex tasks require collaboration from multiple AI experts:

Task Decomposition:

Financial analysts handle numbers.
Legal experts review compliance.
Writing specialists prepare the final report.

Information Isolation:

Each expert receives only relevant information.
Prevents information overload and confusion.

Result Integration:

Main AI synthesizes input from all experts.
Makes the final decision.

IX. Return on Investment: Why Context Engineering More Cost-Effective than Upgrading Models

Impressive Cost-Benefit Ratio

Industry data shows that the return on investment for context engineering far surpasses that of model upgrades: [Source: Aggregated multiple case studies]

Context Engineering:

Constitutes 5% of AI budgets.
Yields 40-90% performance improvements.
Implementation period: 2-3 months.

Model Upgrades:

Constitute 60-70% of AI budgets.
Yield 10-20% performance improvements.
Implementation period: 6-12 months.

Real Bills from a Tech Company

Data from a medium-sized tech firm reveal:

After implementing context engineering, they saved $23,000 per month in computing costs.
Input size reduced by 80% through context trimming.
API call costs correspondingly decreased by 80%.
Performance actually improved by 15%.

It’s like better traffic planning, saving both fuel costs and commuting time.

X. Outlook for 2025: The Crucial Step from “Demonstration” to “Production”

Industry Experts’ Consensus

“The failure of most AI agents is no longer due to model failure, but context failure.” This has become an industry consensus.

Cognition (the team behind Devin AI) explicitly states: “Context engineering is the primary task in building AI agents.” [Source: cognition.ai blog]

Three Action Recommendations for Businesses

1. Conduct an Immediate “Context Health Check.”

Document specific scenarios where your AI system has failed:

What information was lacking when the AI provided incorrect answers?
Where are there gaps in information?
What data sources can the existing systems access?

2. Choose a High-Value Pilot Project.

Don’t attempt to overhaul all systems at once—select one:

With high usage frequency.
With significant cost of failure.
With obvious potential for improvement.

For example: customer service, order processing, report generation.

3. Establish Cross-Department Collaboration.

Context engineering requires:

IT Department: technology support.
Business Department: defining information needs.
Data Team: ensuring data quality.
Compliance Team: ensuring information security.

Avoid Common Pitfalls

Pitfall 1: Blindly Pursuing Large Models.

Incorrect thinking: The bigger the model, the better.
Correct approach: Optimize context first, and then consider upgrading the model.

Pitfall 2: More Information is Always Better.

Incorrect thinking: Provide AI with all possible information.
Correct approach: Provide relevant information precisely.

Pitfall 3: Ignoring Information Quality.

Incorrect thinking: As long as there’s information, it’s fine.
Correct approach: Ensure information is accurate, timely, and structured.

Conclusion: The Beginning of a New Era

The years 2023-2025 will be remembered as the “Year of Context Engineering.” From Google’s $100 billion lesson to the successful practices of Tesla, Amazon, and Walmart, we have observed a clear trend:

The success of AI no longer depends on a “smarter brain,” but rather on a “better memory system.”

Companies mastering context engineering are gaining a sustainable competitive advantage:

Vastly improved operational efficiency.
Significantly enhanced customer experience.
Multiplicative return on investment.
Substantially reduced risks and errors.

Conversely, those ignoring this trend could end up like companies that once missed the Internet revolution, left behind by the relentless march of time.

As one industry leader aptly noted: “In the era of AI, context engineering could be the most profitable part of your AI investment.”

It’s time to reassess your AI strategy—not by asking “Do we need stronger AI?” but by inquiring “How can we enable our existing AI to better understand and remember key information?”

The answer lies within context engineering.

This article is based on the practical cases of internationally leading companies from 2023 to 2025, with all data sourced from public reports and official releases.