OpenAI has released a report titled "AI in the Enterprise," which is very insightful, covering how to bring AI into work, how AI is reshaping new work models, how AI unlocks developer capabilities, and how to evaluate and fine-tune models. It shares lessons learned from collaborating with 7 "frontier companies." It's not just a technical showcase, but also includes practical implementation strategies. Having read this report immediately, here are the key takeaways:
Address:
https://cdn.openai.com/business-guides-and-resources/ai-in-the-enterprise.pdf
OpenAI observes that AI is bringing significant, measurable improvements to enterprises on three major fronts:
First, improving workforce performance, enabling employees to produce higher-quality work in less time;
Second, automating routine operations, freeing people from repetitive tasks to focus on high-value work;
Third, powering product innovation, providing more relevant and faster-responding customer experiences.
However, it's important to note that using AI is different from traditional software or cloud applications. Successful companies often view it as a new paradigm, embracing an experimental mindset and iterative approach to see value faster and gain support from users and decision-makers.
OpenAI itself also uses iterative development, rapidly deploying, gathering feedback, and continuously improving model performance and safety. This means that partner companies can use new technologies earlier, and their feedback directly influences the future shape of AI.
Seven Core Lessons: Real-world Experience from Frontier Companies
The report summarizes 7 key lessons, each accompanied by specific examples, packed with practical insights:
Lesson 1: Start with Evaluations (Evals) to Ensure Quality and Safety
Core idea: Before deploying to production, a systematic evaluation process must be used to measure the performance of AI models in specific scenarios. This is not just a "test" but the foundation for continuous improvement.
Case Study: Morgan Stanley
Scenario: Financial services, highly sensitive and personalized. The core need is to improve financial advisor efficiency.
Approach: Strictly evaluate each AI application (evals), specifically including assessing the accuracy and quality of language translation; assessing the accuracy, relevance, and coherence of content summaries; and comparing AI output with human experts to judge accuracy and relevance.
Effect: 98% of advisors use OpenAI daily; document information retrieval rate jumped from 20% to 80%, significantly reducing search time; advisors have more time to maintain customer relationships, follow-ups that used to take days are now completed in hours.
What are Evals? This is a process for verifying and testing model output. Rigorous Evals ensure stable and reliable applications that are more resistant to changes. It focuses on specific tasks and measures model output quality against benchmarks (such as accuracy, compliance, safety).
Lesson 2: Embed AI into Products, Create New Experiences
Core idea: Leverage AI to process massive data and automate tedious tasks, thereby creating more humanized and personalized customer experiences.
Case Study: Indeed (the world's largest job site)
Scenario: Optimize job matching and enhance user experience.
Approach: Use the GPT-4o mini model not only to recommend jobs but, more importantly, to explain to job seekers "why" a particular job is suitable for them. AI analyzes candidate background and experience to generate personalized "Invite to Apply" reasons.
Effect: Compared to the old engine, the new version saw a 20% increase in job application initiation rate and a 13% increase in downstream success rate (employers being more likely to hire). Considering Indeed sends over 20 million messages monthly and has 350 million monthly active users, this improvement has a significant commercial impact.
Optimization: To control costs and improve efficiency (due to high volume), OpenAI assisted Indeed in fine-tuning a smaller GPT model, achieving similar results while reducing token consumption by 60%.
Lesson 3: Act Immediately, Invest Early, Enjoy Compound Returns
Core idea: AI is not a plug-and-play solution; its value grows continuously through iteration. The earlier you start, the more the organization benefits from "knowledge compounding."
Case Study: Klarna (global payment and shopping platform)
Scenario: Optimize customer service.
Approach: Introduced an AI customer service assistant. Through continuous testing and optimization, within a few months, AI handled two-thirds of customer service chats, equivalent to the workload of hundreds of human agents, and the average resolution time shortened from 11 minutes to 2 minutes.
Effect: Expected to bring $40 million in profit improvement, while customer satisfaction is on par with human customer service. More importantly, 90% of Klarna employees use AI in their daily work, accelerating internal innovation and continuous optimization of customer experience across the board, allowing AI's benefits to "compound" throughout the business.
Lesson 4: Customize and Fine-tune Models to Unleash Specific Value
Core idea: Customizing or fine-tuning models for specific business data and needs can significantly enhance the value of AI applications.
Case Study: Lowe's (Home Improvement Retailer)
Scenario: Improve the accuracy and relevance of product search on the e-commerce platform.
Challenge: Numerous suppliers, incomplete or inconsistent product data.
Approach: Collaborated with OpenAI to fine-tune models. This requires not only accurate product descriptions and tags but also understanding the dynamic changes in consumer search behavior across different categories.
Effect: Product tag accuracy increased by 20%, and error detection capability increased by 60%.
What is Fine-tuning? If the GPT model is "off-the-rack," then fine-tuning is "tailor-made." Use your unique data (such as product catalogs, internal FAQs) to train the model, so it understands your business terms, style, and scenarios better, producing results that are more relevant and aligned with your brand tone, while reducing manual editing and verification and improving efficiency.
Lesson 5: Put AI in the Hands of Frontline Experts
Core idea: Those who best understand business processes and pain points are often best positioned to find uses for AI. Letting frontline experts use AI directly is more effective than building generic solutions.
Case Study: BBVA (Spanish Bank)
Scenario: Promote AI applications among over 125,000 employees globally.
Approach: Worked closely with legal, compliance, and IT security teams to ensure responsible use before deploying ChatGPT Enterprise globally. Employees were then encouraged to explore application scenarios independently and create Custom GPTs.
Effect: Within 5 months, employees created over 2900 Custom GPTs, shortening project and process timelines from weeks to hours. Applications span multiple areas: Credit risk teams use it to assess credit faster and more accurately; legal teams use it to answer 40,000 questions annually on policies, compliance, etc.; customer service teams use it to automate NPS survey sentiment analysis. Successful AI applications have expanded to marketing, risk management, operations, and more.
Lesson 6: Unburden Developers, Accelerate Innovation
Core idea: Developer resources are a bottleneck for innovation in many organizations. Using AI to build a development platform layer can unify and accelerate the construction of AI applications.
Case Study: Mercado Libre (Latin America's largest e-commerce and fintech company)
Scenario: Address the issue of overburdened engineering teams and slow innovation.
Approach: Collaborated with OpenAI to build a development platform layer called "Verdi" based on GPT-4o and GPT-4o mini. The platform integrates language models, Python nodes, and APIs, with natural language as the core interaction method, helping their 17,000 developers build high-quality AI applications faster and more consistently without delving into source code. Security, guardrails, and routing logic are all built-in.
Effect: AI application development significantly accelerated, empowering multiple businesses, such as increasing inventory capabilities by 100 times through GPT-4o mini Vision; improving fraud detection accuracy to nearly 99%; customizing product descriptions to adapt to different dialects; increasing orders by automating review summaries; and personalizing notifications to boost engagement.
Future: Plans to use Verdi to optimize logistics, reduce delivery delays, and take on more high-impact tasks across the organization.
Lesson 7: Set Bold Automation Goals
Core idea: Many processes contain a large amount of repetitive work, which is fertile ground for automation. Don't settle for the inefficient status quo; dare to set ambitious goals.
Case Study: OpenAI itself
Scenario: Internal support teams spend significant time accessing systems, understanding issues, writing replies, and executing actions.
Approach: Built an internal automation platform layered over existing workflows and systems to automate repetitive work and accelerate insights and actions. The first use case works on top of Gmail, automatically drafting customer replies and triggering subsequent actions (such as accessing customer data, knowledge base, updating accounts, creating tickets).
Effect: The platform processes hundreds of thousands of tasks monthly, freeing up human resources for higher-value work. The system is being rolled out to other departments.
Concluding Remarks
The common threads in these cases are: an open, experimental mindset, rigorous evaluation, and safety guardrails. Successful companies don't inject AI into all processes overnight but first focus on high-return, low-barrier scenarios, learning through iteration, and then scaling the experience to new areas.
The results are clear: faster processes, higher accuracy, more personalized experiences, and more valuable work.
OpenAI also observes a new trend: enterprises are starting to integrate AI workflows, leveraging tools, resources, and Agents to automate increasingly complex processes. The report mentions Operator, a "virtual employee" that can autonomously browse the web, click buttons, fill out forms, and work across systems, achieving end-to-end automation without custom integrations or APIs. For example: automating software testing and QA, interacting like a real user and marking UI issues; and updating record systems on behalf of users without technical instructions or API connections.
Hope the experiences shared by OpenAI can provide you with some inspiration.
Reference:
https://cdn.openai.com/business-guides-and-resources/ai-in-the-enterprise.pdf
Like below 👇👇