AI in the enterprise

A new way of working

💡 This is an official document from OpenAI, translated into CIS. Reading time: 12 min. Lessons from seven leading companies. As an AI research and deployment company, OpenAI makes partnering with global companies a priority, because our models perform best in complex, interconnected workflows and systems. We see AI delivering meaningful, measurable improvements in three areas:

Boosting employee productivity — helps people achieve higher results in less time.

Automating routine operations — frees employees from repetitive tasks so they can focus on creating value.

Enhancing products — delivers a more relevant and responsive user experience.

But using AI is not the same as developing software or deploying cloud applications.

The greatest success goes to companies that treat AI as a new way of working. This leads to experimental thinking and an iterative approach that delivers value faster and drives greater engagement from users and stakeholders.

Our approach: iterative development

. OpenAI is organized around three teams: the Research Team advances the fundamentals of AI by developing new models and capabilities. The Applied Team turns these models into products such as ChatGPT Enterprise and our API. The Deployment Team brings these products into companies, solving their most pressing challenges. We use iterative deployment to learn quickly from real use cases and accelerate product improvement.

This means frequent updates, gathering feedback, and improving performance and security at every stage. The result: users gain access to the latest AI advances early and often — and their feedback shapes future products and models.

Executive summary

Seven lessons for adopting AI in the enterprise:
Start with evals — use a systematic evaluation approach to measure how models handle your tasks.
Embed AI into products — create new customer experiences and more relevant interactions.
Start now and invest early — the sooner you begin, the more compounding effect you gain over time.
Configure and fine-tune models — adapting models to your specific cases can significantly increase value.
Give experts access to AI — those closest to a process understand best how to improve it with AI.
Unblock developers — automating the development lifecycle multiplies the return from AI.
Set bold automation goals — most processes include routine work suitable for automation. Set ambitious goals. Next, we will look at each lesson in more detail using customer examples.

How Morgan Stanley ensured quality and safety through iteration

. As a global leader in financial services, Morgan Stanley is relationship-driven. No surprise that questions arose across the company about how AI could add value to such a personal and sensitive line of work. The answer was to run intensive evals for every proposed AI use case. An eval is a strictly structured process for measuring how an AI model handles a specific task against defined metrics.

It is also a way to continuously improve AI solutions with experts involved at every stage.

How it all began

. Morgan Stanley's first eval aimed to boost the effectiveness of financial advisors. The idea was simple: if advisors could get information faster and spend less time on routine tasks, they could give clients more and better recommendations. They ran three model evals:
Text translation — evaluating the accuracy and quality of translations produced by the model.
Summarization — analyzing how the model condenses information against agreed metrics of accuracy, relevance and coherence.

03 Comparison with experts — matching AI outputs against the answers of professional advisors, assessing accuracy and relevance. These and other evals gave Morgan Stanley the confidence to deploy AI use cases into production.

Where things stand now

. Today 98% of Morgan Stanley advisors use OpenAI daily; document access grew from 20% to 80%, and the time spent searching for information dropped significantly. Advisors spend more time talking with clients thanks to task automation and faster insights. Advisor feedback is overwhelmingly positive.

They became more engaged, and actions that used to take days now happen in hours. — Caitlin Elliott, Head of Generative AI Across the Company. What is an eval? An eval is the process of checking and testing the outputs your model produces. Rigorous evals lead to more stable and reliable applications that withstand change.

They are built on tasks that measure the quality of model output against a reference: is it more accurate? does it meet requirements? is it safe? Key metrics depend on what matters in your specific case.

Lesson 2: Embed AI into your products — the Indeed case

How Indeed makes job matching more human.
When AI automates and speeds up boring, repetitive work, employees can focus on what only people can do.
Thanks to its ability to process vast amounts of data, AI can create customer experiences that feel more personalized and human. Indeed, the world's #1 job search site, uses GPT-4o mini for new ways of matching job seekers and openings.
Simply suggesting a suitable job is not enough — it is important to explain why this particular job is recommended. Indeed uses GPT-4o mini's data analysis and text generation capabilities to craft such explanations in emails and messages.
The popular "Invite to Apply" feature now includes reasons why a candidate is a good fit for the role, based on past experience and skills.
Result after introducing AI into recruiting: —
A 20% increase in started job applications —
A 13% increase in end success (getting hired). With over 20 million messages per month and 350 million site visitors, this means a significant business impact.
To increase efficiency, OpenAI and Indeed jointly fine-tuned a smaller model that delivers similar results but with 60% fewer tokens. —
CEO: "Matching people to the right jobs is a deeply human outcome.
The Indeed team uses AI to connect people with jobs faster — a win for everyone."

How Klarna benefits from its accumulated AI experience

. AI is rarely a plug-and-play solution — real use cases grow more complex and more valuable through iteration. The earlier you start, the faster and the more value you gain from compounding improvements. Klarna, a global payments network and shopping platform, rolled out a new AI assistant to optimize customer service. Within a few months the assistant was already handling two-thirds of all chats, doing the work of hundreds of agents and cutting the average response time from 11 minutes to 2.

This project is projected to generate $40 million in profit, while satisfaction scores remained on par with human support. And none of it happened overnight. Klarna achieved these results through continuous testing and improvement of the assistant. Just as important, 90% of Klarna employees use AI every day. Mass familiarity with AI made it possible to launch internal initiatives faster and continuously improve the customer experience.

By investing early and encouraging broad adoption, Klarna sees compounding returns from AI across the business. — Sebastian Siemiatkowski, Co-founder and CEO: "This AI breakthrough in customer interaction means better service quality at a better price, more interesting work for employees, and higher returns for investors."

Lesson 4: Configure and fine-tune your models — the Lowe's case

The greatest success with AI goes to companies that invest in internal adaptation and training models on their own data. OpenAI has invested heavily in its API to simplify customization — both for self-service use and with support from OpenAI.
We worked closely with Lowe's, a Fortune 50 home improvement company, to improve search in its online store. With thousands of suppliers, Lowe's often deals with incomplete and inconsistent product data.
The key was improving product descriptions and tags, plus understanding how shoppers behave when searching, which differs across product categories.
Fine-tuning models is especially important here.
Fine-tuning results on Lowe's data: —
A 20% improvement in product tagging accuracy —
A 60% increase in error detection efficiency —
Senior Director of Data, Analytics and Computational Intelligence: "The team was thrilled when it saw the results of fine-tuning GPT-3.5 on our product data.
Product note: OpenAI launched Vision Fine-Tuning to further improve product search and address tasks in medical imaging and autonomous driving.

What is fine-tuning?

If a GPT model is an off-the-rack suit, then fine-tuning is bespoke tailoring: adapting the model to your data and needs.
Higher accuracy — a model trained on your data (for example, catalogs or FAQs) produces more relevant, on-brand answers —
Industry expertise — the model better understands professional terms, style and context —
Consistent tone and style — whether legal references or branded descriptions — everything is formatted the same way —
Faster results — less manual editing, so employees can focus on what matters

We'll curate materials for your task

clients@kt.team Telegram @kt_team_blog

We'll reply within 30 minutes and send relevant cases, diagrams, or analyses tailored to your context.

BBVA takes an expert-driven approach to AI adoption

. Employees know the company's internal processes and problems best — and often they are the ones able to find the optimal AI solutions. Putting AI in the hands of these experts can be far more effective than building one-size-fits-all solutions from the top down. BBVA, a global banking leader, has more than 125,000 employees, each facing unique challenges and opportunities.

The company decided to give employees worldwide access to AI — working closely with the legal, compliance, and IT security teams to ensure responsible use. They rolled out ChatGPT Enterprise across the company and then let people find their own use cases. — Elena Alfaro, Head of Global AI Adoption at BBVA: "Normally, building even a prototype in our business takes technical resources and time.

With custom GPTs this became easy — anyone can build an app for their own task."

Results over 5 months:

— Employees built more than 2,900 custom GPTs — Many of them cut the duration of projects and processes from weeks to hours Use cases: — Credit risk team — assesses creditworthiness faster and more accurately — Legal department — handles 40,000+ requests per year on policy, compliance, and other matters — Customer service — automates sentiment analysis in surveys

NPS AI tools are also widely used in marketing, risk management, operations, and other departments — all because employees themselves found ways to apply AI in their work. — Elena Alfaro: "We see investment in ChatGPT as an investment in our people.

AI amplifies our potential and helps us be more efficient and creative." Product note: ChatGPT can perform deep research. You enter a query and it synthesizes hundreds of sources, producing detailed, expert-level overviews — in just minutes. Internal evals showed that such research saves an average of 4 hours per complex task.

How Mercado Libre builds AI software faster and more reliably

. At many companies, developers are the main bottleneck and a brake on growth. When engineering teams are overloaded, innovation slows and ideas pile up in the backlog. Mercado Libre, Latin America's largest e-commerce and fintech company, partnered with OpenAI to build a development platform on GPT-4o. The result was Verdi — a platform layer that helps Mercado Libre's 17,000 developers speed up and unify the creation of AI applications.

Verdi combines language models, Python nodes, and APIs into a single scalable system where natural language is the primary interface. Developers can now build quality applications faster without diving into source code. Security, routing logic, and guardrails are already built in.

What was achieved:

— Higher cataloging throughput: GPT-4o mini Vision helps tag and describe products, filling the catalog 100x faster — Fraud detection: AI processes millions of product cards, reaching up to 99% accuracy on suspicious cases — Localizing product descriptions: translating and adapting to the regional language nuances of Spanish and Portuguese — Boosting orders: automatic review summarization helps shoppers

grasp the essence faster — Personalizing notifications: tailoring push messages boosts engagement and recommendation quality — Sebastian Barrios, Senior VP of Technology: "We built our ideal AI platform on GPT-4o mini, focused on reducing cognitive load and enabling the whole company to develop and ship innovations."

How we automate our own work at OpenAI

. At OpenAI we work with AI every day and constantly look for new ways to automate routine processes.

Example: customer support

. Our support teams spent too much time accessing systems, analyzing context, drafting responses, and taking the necessary actions on behalf of the customer. So we built an internal automation platform that runs on top of our existing processes and systems. It automates routine work and speeds up getting insights and taking actions.

First case: automating email handling in Gmail

. The platform accesses customer data and relevant articles, then uses the results: — to draft an email response — to take actions: updating an account, creating a ticket, and so on.

Result: — teams became more efficient, faster and more customer-oriented — the system already handles hundreds of thousands of tasks per month, freeing people for higher-value work. All of this became possible because from the start we set ambitious automation goals instead of accepting inefficient processes as a "cost of doing business."

Learning from each other

As the previous examples show, every company has opportunities to apply AI for better results.
Use cases may differ by industry and scale, but the principles remain universal.
AI adoption delivers the greatest return when paired with an open, experimental mindset combined with rigorous evaluations and safety measures. Companies that succeed do not rush to embed AI into every process — they start with simple but profitable use cases, learn from them, and then carry that experience over to new areas.
A more personalized customer experience —
More meaningful work — employees do what people do best
We now see companies integrating AI into complex processes, often using tools and agents to achieve results.
We will keep sharing observations from the front lines so you can apply this knowledge in your strategy.
Product note: Operator is an example of OpenAI's agentic approach.
Its own virtual browser lets it navigate sites, click buttons, fill out forms and collect data the way a human would.
On top of that, it can run processes across different systems and tools — with no integrations or APIs.
Use cases in companies: —
Automated software testing — Operator works like a real user, surfacing bugs in the interface —
Updating accounting systems on behalf of users — without technical instructions or APIs. Result: full end-to-end automation that frees teams from routine and boosts overall efficiency.

Security and privacy

For our enterprise customers, security, privacy and control matter most.
Your data stays yours — we do not use your content to train models; your business retains full ownership.
Compliance with corporate standards — data is encrypted in transit and at rest.
Compliance with SOC 2 Type 2 and CSA STAR Level 1 standards.
Flexible access control — you decide who can view and manage the data.
This ensures compliance with governance requirements.
Data retention settings — flexibly configure logs and storage in line with your organization's policy.

Additional resources

OpenAI for Business Customer Stories OpenAI ChatGPT Enterprise OpenAI Security API Platform OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.

Contacts

Let's Discuss Your Project

Leave your current contact details and describe your task. We will come back with clarifying questions and a proposal for the next step.