GPT-4o, introduced by OpenAI in 2024, was a notable step in multimodal artificial intelligence. It handles text, images and audio within a single model and offered faster responses and lower cost than earlier flagship models. Businesses used it to build customer support assistants, document analysis tools, voice interfaces, accessibility products and creative applications. Although newer models have since appeared, many production systems were built around GPT-4o, and understanding the companies that develop with it remains useful.
This guide from Software Outsourcing Journal highlights five organisations that are prominent in building with or around GPT-4o. If you want to explore other AI and software stacks, you can browse our directory of Software Development Technologies.
What Makes GPT-4o Different
The defining feature of GPT-4o is its multimodal design. Instead of chaining separate models for speech recognition, language understanding and speech synthesis, a single model can process several input types and generate natural responses. This reduces latency and enables more fluid experiences, such as real-time voice conversations, image-based question answering and tools that explain what a camera sees.
For developers, the model is available through APIs and through cloud platforms that host OpenAI models. Applications can use function calling to connect the model with databases and business systems, structured outputs to produce reliable data formats, and retrieval techniques to ground answers in company documents. These patterns turn a general-purpose model into a practical business tool.
It is important to note the pace of change. Model families are updated frequently, older versions can be deprecated, and pricing and capabilities shift over time. Teams should design applications so that the underlying model can be swapped, with evaluation tests that measure quality before and after any change. Check OpenAI’s current documentation for model availability and lifecycle information.
Common Business Applications
Customer service is one of the most common uses. Assistants can answer questions, summarise tickets, draft replies and route complex cases to human agents. When connected to a knowledge base, they can provide accurate, up-to-date answers and cite sources.
Document and image understanding is another area. Companies use multimodal models to extract data from invoices, interpret charts, review forms and describe visual content. Accessibility is a particularly meaningful application: tools can describe scenes or documents for people with visual impairments.
Education, productivity and content creation also benefit. Tutoring assistants can explain concepts and adapt to learners, while internal copilots help employees search knowledge, write reports and analyse data. In every case, governance matters: human review, privacy protection and clear usage policies are essential.
How We Selected These Companies
We selected organisations with public, documented involvement with OpenAI models, including GPT-4o, either as creators, platform providers or product builders. The list is editorial and not a measure of customer count or quality, and some uses involve several models rather than GPT-4o alone. Verify current model usage, features and pricing directly with each company.
Top 5 Companies Building with GPT-4o
1. OpenAI
OpenAI created GPT-4o and provides the API, tools and documentation that developers use to build with it. It offers capabilities such as function calling, structured outputs, fine-tuning options for some models and enterprise features aimed at security and administration. Its research and product teams set the direction for the ecosystem, so organisations planning serious AI work should follow its updates closely. Working directly with OpenAI can make sense for large deployments that need support, scale and early access to new capabilities.
2. Microsoft
Microsoft offers OpenAI models through Azure OpenAI Service, giving enterprises access within the Azure security, compliance and networking framework. It also integrates OpenAI technology into products such as Copilot experiences across its software. For companies already using Azure, this route can simplify procurement, identity management and data governance. Microsoft’s partner network provides additional implementation help, making it a practical choice for regulated or large-scale deployments.
3. Be My Eyes
Be My Eyes is an accessibility company that connects blind and low-vision people with sighted volunteers through a mobile app. It has worked with OpenAI to bring image understanding to its Be My AI feature, which describes photos and surroundings to users. The project illustrates how multimodal AI can serve a real human need, and it shows the value of designing with the target community. Teams building assistive products can learn from its emphasis on safety, accuracy and user feedback.
4. Khan Academy
Khan Academy is a non-profit education organisation that developed Khanmigo, an AI tutor and teaching assistant built using OpenAI models. The tool is designed to guide students through problems rather than simply giving answers, and it includes safeguards and teacher visibility. This approach shows how careful product design can align AI with learning goals. Education companies exploring tutoring or assessment features can look to this work for design principles, while recognising that classroom products require strong privacy protections.
5. Duolingo
Duolingo is a language-learning company that has introduced AI-powered features for practice and explanations using OpenAI models. These features allow learners to practise conversation and receive feedback in more natural ways than static exercises permit. The company’s work demonstrates how a consumer app can integrate generative AI into a subscription product while maintaining a structured learning experience. Consumer app developers can study how it balances automation, content quality and user engagement.
How to Build a Reliable GPT-4o Solution
Start with a narrow use case and clear success metrics. Rather than launching a general assistant, choose a workflow with measurable outcomes, such as reducing average handling time or improving first-contact resolution. Create a test set of realistic examples and evaluate responses systematically.
Ground the model in your data. Retrieval-augmented generation lets the system look up relevant documents before answering, which reduces fabricated information and keeps responses current. Keep source references so users can verify claims.
Build guardrails. Filter sensitive inputs, validate outputs, restrict actions the model can take and log interactions for monitoring. For high-stakes domains such as finance, health or law, keep a human in the loop and make limitations clear to users.
Manage cost and latency. Choose the smallest model that meets quality needs, cache repeated results, limit context length and monitor token usage. Plan for model changes by abstracting the provider interface and keeping evaluation tests ready.
Frequently Asked Questions
Is GPT-4o still the best choice?
Not necessarily. Newer models may offer better quality, cost or features, so evaluate current options before starting a project.
Can GPT-4o process voice?
It was designed for multimodal input and output, including audio, though available features depend on the API and product version. Check the current documentation.
How do we protect confidential data?
Use enterprise offerings with appropriate data handling terms, minimise sensitive inputs, apply access controls and review the provider’s privacy commitments.
Evaluating Multimodal Applications
Multimodal products need testing beyond text accuracy. For image features, build a library of representative photos, including poor lighting, unusual angles and sensitive content. For audio, test accents, background noise and interruptions. Measure latency, because delays quickly make voice experiences frustrating. Review failure cases with domain experts and define fallback behaviour when the model is uncertain. Accessibility, privacy and consent require special care when cameras and microphones are involved: tell users what is captured, store as little as possible and offer clear controls. Monitor production usage, collect feedback and update tests as models change. Organisations that treat evaluation as a continuous discipline can adopt new model versions confidently, while those that rely on occasional manual checks often discover regressions only after users complain.
Final Thoughts
GPT-4o helped bring multimodal AI into everyday products, and the organisations above show a range of ways to apply it. OpenAI, Microsoft, Be My Eyes, Khan Academy and Duolingo demonstrate platform, enterprise and consumer perspectives. Whatever model you choose, focus on a well-defined problem, rigorous evaluation and responsible deployment. For more vendor guides, visit Software Outsourcing Journal and explore our Software Development Technologies library.