Microsoft Brings OpenAI AI from Azure to Bing and Skype

On February 22, 2023, Microsoft extended the Bing preview to mobile and Skype; together with Azure, the move shows how to separate model, managed service, interface and distribution channel.

Microsoft has gone from renewing its financial commitment to OpenAI to putting its models at the center of several products in just a few weeks. Azure already offers enterprise access to these technologies, while Bing, Edge and, as of today, Microsoft’s mobile apps and Skype are adding conversational and text-generation features.

The move matters because Microsoft is not rolling out a standalone product. It is building a way to bring language models—systems capable of producing and summarizing text or holding a conversation—to the cloud used by businesses and to software with hundreds of millions of users.

Azure Becomes the Gateway for Businesses

Microsoft announced a new multiyear, multibillion-dollar investment in OpenAI in January. The company did not disclose the amount of that investment. The agreement expands a relationship that already included Azure’s supercomputing infrastructure and OpenAI’s exclusive use of the cloud.

The most immediate benefit for corporate customers is Azure OpenAI Service, which has been generally available since January. The service lets a company use models such as GPT-3.5, Codex and DALL·E through Microsoft’s infrastructure, with access controls, billing and compliance requirements that a consumer account does not provide in the same way.

That does not mean any employee can simply upload sensitive information to a chatbot. Organizations still have to decide which data they connect to each tool, who can use it and how they review the responses. The models can write fluently while also being confidently wrong. In a business environment, that combination makes human oversight essential, particularly for legal documents, customer service and financial decisions.

Bing Seeks to Change How People Use a Search Engine

On February 7, Microsoft unveiled a new version of Bing and the Edge browser powered by an OpenAI model adapted for search. Rather than limiting itself to a list of links, Bing can synthesize information from different pages and answer questions phrased in natural language.

The proposal targets Google’s core business: web search and the advertising tied to it. Microsoft starts from a smaller position in that market, but generative AI allows it to offer a different experience. For some queries—comparing products, planning a trip or summarizing scattered information, for example—a detailed answer can save users time.

The risk is that summarizing is not the same as getting things right. Microsoft acknowledged during the limited test that the new Bing could produce unexpected responses in long conversations. It therefore imposed a limit of five turns per session and 50 messages per day while it fine-tunes the system. That is a significant signal: launching a flashy feature early also means exposing its flaws to real users. Primary source

Chat on Mobile and in Skype

On February 22, Microsoft extended the new Bing to the Bing and Edge mobile apps. It also added the option to use Bing within Skype, both in one-to-one conversations and in groups. Users can invoke Bing in a chat to ask for recommendations, plan an activity or request information without leaving the conversation.

The Skype integration carries less strategic weight than Bing or Azure, but it illustrates the company’s ambition: AI is no longer a standalone web page and is beginning to appear inside familiar software. It is an interface shift. Instead of opening a dedicated app to type a request, the assistant is built into the place where people search, browse or talk to others.

A Distribution Advantage, Not Just a Model Advantage

OpenAI has shown with ChatGPT that there is mass demand for these kinds of interfaces. Microsoft brings something different: computing capacity, an enterprise cloud and established distribution channels. That combination could accelerate adoption far more than a chatbot available only through a website.

The race, however, will not be decided simply by adding a dialog box to every product. Microsoft will have to show that its answers are useful, that it cites or links to sources clearly and that customers can control how their data is used. It will also have to justify the cost of running large models for millions of queries.

For now, the partnership with OpenAI has ceased to be an investment deal. It is becoming a bet on redefining products Microsoft already had: the enterprise cloud, the search engine, the browser and communications.

Model, service, application and distribution

A company can use the same model family in products with different responsibilities. Azure OpenAI Service offers a managed route for integration; Bing presents answers to users; Skype places an assistant inside conversation. The model generates, the service manages access and operation, the application defines experience and the channel supplies an audience. Treating them as one thing hides where failure occurs.

For an Azure customer, the record begins with region, identity, logging, retention, filters and contract. Accuracy and latency come next. Bing users need citations, freshness and correction routes. Skype adds group consent: invoking an assistant may send parts of a conversation to another system, and every participant should know that.

General availability is not automatic permission for every dataset. An organisation remains responsible for deciding which collections connect, who has access and which outputs require review. The cloud supplies controls; someone must configure them and demonstrate that they work across accounts with different privileges.

Distribution can be an advantage and a risk

Microsoft could insert a feature into installed software, reducing the effort required to try it. That proximity also makes it easier to use without considering which context it receives. An evaluation records the entry point, text sent, active account, stored history and deletion route. An assistant inside a familiar tool remains an external service with its own limits.

The mobile and Skype announcement described more than one million people in preview and highlighted voice and group chat. Those figures show reach, not answer quality. A task needs different metrics: share of answers supported, time saved after review and severity of errors.

Design a manual route too. If generation fails, classic search, the original link or a human operator should remain available. Integrating AI should not remove the path that lets a person verify or complete work without it.

The transferable skill is to draw four boxes before assessing an AI product: model, service, application and distribution. Record provider, data, control and owner in each. This locates failures and avoids blaming the model for problems caused by permissions, interface or deployment.

This map also improves procurement. Instead of buying “Microsoft AI”, a contract names the service, region, model, interface and support. When one component changes, the team knows which evaluation to repeat. A shared brand no longer hides that a consumer chatbot and an enterprise API may have different terms, logs and controls.

The final comparison is task-based: public query, internal document, group conversation or automation. Each may require a different box in the stack. Choosing the task first prevents buying broad distribution for a problem that actually needed control and traceability.

The diagram should accompany the evaluation and be updated whenever any component changes.