Microsoft infuses Copilot on all levels of the Power Platform

The “autonomous” agents of Copilot Studio

More than 600,000 companies and entities have used generative AI functionalities, including Copilot, through the Power Platform. A statistic multiplied by four if we compare to last year.

Precisely, in Copilot Studio, Microsoft presented the preview of “autonomous agents”, that is to say agents powered by large language models capable of executing tasks independently. “They dynamically plan and learn from processes, adapt to changing conditions, and make decisions without the need for constant human intervention,” says Charles Lamanna, corporate vice president, Business & Industry Copilot at Microsoft, in a blog post.

“These autonomous agents can be triggered by data changes, events and other background tasks, not just through chat,” he adds.

This capability is made possible by the introduction of a tool called “Event triggers”. This triggers a workflow after the agent receives a payload (a JSON file) through a connector. Depending on the event and its variables, the agent developer specifies the actions he can take, for example, identifying the subject of an email, summarizing it and creating a task in Planner (the equivalent of Trello or Asana in the Microsoft ecosystem).

Low-code, RPA, agents: co-pilots for all users

At the same time, Microsoft announced agent library previews, a way to reuse components, i.e. triggers, actions, conversation topics (a list of conditions indicating how an agent should interact in certain situations), sources of knowledge, etc.

The preview of Agent Builder in Power Apps allows you to use the existing knowledge bases, subjects, logic, actions developed with Power Virtual Agents before it was called Copilot Studio. The objective? Do not redesign elements already in place.

In December 2024, Power Apps will also be eligible for the public preview of Plan Designer. Here, it is a question of relying on the “reasoning” capabilities of the model underlying Copilot (GPT-4o) to help the professions and citizen developers to design the architecture of the application (the whole of the components necessary for its operation) and the associated data model from a business problem and the roles that will use it. This process is iterative. Once the different elements meet the needs of the participants, Plan Designer creates the application template that can be modified in Power Apps.

Wells Fargo was able to test the functionality and modified the application interface in order to customize it according to its needs and by connecting it to a data source other than Dataverse. In addition, Copilot is used to generate Power Fx expressions (the Power Apps DSL taken straight from Excel) and to process documents in OCR mode. Having already been able to test Power Apps, LeMagIT understands that Microsoft intends to accelerate the development process, particularly for PowerFX and BPMN neophytes for whom the system avoids starting from a blank sheet.

After two months of private preview launched in September, Microsoft indicates that around a hundred companies were able to test Plan Designer.

An agent will also be dedicated to filling out forms, one of the most common reasons to use Power Apps.

Copilot Studio will offer the possibility of starting from a Power Apps application to create an agent. The main functionalities are translated into prompts in order to deal with repetitive manual processes. “This will be very useful to companies that already have several Power Apps applications in-house,” said Ryan Cunningham, vice president, Power Platform Intelligent Apps at Microsoft. Here again, you have to wait until December to access a public preview.

Additionally, in a few weeks, users will be able to customize the “off-the-shelf” agents offered by Microsoft. “It will be possible to add actions, sources of knowledge, consult activities, etc. This applies to the agents we develop and yours,” promises Ryan Cunningham.

Generative AI to repair RPA bots in Power Automate

As for Power Automate, Microsoft’s automation and RPA suite, it also has the right to a “facelift” with generative AI. Here, GPT-4o models perform data extraction from semi-structured or unstructured documents – emails, invoices, etc. – in PDF, JPEG, PNG format. The results can be saved in a JSON format file.

Similar to Plan Designer, Power Automate will have an early access feature that leverages an LLM to determine the next action in an automation flow. This choice will be made based on conditions expressed in natural language. For example, the LLM will check that all elements required by company policy appear on an invoice before sending it by email.

Power Automate has a set of expressions to customize automated flows. Here again, the cloud giant uses Copilot to generate these expressions from a prompt.

Another specific capability, the preview integration of a Copilot function should make it possible to learn about existing automation flows referenced in the center dedicated to Power Automate. From the Desktop interface, it will also be possible to obtain and apply recommendations to repair supervised and unsupervised RPA bots.

More operational governance and security options

Obviously, infusing AI models known for their hallucinations into business processes seems worrying on paper. To reassure its customers, spokespersons say, Microsoft intends to strengthen security and operations governance features.

First, there is the addition of a supervision interface accessible by “makers” – developers and citizen developers – from the Power Apps UI. This should make it possible to verify the proper functioning of applications created with the low-code/no-code suite. This view gives indicators about the root cause of a problem. Administrators have access to the same enriched dashboard.

“From the same interface, I can access changes made by makers as well as professional developers who recently made a commit from a Git repository. »

Above all, the administration console provides access to a view of all the components used to operate an application. Microsoft had already made it so that citizen developers and developers could collaborate through Power Apps, but the integration with Git gives a “complete view of the source code”. “From the same interface, I can access changes made by makers as well as professional developers who recently made a commit from a Git repository,” explains Clay Wesener, GPM partner, Power Apps Studio at Microsoft.

Deployment can be done from the administration console to deploy changes from a preconfigured form.

A maker can also view all deployments made and obtain a note generated by an LLM regarding the review of changes by an admin and be consulted to validate a change in the same way.

Administrators have a centralized view of application and agent deployments.

From the security tab, this same administration center provides a preview score concerning the level of cyber protection of the deployed tenants. The same provides a list of priority actions in order to reduce the risks indicated by the console. Access to the configuration parameters of these tenants is done from the same interface.

“It is not just a matter of producing a score, but of simplifying the application of best practices concerning the fine management of roles and access to applications and data in Dataverse, the addition of an IP to a fire, configuring VNET or even adding a client’s encryption key,” boasts Ryan Cunningham. “We offer features that don’t exist elsewhere in the low-code market.”

However, advanced security features will be available to customers who have opted for “premium” Power Platform licenses. They cover Power Apps, Power Automate, Copilot Studio and Microsoft Dynamics 365 environments.

Filling the Gaps in the Racket Reported by Gartner

Microsoft is far from being the only publisher of low-code/no-code solutions to infuse generative AI functions. Outsystems, Pega, Appian, Mendix and ServiceNow – not to mention Salesforce, Oracle and SAP – have all moved in this direction.

If Microsoft’s proximity to OpenAI seems to give it a head start in terms of generative AI, in its October 2024 Magic Quadrant dedicated to low-code/no-code, Gartner reports that Microsoft is lagging behind in serving the development of more complex applications with the Power Platform, as well as in governance management. Hence the adjustments presented above.

However, Microsoft is also singled out for the management of its licenses, which are both expensive and complex. Here again, the Redmond giant has some answers. First, the Power Apps admin console allows you to view the licenses relating to the tenant. Finally, Copilot Studio will be accessible under an annual subscription model (as is the case today), or for consumption in a pay-as-you-go mode. These points of attention are more salient given the size of the installed base. “Gartner estimates that Microsoft Power Apps has the largest user base of any enterprise low-code application platform,” the analysts note.