Final time I checked out Copilot Studio, it was a means of extending the unique Energy Digital Brokers instruments to include generative AI and assist extra common conversational interactions. By utilizing Azure OpenAI instruments to work with extra knowledge sources, Copilot Studio grew to become a way more versatile device with improved language understanding capabilities.

At Construct 2024 Microsoft charted a brand new path for Energy Platform’s AI capabilities, aligning them with the platform’s low code and no code improvement instruments and including assist for Energy Automate flows and connectors. It’s an enormous shift for the Energy Platform, however one which takes benefit of Microsoft’s adoption of generative AI as a device for constructing and operating autonomous brokers.

The brand new preview of Copilot Studio thus quantities to an entire refactoring of Energy Platform’s AI technique, evolving past chatbots to AI-orchestrated workflows. Whereas chatbots are nonetheless supported, there’s now far more to construct in Copilot Studio’s web-based, no-code improvement canvas.

Placing the robotic into RPA

On the coronary heart of the brand new Copilot Studio is an improved understanding of how fashions like GPT 4.0 can work with structured interface descriptions, like these utilized by OpenAPI to dynamically generate requests and to parse and format responses utilizing pure language. Right here, as a substitute of utilizing OpenAPI, generative AI is getting used to orchestrate current and new Energy Platform connectors, permitting you to converse together with your agent and see its responses in any supported chat shopper.

There’s rather a lot to love on this strategy. Working with lengthy transactions has at all times been an issue, and the semantic reminiscence instruments on the coronary heart of AI-driven workflows are a promising resolution, particularly after they’re used to maintain the human within the loop.

Crucial side of this redesign is the deliberate potential to make use of a set off to run a circulation that encompasses a collection of various AI-driven duties. As an alternative of being one-shot chat instruments, they’re now a approach to handle long-running transactions, modifying steps based mostly on the final set of outcomes.

How this works is simple. Let’s say you’re triggering a set of actions based mostly on an incoming e-mail. You need to use Copilot Studio to map out a workflow, launched by an occasion within the Microsoft Graph. This would possibly contain pulling within the particulars of the sender from Dynamics 365, mechanically producing a response based mostly on the incoming mail and the sender’s CRM interplay historical past, and sending a message to Groups, detailing the actions taken and itemizing doable follow-ups that require human intervention.

That’s a really completely different set of actions from these managed by the primary era of the Energy Platform’s AI assistant. It’s now a means of working with these long-running, virtually advert hoc workflows that require managing info between operations in addition to a mixture of automated and guide actions. By utilizing an agent to handle this, not solely can we ship pure language responses based mostly on utility knowledge, we are able to additionally route notifications and interactions to the fitting individual.

Constructing brokers in Copilot Studio

This requires integration with Microsoft’s varied clouds, particularly the Microsoft Graph and the Energy Platform’s Dataverse. Consequently, the most recent era of Copilot Studio leans into the circulation design metaphor of Energy Automate. As an alternative of constructing chat apps (or moderately, in addition to constructing chat apps), we’re now utilizing AI to handle and management workflows. The truth is, what we’re doing is utilizing Copilot Studio with current Energy Automate flows, so you may construct AI into current enterprise processes.

Flows are handled as considered one of a set of obtainable actions: conversational, connectors, circulation, and prompts. Conversational actions are one of many extra attention-grabbing new options, working like ChatGPT plug-ins or abilities in Semantic Kernel. They behave like a Copilot Studio subject, however as a substitute of connecting to content material, they permit your Copilot Studio agent to entry APIs and exterior knowledge. You’ll be able to even hook them as much as customized code and enterprise logic, mixing conventional enterprise software program improvement strategies with no-code AI.

Grounding with connectors and actual knowledge

One of many extra essential new options is the Copilot connector. Very similar to the connectors utilized in Energy Apps and Energy Automate, these hyperlink your utility to exterior knowledge and APIs. Instruments like this are extra essential in an AI-based utility, particularly one utilizing generative AI, as they supply the required grounding to cut back the chance of out-of-control outputs.

Usefully Copilot Studio can use current Energy Platform connectors, extending what Microsoft describes as its “data.” It is a set of knowledge sources that embrace the present chatbot instruments and Microsoft info sources like Dataverse and Cloth, in addition to utilizing the Dataverse as a means of getting ready knowledge from different enterprise sources to be used in RAG (retrieval augmented era)-driven outputs. There are limits on what you should use, with solely two Dataverse sources per utility (and solely fifteen tables in every supply). Customized knowledge from line-of-business functions is imported as JSON, prepared to be used.

Which will appear to be not very a lot knowledge, however you’re not utilizing Copilot Studio to construct and run full-scale autonomous functions; these actually require working with frameworks like Azure AI Studio’s Immediate Movement.

Including a connector to an agent in Copilot Studio is easy sufficient. Begin by including data to your utility, including an enterprise connection. These connections inherit the permissions of the consumer, making certain that customers get outcomes with out breaking safety boundaries. This strategy is crucial in case you’re constructing AI functions for regulated industries.

AI-powered workflow with conversational actions

Issues get extra attention-grabbing while you begin to use conversational actions in your functions. That is the place the underlying agent begins exhibiting autonomous behaviors, by parsing a consumer’s request and utilizing it to assemble an orchestration throughout a identified set of actions, connections, and parts, earlier than utilizing generative AI to assemble a pure language response.

Right here the consumer’s request is an orchestration immediate that’s used to begin the interplay. In a future launch the underlying system will use its data of the APIs it’s utilizing to request extra info, as vital. For now, nonetheless, you’re restricted to a helpful, if fundamental, means of including a pure language extension to an current AI utility that you simply’ve already constructed and examined in Copilot Studio.

All you could do is edit your utility, including an extension or an motion, selecting a conversational motion. You’ll then have to set some fundamental configurations earlier than enhancing the motion. A set off is a immediate that defines the motion, describing what it’s used for. That is used to find out when and the way that motion is invoked.

Upon getting the set off in place, you may then construct the motion. It is a course of circulation that has no UI. Microsoft’s enhancing device received’t present any consumer interplay parts, making certain that the method runs inside your copilot and doesn’t interrupt its circulation. As soon as printed, you may add the motion to the Microsoft 365 Copilot catalog, the place it’s handled as a plugin and activated as a part of a consumer dialog with the copilot.

The price of the upgraded Copilot Studio is surprisingly low. Because it’s a background service, it’s not licensed per-user, however makes use of a per-message pricing, with 25,000 messages for $200 per thirty days. A message is a request that triggers a response, with a message that requires generative AI operations counting as two commonplace messages. It’s not clear how one can buy extra capability at this level. There may be another $30-per-user possibility to be used with Microsoft 365 solely.

The preliminary launch of Copilot Studio gave us an uncomplicated approach to construct chatbots, infusing current applied sciences with generative AI. This new replace, now in preview, goes a lot additional, linking trendy AI instruments to course of automation, providing the promise of no-code improvement of autonomous brokers. Mixing acquainted strategies with AI-powered orchestration permits the present era of AI instruments to do what they do finest: working with well-defined, semantically wealthy APIs, and delivering ends in a human-friendly format.