ServiceNow on Orchestrating a Symphony of AI AgentsServiceNow on Orchestrating a Symphony of AI Agents

Today, ServiceNow announced several new agentic AI capabilities for its platform: AI Agent Studio, which allows an organization to build and customize AI agents; pre-built AI Agents designed for IT, customer service, HR, and other workflows thus allowing organizations to deploy these agents quickly; and, the AI Agent Orchestrator, which enables and ensures that the “thousands of AI agents” created via ServiceNow can deliver on the various tasks assigned to them.

These new capabilities emphasize automating workflows and processes. “We've been doing workflow automation for 20 years. We understand the complexity of this environment and that our customers really want to have an end-to-end solution,” said Dorit Zilbershot, GVP of GenAI and AI Experiences at ServiceNow. “We’re moving from simple chatbots where you ask a question and get a summarized answer – which is great and saves time – to AI agents that don't just give me an answer but do the work on my behalf.”

How Do They Work?

Zilbershot says the AI Agent Orchestrator (Orchestrator) is analogous to the most knowledgeable employee in the company – it knows all the policies, the people and their skill sets, and how to solve every problem, while the AI agents are the specific tools in the toolbox.

“We call the AI orchestrator the ‘how,’ the person who plans the work. The AI agents are specialized digital workers that knows how to perform an ‘atomic’ capability,” Zilbershot said.

The ServiceNow AI agents are granted explicit access to all the tools in the ServiceNow platform – e.g., generative AI actions like summarization or generation, rules-based actions, existing workflows and scripts, as well as the millions of interactions stored within the organization’s systems (ServiceNow or third-party).

Source: ServiceNow

As for how the Orchestrator and Agents function, Zilbershot used the example outcome of onboarding a new employee. Presented with that use case, the Orchestrator devises a plan: ‘To onboard a new employee, I need to get their address, assign them a laptop, get them a paycheck, etc.’”

Each one of these sub-tasks belongs to an individual, specialized AI agent. One is ‘expert’ in obtaining and processing a driver license to create an employee record with address and identity; another is expert in hardware allocation while another is expert in making sure the individual’s paycheck is set up.

The ServiceNow platform includes many prebuilt AI Agents, and those agents can be customized – and new ones built – in the AI Agent Studio. Creating or altering agents is a mix of giving them instructions (prompting) in plain language, as well as giving them explicit access to the tools those agents need to fulfill their function. Zilbershot said that the platform also includes the ability to evaluate and test the AI agent before it is deployed, and the evaluation includes showing the full decision log of the Orchestrator and the AI agents so that customers can fully understand what is going on behind the scenes.

“If you take a first principles approach to this whole thing and deconstruct it as far as you can, we’re not talking about something new here. This is not a new invention,” said Bradley Shimmin, Chief Analyst, AI & Data Analytics with Omdia. “This is the application of several inventions that we've been working on for the last 70 years, and most recently, the last five years.”

Shimmin likened the multiple layers of AI agents acting autonomously to highly structured software, like an ‘olden days’ program written in BASIC, where there’s line after line of code and, when a function is executed (if it returns the appropriate result) the ‘go to’ instruction tells the program to jump to a specific line number.

“Software still really works that way, but it does so much more elegantly now. It’s much more adaptive and flexible and performant by introducing things like functional programming, object-oriented programming and parallelism, so that you can have multiple different subroutines functions running simultaneously, each of them returning a value, all of which goes into the ‘hopper’ and then the program says, ‘Ah I have enough now to complete this credit card transaction,’ if that is what it's doing,” Shimmin said. “What we’re talking about with agentic systems is exactly the same thing except we have a subroutine that uses generative AI's ability to do things like create a plan, analyze it, execute on it and then assess the output before it returns the results.”

Another key difference is that instead of a deterministic computer software environment, Gen AI and AI agents are probabilistic systems, “so you don't always get exactly the same output every time you run it, but for a lot of use cases, that really doesn't matter,” Shimmin said.

With respect to that latter point, Zilbershot said that AI agents and their outputs must be predictable, but predictable in a way that if it solved a specific task one day, it needs to be able to solve it the next day.

“We do see that with customers there's still a lot of education that needs to come, because even if the answer is the same, the variance in the language will be different – you won’t get the same exact words every time,” Zilbershot said. “That is something that we feel customers need to get adjusted to and how they measure quality is going to change.”

Orchestrating the Orchestrator

Although the individual AI agents are not aware of each other – any more than a screwdriver is known to the pliers inside a toolbox – they are each known to the tool wielder, the AI Agent Orchestrator. And more than that, Zilbershot said that ServiceNow is home to multiple Orchestrators.

Source: ServiceNow

“Our platform has an ‘uber’ Orchestrator which understands all the use cases. Once you identify the use case, then you get an Orchestrator that specializes in that specific task. We wanted to make sure that there's no chance of accidentally updating a customer record if all you’re trying to do is onboard a new employee,” Zilbershot said. “We wanted a controlled environment, and so each use case has their own AI Agent Orchestrator that does the planning for that specific outcome and has access to different AI agents. Then there's the ‘uber’ Orchestrator that understands which use case and which Orchestrator to apply.”

Zilbershot said that customers can choose to make any action taken by the AI agents or the Orchestrator supervised, meaning it must wait for a human to approve it, or unsupervised.

“We recommend that at the beginning, customers make ‘read’ actions autonomous. Everything that is a ‘write’ action, make it supervised. As customers gain confidence, they can change [those actions] to be fully autonomous, but they have control over that.”

“For us, the agentic platform is more than just the model. The model, obviously, is key. It's how you do the reasoning, the planning and the generation and all that,” Zilbershot said. “But [the agentic platform] is really about having the ability to orchestrate different agents – having the short-term memory and understanding the conversation, as well as long term memory to understand what happened in the past, and having all the context available. All these things coming together is what we feel makes a great agentic platform in the enterprise.

ServiceNow AI Agent Orchestrator and AI Agent Studio will be available to customers on the ServiceNow Platform in March 2025. The full suite of capabilities within ServiceNow Agents, including Orchestrator and AI Agent Studio, will be included at no additional cost for all Pro Plus and Enterprise Plus customers.

Other ServiceNow News Today

In addition to the above, ServiceNow also announced an expanded partnership, including new integrations with Google Cloud, and a new integration with Oracle.

ServiceNow and Google Cloud Expand Partnership

ServiceNow will bring its Now Platform and full suite of workflows to customers on Google Cloud Marketplace and also make its Customer Relationship Management (CRM), IT Service Management (ITSM), and Security Incident Response (SIR) solutions available on Google Distributed Cloud (GDC). ServiceNow will also make these solutions available on GDC air-gapped, for customers in highly regulated industries.

Additionally, ServiceNow will integrate its Workflow Data Fabric with BigQuery which will allow ServiceNow users with real-time, secure access to BigQuery data and enable them to enhance common CRM, ITSM, and SIR solutions, while also adding to AI Agent capabilities. The companies will also enable a zero-copy integration between ServiceNow and BigQuery.

Another new integration involves ServiceNow CRM and the Customer Engagement Suite with Google AI, which will allow customers to automate and personalize interactions across customer service channels, including self-service voice and chat conversations.

Lastly, ServiceNow launched new integrations that allow for one-click export of ServiceNow data in Sheets. Another integration with Chat will enable employees to ask questions and get help through Now Assist without leaving the productivity tools they’re working in.

ServiceNow plans to launch on Google Cloud Marketplace throughout Q2 and Q3 in various regions. The new integrations across BigQuery, Customer Engagement Suite with Google AI, and Workspace are expected to be available later this year. ServiceNow CRM, ITSM, and SIR modules to Infrastructure Operators (IO) in Google-Operated and Partner-Operated models of Google Distributed Cloud are expected to be available later this year.

ServiceNow Launches New Integration with Oracle

ServiceNow said that its Platform will be integrated with the Oracle Autonomous Database. Through this integration, ServiceNow customers can in access traditional structured data along with unstructured data directly from Oracle data sources and Oracle customers can retrieve data from ServiceNow without the need to move or copy data (aka, zero copy data sharing and bi-directional data exchange). On the ServiceNow side, this integration leverages and expands the capabilities of its Workflow Data Fabric.

This ServiceNow-Oracle integration is expected to be available to select customers in the second half of calendar year 2025.

Read more here: