AI agents are coming (but maybe not quite yet)

‘Short-term we’re too optimistic about what AI can do; long term we’re too pessimistic’ says Celonis AI lead

Earlier this year, no tech event was complete without a GenAI announcement. Six months later the story has moved on, and now the talk is all about agentic AI.

At Celonis' Celosphere 2024 event in its home city of Munich this week, the process intelligence (PI) company joined the fray, announcing the launch of AgentC, which VP of product marketing Vivya Krishnan (pictured) described as a "suite of AI agent tools, integrations and partnerships that enables our community to develop AI agents in the leading AI agent platforms."

These include integrations with AWS Bedrock Agents, IBM WatsonX and Microsoft Copilot, with CEO Alex Rinke demonstrating the ability to create bespoke agents to manage business processes using Microsoft's ecosystem during his keynote address.

Assistant or agent?

While the conversational shift from AI assistants to the more buzzy AI agents accurately reflects the growing operational capabilities of the technology, the distinction is more presentational than practical, AI lead Cong Yu told Computing during a press briefing.

"What we care about is whether we can build something AI powered for our customers' use cases. Whether you call it a system or an agent doesn't really matter to us, but we do think the world is going to evolve in a very agentic fashion, with a lot of multi-step reasoning and goal setting."

And Celonis certainly has a better pitch for agentic AI than most vendors, given that its platform links data and events from multiple sources including ERP, CRM, databases, data lakes, data warehouses and collaboration tools, tying business process KPIs to business outcomes. As such, the company has access to all the pieces of the puzzle and can, how they interact and where the bottlenecks are.

See also Unclogging the information arteries of the NHS

"Through process mining we have data about how processes actually flow, and through process modelling we can show how processes should flow," said Krishnan, adding that for the enterprise: "There's no AI without PI."

Illuminating hidden connections and unseen correlations and bridging the gap between real and desired process flows is the raison d'etre for PI, but currently this is largely an analytical process, with the human very much in the driving seat, using that information to bring processes closer to the "happy path".

Nevertheless, there is scope for PI to be more operational, with intelligent automation of actions, particularly in the case of process exceptions. These are deviations from happy path which often involve manual workarounds by employees who, for whatever reason, are not using systems as intended, for example, unnecessarily copying and pasting information from one application into another, or using email rather than automated messaging.

In the first instance, an AI agent could suggest a course of action (“why not enable automated messaging?”) which is then validated by the employee, with the possibility that the agent will be allowed to act autonomously in the future as the approach becomes standardised.

A stepwise approach has several advantages, not least accustoming nervous employees in using AI to make their lives easier. It also manages the unpredictability problem, which is an unavoidable feature of GenAI: outcomes are not deterministic.

"We're still talking about human in the loop," said Yu. "With most customers you want to start with use cases that you can have a handle on. That means that agents are in the position to come up with recommendations. But I do expect to see that evolve.”

He continued: “Going more autonomous will hinge on a lot of different factors, including governance requirements, the accuracy that the organisation is seeking, and just getting the workforce acclimatised to interacting with AI."

Agents as processes

Looking into the future, agents will themselves become just so many processes, to be tracked and monitored and nudged onto the happy path using process intelligence, perhaps acted upon by other agents, he predicted.

And yet for now, as with so much business GenAI, autonomous agents capable of reasoning and acting are a nascent development, an evolution rather than an evolution, and one where the hype is outpacing the reality. But that does not mean they are not coming.

"Short term, I think we are too optimistic about what AI can do," said Yu. "We think AI will do everything, and that's causing some disillusion.

"But in the long term I think we are too pessimistic. It is becoming much more capable. [Business] will be way more autonomous than it is now, and our definition of agents will evolve along with that."