AI that generates gives way to AI that executes? Genpact prepares for autonomous agents in the enterprise

A theme that will be discussed a lot this year and next is preparedness for autonomous agentic AI. At present the focus is on assistive agentic AI, rather than on autonomous agentic AI. Genpact's recently published "Autonomy by Design" research provides a useful check-in as to where organizations are in getting ready for this next phase of agentic AI adoption.

The firm believes that organizations that are transitioning from 'AI that generates' to 'AI that executes' are on the path to becoming autonomous enterprises. The research concludes that optimism outpaces readiness, since 50% of executives questioned, expect autonomous AI decision-making across strategic functions within three years, yet only 12% are positioned to get there today.

AI adoption is now more of a process integration challenge

An issue that features in the research commentary based on the survey is the complexity of technology. For example, the study states:

AI doesn't fail because the models are weak; it fails because integration is," says Anu Dixit, global chief customer officer at Resolution Life, a global life insurance group. "Fragmented data and immature enterprise wiring prevent AI from scaling.

Sanjeev Vohra, Genpact's Chief Technology and Innovation Officer, says that this sentiment indicates:

The constraint is not necessarily data and that organizations are moving on from experimentation managed by data scientists and have passed over data cleaning and curation and are now moving into the mainstream application landscape. They are thinking about their legacy systems and apps, ERP, CRM and so on and how they build new AI systems.

Vohra thinks that the 12% identified as leaders by the research do have a change management structure in place and an advocacy suite of leaders. The research goes on to point out that workforce capability gaps continue to be the most frequently cited organizational constraint to AI adoption, as reported by six in 10 executives - yet only 45% say their organizations offer AI training for all employees. In tandem, 40% identify fragmented ownership and accountability as key challenges. While leaders (the 12%) have done more to overcome these barriers, they've not eliminated them yet.

Vohra explains:

We test and validate our clients' thinking to ensure they understand that everybody has to believe in the change. For example, it is a good idea to think about new labels for roles so that people think about how they are preparing for the new role rather than thinking that the technology is taking their job away. Instead of being an underwriter, you are an underwriter powered by AI and this technology is helping you migrate to that new role."

It is hard to implement AI unless the human workforce embraces it because you need human intelligence to inform AI adoption. As Nelson Repenning, Distinguished Professor of System Dynamics and Organization Studies at the MIT Sloan School of Management comments in the report:

AI poses an interesting paradox. On the one hand, it's main benefit will come from automating the work that humans currently do. On the other, however, you can't automate the process of implementing automation. Humans must still make the tough, messy decisions about where and how to apply AI

AI Maestro - automating the process integration

Unsurprisingly, the research reveals that almost everyone says their organizations lack adequate governance models for autonomous agentic AI systems. Governance thus remains an extremely critical topic. Vohra continues:

AI apps are flourishing and there is pressure to generate value from AI. Organisations have aligned their IT and AI is being used for experimentation. But this means that organisations may be using five different systems to address the same problem. There is also the risk perspective in terms of using new vendors that may not be so mature in their environments. These are concerns for CIOs and CTOs because these roles are now involved, AI adoption is no longer just the preserve of the Chief Data Officer, and that is why the embedding of AI into the workflow is becoming more important.

We think for speed and agility you need a centralised approach to addressing compliance via auditing, and then a more distributed, federated approach to how work is done. Governance itself should be automated. For example, within Genpact we have multiple agents that we are using in different parts of the workflow. We are using AI to govern AI: Genpact's AI Maestro creates a system of work bringing together humans and AI agents. It is an operating system to manage the agent workforce, providing guardrails by design.

AI Maestro is described as providing a dynamic orchestrator to guide both human and digital agents based on learned process intelligence from more than 1,000 AI models and 25 datasets, with a context interface to ensure agents are meeting humans in their preferred apps. This capability is not yet ready to offer as a standalone asset for customers to use across their organisations, rather it provides an accelerator for Genpact to build enterprise architecture for clients.

My take

The successful enterprise preparation for autonomous agents will be a slog and requires an ongoing commitment to change management. Consequently, choice of IT service partner is important, as the partner needs to be able to drive value jointly with the enterprise over time. This is all about outcome-based contracting, flexing with organisational and process change and so joint ownership and accountability is critical. Arguably IT service providers with Business Process Service (BPS) experience have more credibility here than other service providers. Genpact has this pedigree making it one to watch as the enterprise market for autonomous agents develops.