Agentic AI:

Agentic AI appearing on the corporate earnings agenda

Despite agentic AI being new to most companies, research already points to rising real-world momentum. One recent survey of senior IT decision-makers in the US found almost half were already adopting AI agents in their enterprises, while another 33% were actively evaluating agentic AI solutions.

Yet some of the misconceptions surrounding the technology, as well as the investment and challenges involved, argue for caution in agentic AI decision-making and deployment. In this issue of Perspectives, experts with hands-on experience in agentic AI share insights on navigating this evolving space, and balancing its potential with the real risks for enterprises.

Even what constitutes agentic AI – and an AI agent - is still up for debate. But in general, it refers to GenAI-based models that can grasp and work through complex, multi-layered problems with a degree of autonomy. 

As Prasanna Pendse, Global VP of AI Strategy at Thoughtworks, points out, ‘agent’ has  a double meaning, referring to something that takes action in the real world, and has the agency to do so on behalf of others. 

“Robotic agents have been around for a long time; think of the Roomba vacuum cleaners that learn to adapt to their owners’ homes,” he says. “GenAI has allowed us to imagine what a robot could be in the ambiguous, changing, volatile digital world. Agentic AI means the robot is able to deal with uncertainty by itself, to be more resilient and adaptive to the input it receives and still achieve its task. The other development is that the robot doesn't need to be physical anymore – it can live on the cloud. That is how we imagine agentic AI - autonomous robots in the cloud doing things for us, but we aren’t there yet.”   

Shayan Mohanty, Thoughtworks’ Head of AI Research, advises thinking of the current incarnation of agentic AI as a GenAI-powered large language model with a layer of code on top. This governs how the model makes its way through the various elements of a process or system – and critically, enables it to do so reliably, and independently. 

Agentic AI solutions providers are ready for prime-time

This also means there’s a lot riding on agentic AI’s success. “For VC funds to exit their investments, they need to create the narrative that this is the next big thing,” Mohanty says.  “Hardware and cloud companies also need these agents to go live in order for their compute to be consumed. Everybody is aligned in seeing the potential value; the question is how easy, and how expensive that value is to realize. And even though progress is rapidly accelerating, the answer is still largely TBC.” 

One issue is that the engineering effort required to realize the promise of full autonomy is still high. Mohanty advises enterprises keen to put agentic AI to use to look for processes governed by well-defined workflows with a clear succession of steps or states – in engineering terms, ‘state machines.’ 

“In customer support interactions, people are often operating against a script that has a decision tree structure – that’s basically a state machine,” he explains. “Building a model on top of that and engineering a system around it is relatively straightforward, as you’re essentially taking something that's written in narrative form and turning it into code.” 

For Pendse, the biggest gains from agentic AI are realized when it accelerates repetitive tasks and access to information. 

“Let's say that you need to file tickets every day, if you’re, for example, working in a bank analyzing loan applications from wealthy people,” he says. “You need to understand if these people are politically exposed, whether there are reputational risks involved in doing business with them, if the assets they claim to own are really theirs, potentially across languages and jurisdictions and different ways to access and authenticate data. Agentic AI can automate aspects of that, to help you reach conclusions and do your job more efficiently.” 

While the path isn’t always straightforward, Pendse also sees significant potential for businesses when multiple agents are able to work in concert.  

“The reality is always going to lag our expectations, but that doesn't mean there isn’t value to be had,” he says. “Typically a single agent is of limited use as it will only perform one task well; the dream is a multi-agent system that can be assigned various complex tasks. The agents then coordinate to formulate a plan of action, figure out which agents are best placed to address each component; assign roles; verify what works and what doesn’t, and iterate and improve until the task is done.”  

Thoughtworks has helped implement such a system at one major technology company, designed to improve the understanding and optimization of GPU allocations in a computing-intensive environment. Drawing on telemetry from a number of different systems, “these agents have to figure out based on the questions that are asked where to find the answers, and put them together in an accessible way, with accuracy,” Kulkarni explains.  

Thoughtworks also collaborated with the global pharmaceutical company, Bayer, to develop agents that act as research assistants in the painstaking process of drug discovery. The associated research requirements are significant, often involving the scouring of internal preclinical knowledge and other sources for evidence from previous studies in similar fields. However, the sheer volume of information is so extensive that no human can realistically access or remember all of it.

The agents designed by Thoughtworks trawl rapidly through data from historical studies, enabling drugmakers to retrieve relevant information efficiently and accelerate decision-making “In a sense its assistive memory of all the preclinical knowledge that has been generated by Bayer,” Kulkarni says. “Involving agents in the process significantly speeds up time to decision.” 

This is an example of how agentic AI can save not just time, but substantial costs by helping prevent the unnecessary repetition of tests or data-driven de-risking of drug development programs. “Finding the right information, in the right place, at the right time, is crucial in this context,” says Pendse. 

The agents in the system have been developed to handle complex inquiries, synthesize responses, and verify that all necessary information is available before presenting it to the user in the required format.

“Data scientists previously had to run multiple SQL queries that took days or weeks to produce insights that are now generated in seconds,” says Kulkarni. The system has also proven capable of flagging discrepancies or data inconsistencies in study annotations that might otherwise go unnoticed

Thoughtworks (along with Bayer) is also deploying a multi-agent writing assistant. One agent acts as a researcher, accessing thousands of historical reports and the preclinical database When a user needs to generate a report based on these specific studies, the first agent determines what information to extract. Another agent then synthesizes that information into a report tailored to the desired format and target audience. A third agent reviews and fact-checks the report, ensuring accuracy before it is finalized. This collaborative approach streamlines reporting and enhances reliability.

“I see tremendous potential in expanding the use of multi-agent writing assistants to various use cases. By leveraging these agents, we can free up resources for our scientists to focus on important tasks instead of drafting documents. This technology significantly enhances our ability to manage and synthesize vast amounts of data, ensuring that we generate accurate and comprehensive reports efficiently,” says Annika Kreuchwig, Senior Data Scientist in Preclinical Development at Bayer.

The power and diverse capabilities of multi-agent systems are the reason they are a major area of research at Thoughtworks AI Labs, where experts are aiming to work out some of the challenges that come with them.

The most recent edition of public relations firm Edelman’s benchmark Trust Barometer highlights business use of AI as an area where public trust is lacking and that has the potential to fuel more grievances. This underlines the need for companies to tread carefully. 

“If you have an agent that you're enabling, implicitly or explicitly, to perform a bunch of actions touching on a bunch of systems, you have to be able to trust it,” says Mohanty. “If it’s booking a vacation for you, knowing that it’s not going to bankrupt you is an important thing. What if it takes a hard left somewhere in the process and accidentally books everything on a private jet? The trust piece is missing at the moment - even academics haven’t really decided on the metrics that we should care about.” 

One of the first problem areas many businesses run across is privacy and security, given how essential data is to building competent AI systems. 

“A lot of times when models are training, they'll need access to specific data,” notes Kulkarni. “If your organization has personal information stored where it’s not supposed to be, the agent might gain access to all of it, and pull that data into the training process.” 

Pendse also points out agentic AI can come with perception risks. “If you’re saying that you’re going to lay off so many people, or if your employees feel like their jobs are under threat, that will almost certainly impact their morale and productivity,” he says. 

In using agentic AI some amount of downside is inevitable. The question is whether the risk is acceptable, or worth the payoff, and the answer may depend on the circumstances. 

“It might be yes if it’s being deployed in the back office, but not in the context of customer support, or the front end of your website, where you’re interacting with the public and sales prospects,” says Mohanty. 

“Every organization has its own risk profile and appetite,” he adds. “Financial services and pharmaceuticals for instance, are often more conservative, but interestingly, are also the most forward-leaning in terms of experimentation with agentic AI, because they don't want to be caught unaware. There’s a lot of experimentation but nothing’s been rolled out in a big way because the risks are very, very hard to quantify at the moment.” 

“There’s still an evaluation gap, in terms of knowing whether the AI is doing what it's supposed to do, and whether it’s reliable or not,” agrees Pendse. “In the industry, there’s no right answer or one way to get to this point. Some of our research is focused on finding the mathematical underpinnings of this in a way that you can put a number on the risk, so you’re not dependent on interpretation to say whether it’s real or not.”   

For now, Pendse advises enterprises to adopt a risk framework for agentic AI, much like those that are seen in other critical functions. 

“Think of it as a chart, with probability on the y axis and impact on the x axis, where you plot each decision or event in terms of the likelihood and severity,” he explains. “If there’s a high probability something is going to go wrong or if it would have a huge impact – like a data leak that could cost 4 percent of your annual revenue, or result in someone going to jail - it should probably be avoided. Safety is more about where you use AI and the controls you put in place than the AI itself.” 

Frameworks can help enterprises map out AI risks

“You can get to a simplistic solution that looks and feels good very quickly,” says Pendse. “But the challenge lies in taking that into production and making it scale, cost effectively, accurately and reliably. That’s likely to require skill sets that you probably haven’t emphasized before. Suddenly you need to understand statistics, calculus, differential equations, linear algebra.” 

One obvious solution is to source new talent – but for cost and other reasons, this is rarely as easy as it may seem. 

“When you hire new people who bring specialized skills, you need to keep them busy and engaged, on the right number of projects,” Pendse points out. “It requires budgets, linked to business use cases with ROI that, in this context, may be ambiguous. That’s why the answer is likely to be a mix of upskilling internal people, hiring, using contractors, and just buying finished products.” 

‘Buy versus build’ is a classic technology dilemma, “and frankly, why companies like Thoughtworks exist to begin with,” says Mohanty. “The market is evolving very rapidly, so it's a question of whether you buy today and potentially get locked into a year-long contract, during which time the entire market can shift. The AI waters are very choppy at the moment.”  

Building, meanwhile, has to be done in a relatively flexible way that allows solutions or components to be pulled out or plugged in as new innovations emerge – but this feeds straight back into the talent and capability issue. “Doing that requires context,” Mohanty says. “It requires someone who's been there and done that, and that’s frankly missing from a lot of the industry.” 

DeepSeek’s sudden emergence, and the questions it has raised about the products - and pricing models – of the leading incumbents in the AI space served as a timely reminder that most vendors “are trying to sell a particular view of the world that may change in the near future, when there's a step function increase in model capability,” notes Mohanty. 

“If you buy off the shelf, you're dependent on how the vendor is growing the product, and their product vision,” agrees Kulkarni. “But if you build, you can define your own trajectory. As the models underneath improve, your system is going to improve automatically, so your ROI is likely to grow drastically over time. If you’re talking about a use case that’s custom to your business, if you want to automate a specific business process - you probably want to think about investing in custom software.”  

The best things agentic AI brings, according to employees