For enterprises trying to make sense of the best way to benefit from using agent-based systems, one thing should be clear: there is true benefit in optimising existing processes but now more than ever is the biggest opportunity to re-think how you build new products and offerings for your customers and your own teams- and create an “unfair advantage” for your business in the market.
It is still early days, but understanding how enterprises can leverage AI agents is quickly becoming a scramble for the c-suite. Those who figure it out now will set themselves apart from the competition.
Agentic AI is the buzz of the moment thanks to reasoning models such as OpenAI o1 and Anthropic’s Sonnet 3.7. That is to say it is an extension of the current AI conversation and many of the prominent model suppliers will be part and parcel of this industry dialogue.
The advent of the reasoning model marks an evolutionary step in capability. Now LLMs can plan a sequence of tasks geared towards achieving a goal - and execute them. In contrast, over the last several years developers and AI engineers have focused on encouraging reasoning through other (more complex) techniques such as mixture of experts (MoE), retrieval augmented generation (RAG) and extended context windows.
We also see recent agent-specific providers entering into the market that have yet to prove themselves as enterprise-grade. Moreover, we see some AI-first companies introducing agentic AI as a new feature e.g. copilot, chat and ‘agent’ are all ways of using large language models (LLMs).
The route to value depends on the type of reasoning chosen along with two new levers to consider – autonomy and granularity. How these levers are implemented will determine speed to execution and ability to acquire new market share.
This article will offer a mental model to understand the different approaches and unpack some key levers, with some real-life examples to show the value of Agentic AI across the financial services and energy industries.
A quick refresher. LLMs have two major vectors of advancement - domain specificity and task planning autonomy. Most enterprises will need some kind of balance between the two to maximise value and take advantage of the latest and greatest advancements. AI agents occupy that space.
The below shows where the opportunity for LLMs lies for the enterprise - where bespoke AI agents have the highest level of domain specificity and autonomy. Moreover, the use cases that actually achieve ROI tend not to be customer-facing chat interfaces for consumers. Rather they address highly specific problem statements that have some facet of repeatability - something that is a signal of a scaling challenge. Dario Amodei, CEO of Anthropic, alluded to a similar dynamic on an episode of Hard Fork, where Dario points to the idea that ROI lies beyond the consumer-facing products that most people are exposed to.
So, the top right hand corner is best, right? In short, yes. However, the operating space within the ‘Bespoke AI Agents’ is actually broad with at least three axes you should have in your AI vocabulary: reasoning, autonomy and granularity. The below is what you see when you zoom into that bespoke AI agents section - three agent types.
They represent the classes of agents currently available. Each has something to consider with regard to those new vocab words.
AI engineers over the past several years have encouraged reasoning through specialised techniques but now with the advent of reasoning models a form of reasoning occurs within the model instead of outside of it. Importantly, this can be mixed and matched as the use case requires. And more importantly, reasoning within the models opens the door to autonomy.
Autonomy can be measured roughly in degrees of freedom. One version of an agent has many degrees of freedom - access to all the APIs and files at your company and access to the web as well. This highly privileged agent is one you must trust, which is a quality that can be regulated by lower degrees of freedom. An agent with less autonomy might have limited access to APIs, files and the internet. It might be more robustly monitored through logging and explainability mechanisms. Again, like reasoning, autonomy is another vector upon which to build your enterprise agentic AI approach.
Yet again, granularity is another axis to play with. Humans have reasoning and autonomy. They also have specialisation, which is one route to trust. An agent might specialise on specific areas and connect with other agents that are also highly specialised, swarming together to achieve individual action that collectively sums up to communal AI-driven action.
The opportunity presented to the market is:
These risks can be seen as equivalent to the risks of any employee. However! That is not to say an AI agent equals an employee. From a practical point of view, we don’t believe AI systems will evolve in the enterprise in this way.
From an ethical point of view we risk undermining ourselves! Not to mention we need to meet the regulatory burden that requires human oversight.
We intentionally say ‘and’ instead of ‘versus’ because the dynamic between opportunity and risk in this context runs in parallel i.e. as we take on the greater opportunity granted via reasoning and autonomy we see the need for risk mitigation increase.
The broad value-add that reasoning models, and therefore Agentic AI, bring to the table is they can plan multiple steps that may otherwise require more development effort e.g. techniques that encourage reasoning. This captures the imagination as it increases the degree of freedom that AI can possess, aiding its ability to meet more complex goals. Bear in mind that goal might be more narrow than the hype might suggest - at this point in time.
This shouldn’t prompt leaders to wait for the right time to experiment and utilise agent-based systems. Early-adopters will, naturally, be better equipped to pivot as the technology and market trends shift given their technical foundations and understanding of how to apply agentic AI in their organisations.
Like any technology class there is going to be hype, disappointment and then re-finement towards ROI. As organisational leaders and stewards it can be challenging to leap-frog from hype to ROI without meeting disappointment that derails innovation.
The key to avoiding that failure will be through articulating, very well, your challenges as a business. This informs strategy, which will inform how Agentic AI can support. As you can see above there are many levers to pull on and therefore several potential architectures.
However, whatever architecture you design it must align to your strategy. Once you’re on that path, the next big hurdle is a bit different.
The single biggest challenge for CIOs is fostering and acquiring the skills required to build, improve and maintain these systems. This along with the newness of reasoning models and alignment to business challenges is why we see a chasm between enterprise-grade organisations and those operating in the top-right – currently no one.
The reality is that most enterprises are experimenting with various use cases and a wholesale agentic driven transformation is a way off.
How much value enterprises can derive from agent-based systems will depend on the strength of two factors: their data foundations, and the skill set and literacy of the workforce. Businesses can either experiment with POCs of increasing complexity and then develop their maturity, or vice versa. However, we don’t recommend just throwing models at your problem. Consider what level of maturity you have in data foundations first or in parallel – adopting reasoning models doesn’t signal maturity but putting AI in production does.
The majority of our customers have some form of AI in production, having built foundations through product-strategy alignment and data platform enablement they continue to strengthen and iterate on. Some are experimenting with expert agents and others with more complex reasoning agents, with some already in production.
Here are two standout examples we’re helping to build across Financial Services and Energy & Utilities:
Generating personalised market commentaries is costly and time-consuming for advisers. Mesh-AI's solution automates this process, using agents to summarise and refine content from trusted news feeds, while a human validates the final output. This approach enhances customer satisfaction and increases wallter share driven by personalisation. In a competitive financial market, the customer has improved portfolio risk analysis. By embedding autonomous agents into their investment platform, advisers can identify market trends faster and improve portfolio performance through real-time adjustments.
In the long-term, this technology could be expanded to include automated portfolio rebalancing based on market signals, allow advisers to cover a broader range of products or equities, and give clients more diverse portfolios to meet individual client needs and risk profiles.
We’ve built and implemented an agent-based bid response system to enhance efficiency in bid writing, increasing speed, reducing costs, and improving content quality. The system utilises information from across the organisation and enhances bid responses, making them more efficient to produce and elevating the quality of responses.
Agents specialised in both generating and then validating the responses provide an initial first draft needed by technical authors, who can then produce bids more efficiently and at a higher quality. By reducing the human oversight, this will uplift productivity by 30% and, with increased job satisfaction, reduce the multi million pound employee attrition costs.
Navigating the world of agentic AI for businesses isn't easy. There's no single solution or clear roadmap. Success depends on understanding the choices that will shape your path to value and matching those with your company's specific risk appetite and goals.
How ambitious and risk-tolerant your organisation is plays a big role. As these factors increase, so does the potential impact of your company culture and operations on successfully using agentic AI. So, it's crucial to take a careful and deliberate approach to these aspects.
This means creating a culture of innovation and learning, where employees are encouraged to try new AI tools and processes. It might also mean rethinking traditional ways of working to fit the unique capabilities and needs of agentic AI systems.
In the end, successfully adopting agentic AI in business is a complex challenge that requires strategic vision, cultural sensitivity, and operational agility. It's not just about implementing new technologies; it's about transforming your organisation to thrive in an AI-driven future.