Beyond the hype, an emerging approach to transform your business with agentic AI
Two conflicting hypotheses about the future
If you are wondering how your business should approach the adoption of generative AI, you are probably wrestling with two conflicting hypothesis about the future of technology. According to the first one AI is about to completely reshape our world, while the second recognises the signs of an over-inflated bubble that is about to burst.
They both make a lot of sense.
We expect AI agents to transform every aspect of life, democratising services that today are too expensive for most people. Every child will have access to a personal tutor delivering lessons tailored to their needs and interests. At work an agent will act as an executive assistant aligning calendars, scheduling meetings and taking care of most of our administrative tasks. If you need to create a new app, or digital service, you will not have to write any code, you will just tell an agent what you want. An AI agent will take care of organising your holidays, identifying the right combination of flights, hotels, dates, taking into account your own budget and preferences, as well as the ones of the friends you travel with.
As an end user, I would want to own as much as possible of a powerful AI agent acting on my behalf. In reality the sheer amount of information and technology resources required to create and keep one up and running will prevent me from building my own. This opens a large market for a new generation of digital services. Everyone wants a piece of it. Every start-up is now an AI start-up, every established technology vendor has turned into an AI vendor.
However AI agents will be the more effective the more they know about me, and the more they are empowered to take action. This opens a minefield in terms of privacy, data ownership and accountability. What happens when something goes wrong? Who is responsible for that? In his piece “What kind of bubble is AI?”, Cory Doctorow points to the slow adoption of self-driving cars as a case study demonstrating how these challenges can derail the large scale implementation of a promising technology. He argues that safety issues will prevent exploiting the full power of AI in high-value, high-stake use cases. This will limit the range of actionable business opportunities to low-risk use cases. How much customers will be willing to pay for those? Will that be enough to fund for the expensive infrastructure required to train and run powerful AI models?
Generative AI will change everything. Generative AI is a bubble.
Two opposing hypotheses can coexist, provided we don’t try to place them within a fixed timeline.
As humans, we have prospered as a species through our mastery of language and storytelling. Now, we’ve created machines capable of similar feats. Sooner or later this development is bound to profoundly influence our way of life. Sooner or later we are also bound to get over-excited about it.
What practical steps can you take today?
Beyond theoretical speculations, what practical steps can you take today? How can you balance risks and opportunities, place your business at the forefront, and derive real value from the emerging technology?
I think a really good place to start is re-thinking your back-office processes.
Historically, business transformation investments, guided by a sensible cost-benefit analysis, have focused on high throughput processes and on tasks that could be easily standardized and automated. Not every process falls into those buckets though. Organisations have frequently failed to truly streamline their entire operations because of the inability to automate complex tasks, where company policies need to be interpreted and applied on a case by case basis.
Think about legal and compliance activities, human resource management, procurement and vendor management processes, to name just a few. Technology has not been able to replace subject matter experts that, using their domain knowledge and experience, are able to make critical decisions applying policies and directives with the nuance required by each individual case.
That has now changed.
AI can now work with information without being constrained by its format. It can process, research and create complex documents. We can now build solutions that can assess the impact of new regulations, automate compliance and incident reporting, screen vendors, compare bids and manage contracts. The list goes on.
We still want to keep the “human in the loop” for oversight and accountability. You could argue that this significantly reduces the benefits of your investments in AI. If you still need to have an expert overseeing the process, where is the promised cost-saving?
Actually, the true value of the new art of the possible is not just running the same operations faster and cheaper. An experienced customer service agent, skilled in handling complex customer complaints, or an experienced layer drafting a corporate contract, are more productive than their junior colleagues because they have dealt with similar cases before. You might think that, when you build a generative solution to automate those tasks your aim is to re-create the “experience” of a domain specialist. It is not.
You don’t start from training an AI model on thousands of complaints and their resolutions, or a large pool of corporate contracts, aiming to create something that can handle new cases in line with what has happen in the past. That would make your business worse off. Your processes would lose auditability, transparency and agility. How would you demonstrate how decisions were made? How would you process adapt to the internal and external forces that are continually reshaping your business?
Business processes are fluid, they create value through the interaction of multiple actors that have complementary, well defined roles. This is still valid when you bring generative AI into the end to end solution.
In the field, I am noticing an emerging approach built around at least four separate AI agents: a Planner, a Researcher, an Editor and a Reviewer. Let me use an acronym and refer to it as the PRE-view pattern.
Let’s look at the responsibilities of each components:
- The planner identifies the actions that need to be taken to handle the process input. As an example, given a customer complaint, a planner can identify, through summarisation, classification, entity extraction and chain of thought reasoning, the key points raised by the customer and what information is required address them.
- The Researcher is responsible for integrating with all the available knowledge bases, within and beyond the boundaries of the organisation, using both structured queries and similarity search, to identify and retrieve the content requested by the planner. That might include customer account details, product specifications and refund policies.
- The Editor executes the bulk of content generation. The Planner has identified what needs to be done, what format the output of the process needs to take, or what system updates are required. The Researcher has collected the relevant information. The responsibility of the Editor is to create a first draft of the process output, for example an email that responds to the customer complaint or an update to the customer’s profile in the CRM system.
- The Reviewer provides feedback to the editor, validating that the output of the process addresses the needs triggered by the input, in line with the strategy defined by the Planner.
The way in which the components interact, and how many times an end-to-end flow loops through them, depends on the process and its complexity.
In simple scenarios you won’t even need Planner or a Reviewer: the pattern boils down to a Retrieval Augmented Generation (RAG). The Researcher retrieves the relevant information and the Editor performs the output generation.
In other situations the Planner and the Researcher will have to interact with each other a few times before the process can move to the Editor. Take the case of preparing a business contract. A Planner cannot just pick a corporate template with slots to fill. It has first to work with the Researcher to identify relevant laws and regulations, industry standards and companies guidelines. Only after an analysis of those it will be able to identify how to proceed and what specific information is required to create the document.
The term “PRE-view” is also a reminder of the fact that the human remains in the loop. Every AI agent must provide visibility (a “preview”) on what happens next , so that the process owner can review and influence every step of the process. Note that, instead of having to rely just on her experience, she can now leverage an AI framework that is firmly grounding the process behaviour in the latest company guidelines and corporate standards.
This is the core benefit of the approach. It is not just running the same operations faster and cheaper, it is the fact that the process becomes more auditable and transparent.
This transparency will inevitably sheds light on inconsistencies and gaps within existing business practices but you will be equipped with the necessary tools to rapidly refine your operations, enhancing output quality and efficiency.
The role of the ‘human in the loop’ goes beyond ensuring accountability; it’s about continuous management and iterative redesign of the process, aligning AI operations with evolving business goals and ethics.
This is the main difference between the PRE-view pattern and a fully automated “sinergising of reasoning and acting” promoted by the ReAct approach. A fully automated system in which an AI agent decides how to tackle a problem, takes action and then assesses the results, may expedite the process, but it will not offer the same depth of organisational learning.
Looking at the bigger picture
Let’s take a step back. We were wondering how to place our bets in a fast-moving, uncertain environment. Starting your generative AI adoption from the back office can deliver immediate benefit and offer a lower risk exposure than a customer-facing application, but how would it prepare you for the scenario in which AI agents will end up mediating most customers interactions, transforming the front office by streamlining and integrating digital experiences?
This shift could mark a significant change in how businesses operate and collaborate. Quoting Bill Gates: “AI Agents will replace many e-commerce sites because they’ll find the best price for you and won’t be restricted to just a few vendors….Businesses that are separate today — search advertising, social networking with advertising, shopping, productivity software — will become one business.”
It is probably difficult to figure out now your role in that new value network; however, having injected AI into your core operations will give you the agility to adapt and protect your position as a value supplier. Consider a business that uses AI to master complex legal compliance workflows. That capability can become a crucial value-added service as compliance grows more essential yet daunting across the network. Your business can emerge as the expert provider that partners rely on for efficient, AI-powered capabilities.
In short, enhancing back-office operations with AI now is not just about immediate gains. It’s about securing a role in the interconnected, AI-driven business ecosystem of the future.
References
[1] The future of Agents : AI is about to completely change how you use computers (Bill Gates)
[2] What Kind of Bubble is AI? (Cory Doctorow)
[3] What is Retrieval Augmented Generation?
[4] ReAct: Synergizing Reasoning and Acting in Language Models
