The future role of agentic AI in business: orchestration, automation or this…
What are the potential applications of true agentic AI in business? What can we learn from OpenAI Operator, Anthropic’s Computer Use and others? The future? Or red herring?
I’ve recently written about the definition of agentic AI, according to the original leaders in the field that date as far back as the 90s. In that piece, I covered how pre-determined workflow automations do not qualify as AI agents and how the hype surrounding agents today is led by technology companies with vested interests. A more appropriate question for you to ask now isn’t ‘what is agentic AI’, but ‘how can I use agentic AI’ or ‘should I use agentic AI’?
Those are the questions that I’ll answer here, with a novel take on arguably the best use case for agentic AI in business applications that I’ve heard to date, towards the end.
Agentic AI: consumer vs enterprise use cases
The very few existing genuine agentic AI examples today are found in the consumer space. A good example is OpenAI’s Operator, which can use websites on your behalf to complete tasks. Although I think this initial release is clunky, the question is; does this type of agentic AI have relevance for businesses? Will enterprises have their own AI agents like this, too?
Does Operator-like AI agents have potential in the enterprise?
Bill Hawks, Vice President and Senior Digital Product Manager, Citizens Bank, and a great acquaintance of mine, recently posted on how agentic AI could fundamentally change how banks and financial institutions serve customers.
Bill highlights a key challenge in banking: chatbots that have traditionally been limited to fixed positions within apps and websites, offering little added value if users already know how to complete tasks in those environments. Bill’s perspective is that an AI agent that can complete tasks ‘on behalf of’ users offers an opportunity to break out of the ‘app’ mentality and potentially move away from apps or websites altogether.
This is a perspective I’ve shared for some time. The notion that ambient computing platforms and micro services consumed via conversational front-ends will enable end users to get done whatever they need to get done, from wherever they are, using any device, and the most convenient modality of the moment.
However, this highlights two separate concepts that need to be unpacked to establish the role of agentic AI in buisness: ambient computing platforms and business applications. There’s an argument for and against the role of agentic AI in both areas.
Ambient computing as an entry point to business applications
The first paradigm is the ambient computing layer. The personal assistant that enables end users to access their preferred services from a single front-end. Amazon Alexa, for example, could easily have become this platform. Siri and Google Assistant, too. The technology architectures of those early platforms had limitations though; namely understanding and discoverability.
Generative AI can obviously improve the former, and some degree of agentic AI could potentially improve the latter. ChatGPT is shaping up to become that new platform, and so Bill and I are in agreement that agentic AI could allow us to break out of the siloed app paradigm.
The second paradigm that Bill alludes to is that the apps and services that this ambient computing layer connects to, such as your business, will also be agentic in nature. Technically speaking, to understand whether agentic AI is required in this instance, we need to consider a further two concepts that exist at the business application layer.
The two paradigms at the business application layer
In the same way as this ambient computing platform has two layers - the front-end of understanding and the backend of finding and triaging to apps and services - the same two layers exist at the business application layer.
You have the front-end of your business application that would connect to the ambient computing platform, taking intent and context from the platform and deciding what to do with it. And the backend of your business application that will fulfil user requests.
Is agentic AI a good fit for business process orchestration?
The orchestration layer will operate in the same way as the front-end of a current chatbot operates today, with the primary goal of understanding the user request and deciding what to do with it. At this moment in time, I contest whether agentic AI is required in this orchestration layer. Generative AI, yes. An agent? I’m not convinced.
This is because the business layer is far less complex than the ambient computing layer. The ambient computing platform (say, ChatGPT), is required to orchestrate among every type of possible thing a user wants to do. Some of it, it can do itself, such as write you an email. Others, it needs third party services, like access to banking applications. That’s potentially billions of different user needs across the billions of potential users accessing the platform.
Your business, although complex, is significantly less complex from an orchestration perspective than the ambient computing layer. You might service 300 unique customer needs. Each of those 300 needs has a specific process that underpins it. And, although each of those processes have specific nuances and conditions that create variability, the resulting complexity is still orders of magnitude smaller than the ambient computing layer.
The orchestration requirements at the business application layer is to understand and triage to the most appropriate business process based on the user request. Generative AI can be used to enhance understanding here, but it doesn’t require a completely autonomous agent to marry those things together.
Also, don’t let the integration with some potential ambient computing platform in future lead to you believe that you’ll only need your business process orchestration layer at that point. You need it now. Every touch point with your customers is an environment where conversations happen. Your AI orchestration capability should be that first touch point to triage users to the most appropriate service, channel or app.
Is agentic AI a good fit for enterprise business process automation?
Then, you have the back-end of your business application that will deliver and fulfil the user’s request. This is the areas within the enterprise where agentic AI is hotly contested.
To understand whether agentic AI is a good fit for business process automation, you first need to properly understand the real definition of agentic AI because, what you might think of as agentic AI is probably just a program that uses generative AI in some places.
The whole purpose of agentic AI is to create systems that exist in complex environments and can operate autonomously. The word complex cannot be overstated. The whole point is that the environment, and the requirements, are so complex that rules cannot be defined.
The value of AI in complex environments
A way to think about whether you need agentic AI is to consider how well you understand and can define your business processes.
Laurence Moroney, Director of Artificial Intelligence at Arm, defines the difference between machine learning and traditional software programming in a way that draws parallels between how you should think about the technology requirements needed to automate your business processes.
The difference between traditional programming and AI
In a talk I saw Laurence give a while back, he described traditional programming as something akin to:
“Developers specifying the rules that the program should follow, as well as the data that is required to make the program work. Provided that the data exists and the rules are followed, the system will provide you with an answer.”
For example, let’s say that you have an online form that users complete on your website. The data is the information in the form fields. The rules would be that, when the Submit button is pressed, do {something}, such as push the data into a CRM. This would give you your answer; data entered into a CRM. That’s an overly simplistic way of describing it, but that’s pretty much how it works.
On the other hand:
“Machine learning is concerned with providing data and answers, with the algorithm deciding the rules.”
The reason why this has been so useful is that, in many cases, you cannot possibly hope to define all of the possible rules that exist in a given environment to be able to define the outcome that you need.
For example, perhaps there is a world where you could describe, in rules, all of the different permutations that need to be considered when trying to convert speech into text. After all, every piece of speech can be broken down into phonemes that could potentially be described in rules. Every phoneme is made up of a bunch of frequencies (hundreds of thousands), and the combination of frequencies aligned to the phonemes will tell you what the phoneme is. Then, you need to piece together all of the different phonemes to work out what words have been said, and how to divide them up and put spaces in between. Lastly, you have the addition of grammar.
Now, if you were to try and describe, in rules, all of the different considerations and rule requirements to be able to take speech and transcribe it into text, you would probably be writing rules for the rest of your life, and possibly even until the end of time!
So, even though there might be a world in where you’d consider writing all of the necessary rules, it’s not this one.
Instead, it’s far better to feed an algorithm data i.e. lots and lots of speech, combined with answers i.e. lots and lots of descriptions (labels) of what the speech means (transcripts), and then have the algorithm work out what phonemes and frequencies are combined in which order to produce the text that you’ve given it. If you can do this with enough degree of accuracy, then you can generalise that model so that it can predict what the text equivalent is for any speech that it is given.
This is an overly simplistic way of describing it, but fundamentally, the only thing you need to remember is that traditional programming defines data requirements and rules in order to produce answers. Whereas machine learning is provided data and answers in order to arrive at rules.
Business processes can be clearly defined, mostly
The vast majority of business processes have logic requirements that can be defined. They’re either already defined and documented somewhere, such as in your policies, or they’re sitting in some business system somewhere, or in the heads of your people. So the biggest questions that business leaders need to ask themselves is; if I can define the process (the rules), then what value will agentic AI bring me? And the answer is, probably none.
Therefore, the only space for agentic AI in business process automation is in the areas where the environment is too complex to define.
Going back to Bill Hawks’ post; Bill alludes to this complexity in the banking space being related to having multiple different line of business systems that each do specific things, and that don’t talk to each other. This is a common situation for many organisations across industries. Bill’s argument is that a banking AI agent on the backend could potentially join the dots between these systems to enable the automation of processes and use cases that were previously off-limits.
Both Bill and Hans-Joachim Belz allude to how OpenAI’s Operator (or similar technology) has the early potential to bridge this gap in lieu of more robust system integrations.
While I don’t contest this in theory, I’m less convinced that businesses should spend efforts here, as opposed to building those integrations in the first place.
Just because OpenAI does it, doesn’t mean you should too
Operator and Anthropic’s recent example of Computer Use, for me, are short-term solutions. They’re also not strictly applicable to the business world.
These solutions are built by companies that have the goal of creating Artificial General Intelligence. For that to happen, they need to be able to demonstrate that their platforms are capable of doing multiple things as well as a human. They therefore have to solve the problem of third party integrations, and they’re doing so in the only way they can.
Amazon tried this with Skills. Google Assistant with Actions. And the biggest stumbling block was getting enough businesses active on the platform and building supported applications, which is partly what Hans is referring to above.
OpenAI tried the GPT store, which fell flat on its face much faster than the Alexa Skills Store did. The reality seems to be that relying on third party application developers to bolster the capabilities of ambient computing platforms is much easier said than done. So, what’s the alternative? For Anthropic, it’s computer use and for OpenAI, it’s browser use.
That doesn’t mean that we should look at what they’re doing and automatically apply the same logic or solutions to solving business problems. They are solving a totally different problem. Their aims are different to yours.
Computer/website use isn’t sustainable, secure or scalable for your business, in my opinion. It’s RPA with some generative AI flair.
Ron Ashri, CPTO, OpenDialog (podcast with me and Ron coming in the next week, subscribe to VUX World to get it when it drops), summarises it better than I could:
It’s an incredibly inefficient way of doing things. running a browser in a browser, applying vision models to move a pointer, etc - the energy consumption must be off the charts
The user experience currently is terrible, too slow, etc but even if it improves shouldn’t instacart just communicate with my personal agent via APIs, surely this is not the future of online shopping - humans staring at a pointer moving on a screen manipulated by an LLM
The security implications are terrifying - this tool is a scammers dream as they pair it to a database of stolen passwords
Why is openai doing it? - competition, owning the narrative and DATA so they can keep training their models
That’s not to say that AI agents have no place in business applications. It’s to say that I don’t think the comparisons made between Operator, and how that type of agent may be used in enterprise, are worth making right now
Where might agentic AI be useful in business process automation, then?
I started this piece by offering a novel approach to agentic AI for enterprise process automation, and here it is. It comes from a conversation I had with Alan Nichol, CTO, Rasa, on the VUX World podcast.
Alan’s perspective is that, if you’re in a situation where you have a process that is so complex that you cannot describe it, then yes, this could be a perfect candidate for an agentic AI application, but not one as you might think.
Instead of an agentic AI application that persists perpetually, forever; instead, Alan’s perspective is that, perhaps, you could use agentic AI in order to define the process and execute the task a number of times until you can identify a repeatable pattern. Once you can identify the pattern i.e. the rules, then you can move away from the unpredictability and probabilistic nature of large language models and define it deterministically.
This would move a business away from processes being made-up on the fly every time, and into a stable, more predictable and reliable environment. Of course, this is assuming that the complexity of the process isn’t as complex as say, orchestrating billions of user queries or translating speech to text! But if any business has processes that are so complex, I’d argue a redesign is in order in the first place!
Why use probability when you don’t want unpredictability?
Another way of describing it, Alan states, is that if you have something that is the same for every single customer, every single time, like a business process, then why take the risk on a probabilistic technology that will make-up something new every time? Instead, Alan says, you are much better off using AI in the areas that are not the same for every single customer interaction. In Laurence Moroney’s language; use AI for the areas wherein you can’t define the rules.
One of the areas that is definitely different in each customer interaction, that is impossible to define the rules for, is language and managing conversational pathways. You don’t need an agent for that, either.
In summary
The conversation around agentic AI is often driven by hype rather than practical application. While consumer-facing solutions like OpenAI’s Operator offer an early glimpse into one potential future of AI-driven task automation in the consumer space, they don't necessarily translate into valuable enterprise use cases in their current form, and are more attempts to circumnavigate integration limitations for those scale-ups.
The key question for you shouldn’t be whether agentic AI is possible, but whether it’s necessary for your business. Most business processes can be defined and automated using structured logic and traditional AI-driven workflows. If you can define the rules, you don’t need an agent. Instead, agentic AI should be reserved for scenarios where complexity is too great for pre-defined rules to function effectively. Even there, a process redesign is probably in order first.
And rather than trying to have your own version of computer use in your business for automating disparate systems, the questions is whether those systems are fit for purpose in the first place in today’s world, and whether you’d find more value in replacing them with something that has the connectivity you need.
Alan Nichol’s perspective provides a fresh take on how agentic AI could be used strategically in business. Rather than deploying AI agents indefinitely to handle tasks unpredictably, you could use agentic AI temporarily to observe, learn patterns, and extract structured rules from complex processes. Once those patterns are identified, companies can transition to a more predictable, rule-based automation approach.
This shifts generative agentic AI from being an unpredictable black box to a tool for discovering and structuring processes that were previously too chaotic to automate.
Ultimately, the future of AI in business isn’t about replacing structured automation with agents, it’s about knowing when and where agentic AI actually adds value. For now, its role may be in helping businesses uncover new automation opportunities, rather than acting as a permanent decision-maker in mission-critical workflows.