Designing process-oriented AI solutions

When we take an AI solution to a business, it’s easy to show cool stuff, but it’s tough to show what any of it will really mean. Within the companies that work with these solutions, there sure is a lot of hype, but when you’re a cash-strapped business owner looking to make a tangible impact, well…it can be hard to clearly envision one.

As I see it, there are grades of AI solution. If we move up enough, we can show true ROI reliably, without relying on user adoption rates. I’ll just talk through what I see as the first three. There are definitely more, such as at the strategic layer, but I want to focus on core processes.

AI Solution Steps (1-3)

Step 1 - generic AI

Think ChatGPT. An intelligent chatbot. It can respond to prompts and event refer to documents and images. Models like o1 can even apply reasoning to “think” through problems with greater depth. In the end though, the output is generated within the same interface, whether it be a video, image and/or text. The user needs to provide the complete prompt, including any context, within the one thread. From there they either read the response or read it and copy-paste it somewhere else. It’s taking the most advanced technology ever developed and bottlenecking it with something as basic as a clipboard.
Fundamentally, the biggest constraint from an ROI perspective is that this kind of tool is all about adoption. People need to actively seek it out and use it effectively to drive results, and such results can only be at best inferred post-hoc.

Step 2 - agentic, integrated AI

This is where it starts to become a little more exciting, and the benefits start to become a little more tangible. AI can become an agent, working on behalf of the user.
To achieve this, we provide system instructions and files to AI models, giving them the context and knowledge to emulate various personas, such as a case worker at a financial institution. We then provide the connectors to link these agents* into various systems, completing actions such as getting the weather, generating a custom report, or creating a financial journal entry.

* at this stage of AI development, we’re using the term ‘agent’ fairly loosely, giving instructions and context to a pre-trained model.

Step 3 - process-oriented, agentic, integrated AI

The titles here are getting longer, but this is quite a simple step up from 2. Here, we don’t need users to trigger or progress AI responses or even their system actions. Humans become an important part of the picture (e.g. for review, approval and refinement) but AI is the driver. So we’re really moving from a user perspective to a business perspective.
It’s not that big a step - all the hard yards were done in 1 (creating the model). Here, we just need a structure that treats the thread as more of a workflow, or context window. Then we can have event-based triggers that kick-off automated processes. These can have AI, human, and standard automation components. This needs more space to discuss, but trust me, **it’s simple.**

How we can get to step 3 - a process-oriented AI solution

Let’s just go through each of the fundamental puzzle pieces, and the full picture will form pretty fast.

Threads

We need to be able to create and maintain chat threads automatically. The thread serves as the working memory of the workflow, meaning we don’t need to give everything at once to the AI, it can pick up the pieces as the story progresses. So generally, what we’ll do is trigger thread creation off the first step in the process - some kind of event trigger like an enquiry being received.

Messages

Threads consist of messages, and we need to build automations that create these message records when a user creates a message, when a system does, and when the AI provides a response.

Runs

These are the bits where the AI is processing information, or ‘thinking’. We just need to execute these, really, and at the right points.

Prompts

Now we start taking the steps towards process automation. Because we don’t have people writing into the thread all the time to kick-off AI responses and actions, we want to have a library of the prompts we’ll need throughout the process.

Assistants

The ‘system instructions’ part of a prompt, essentially saved to a clipboard and repeated with each prompt. Think of it each time telling the general AI “When you are responding, you are responding in this capacity”. I like to think of it as telling the actor (the AI model) what role it’s playing in this scene. “Here, you are a diligent case worker who has a great knowledge of banking regulation in Australia. Here’s all the key files that’ll help you sell the role.”

Functions

These are also known as actions - basically structured payloads provided to integrated systems. More simply: connecting to systems. For instance, if I want to get the weather, I need to provide the location and if I want Celsius or Fahrenheit. So I’ll put that in when I ask for the weather, the AI will determine to use the “get_weather” function, parse a json format that has the data I gave it to the weather system, and will get back that it’s 30 degrees Celsius in Sydney.

Process hierarchies

Now we reach peak excitement. We all know business process structures, e.g. with level 1,2,3,4 etc. defining at each level of granularity how a business operates. To automate with a process-oriented AI, we just need to load those up, and link AI agents and prompts to specific steps in the process. There’s a beauty to this, because we can decide where we want to deploy AI and at what level. Maybe I don’t trust it yet - that’s fine, let’s put it at level 7 of a process and only for a few steps. But as that capability grows, as our trust grows, we can expand that out with the same tech.

Process executions

With the processes defined, we now just need to have them run. Based on various triggers, we can have humans and AI agents collaborate through to completion of processes. We have great traceability throughout as well, with every AI action, response and decision captured.

So what?

If we can automate processes, we can very easily determine ROI. Just think:

What’s the FTE commitment to maintain that process end-to-end right now?
What elements can we automate using a combination of AI and traditional means?
What are the costs to implement and maintain this?

This equation is quite simple for most businesses, because:

To be blunt, people work in the places we haven’t been able to automate. As such, a significant portion of employees within any organisation will be maintaining core processes.
We can automate a larger chunk of business processes than we could at any point in history, and by a significant leap.
The costs scale as the automation does, but with significantly lower prices than humans and a competitive market driving model prices down, this is not complicated.

Lastly, one point I’d like to add is that, unlike previous AI solutions, a process-oriented solution doesn’t rely on user adoption. It drives the process, and team members respond.