Go further, faster:
Many of us thought AI was just going to be something you could ask any question of, and get the right answer in return. It turns out that many of our first interactions with an AI-based system were with one that generates content, instead: ChatGPT.
Now we can use that same tool to generate the code that we use to consume it, and to extend it. Now we can write automations that ask ChatGPT (or a similar model) for code they use to extend themselves.
This has had the effect of accelerating application development to the point that even application developers are wondering if they will be put out of work by it in the long run (we think yes, they will).
Low-code and no-code solutions have proliferated, allowing for rapid prototyping and getting ideas to market in days or weeks, not months and years.
Why not take advantage of that to help your business accelerate towards a specific application that does exactly what you need to serve your customers just the way you want to?
Get more from your data:
Modern businesses generate vast volumes of data on a daily basis. Businesses that have existed since prior to computers are often trailing literally tons of paper content, too. There’s too much for humans to process.
With the knife of robotic process automation and fork of machine learning, this workload is now more consumable than ever. There is another ton of knowledge locked within that data, and with AI we can help unlock it.
Even basic summarization of content can make it more accessible. Exposing it to search without having to manually tag the documents can unlock massive insights, without having to anticipate in advance the fields we want to search for.
If all your company has to do is expose your existing content to the right machine learning model (s) to unlock question-and-answer or semantic search, why not investigate the opportunity further?
Make Decisions in Real Time:
A good place to look at this capability is snap decisions users, often contact center-agents or customer-facing employees, have to make about flexible policies during the course of their work.
For example, consider a contact center agent on the phone with a customer pleading for a refund of their third overdraft fee this quarter, or an insurance assessor in the field with a live customer begging them to approve a claim, or an underwriter with a customer in front of them asking them to approve a loan.
Are these the types of decisions that employees in your business are allowed to make? If not, how long does it take them to find someone who can make it today?
A properly trained AI model is able to make these decisions in an unbiased, objective manner. This lets the employee receive a decision that’s guaranteed to follow the business policies and reduce shrink due to employee non-adherence. The employee is able to assert that “the system made the decision”, so the customer has little to argue with.
Content Generation:
This is where some people start to have some cause for concern, but depending on the need having an AI generate the content may not be such a bad thing. From a marketing standpoint, a properly trained generative model can be providential, delivering a steady stream of social-media ready post content, and a wide array of things like slicks, blog articles, etc. Well-trained generative models will easily speak in the voice, tone and persona your business wants to project. From our standpoint, the intellectual property is the ability to engineer excellent prompts that result in the desired outcome the first time (and a reproducible output) instead of hours of frustration trying to get the robot to create the output right (this is even more true in code generation exercises).
If the business doesn’t want the entirety of the content to be machine-generated, then it can be tuned to generate outlines or skeletons of posts that marketing analysts fill in from their own creative minds. Generative models can also speed ideation, for example “please generate a list of all the ways to say yes or affirm something in English”.
Generative AI can also be used to create learning materials, like lists of questions that a typical person might ask about a particular piece of content, or a learning plan to educate a person about a topic. These can arguably be used to fuel additional research into the topics as well.
Pattern and Anomaly Detection:
This is often known as clustering, or conflated with, depending on your viewpoint.
Use Cases for Large Language Models:
It might not yet be clear, but there are wide varieties of types of “Artificial Intelligence” and an array of learning and retrieval models for each. One of these that is recently proliferating is known as a Large Language model. These models are trained on sometimes billions of instructions (like ChatGPT was) and can be trained on general knowledge or to perform specific tasks.
On the one hand, it’s easy to dismiss all Machine Learning as chatbots, and position the idea that a chatbot is not the right interface to solve every business problem. That’s true. However, not every Machine Learning/AI/AW solution demands a chatbot interface. Sometimes the output of the model could be a report or a series of simple alert if certain events are seen within an environment.
A good Autonomous Cognitive Entity generates the user interface on the fly to suit the use case required. For example, if an ACE interacting with a user via text (we would call it a chatbot) decides it needs complex entry from the user, it can generate the form on the fly based on the type of data required and present the form instead. A method like this reduces the likelihood of error in the form submission, it actually facilitates the robot validating the form input much more precisely than it could a natural language input.
There are many use cases for LLM’s within business though, none of which necessarily involve presentation via a chatbot-style interface, and virtually none of which require presentation of an interface that’s available to customers or the public at large:
Sometimes the data in the document (considered unstructured data) needs to be in a database or a format more usable to legacy computer applications. Using the right technologies, we can avoid the tedious steps of creating document templates and naming specific fields in OCR applications. Large Language Models are able to understand the content in a, well, natural way, and answer questions about the content that allow for properly formed data sets to end up in the target datastores.