![Knowlegde](/sites/default/files/2019-10/2_0.png)
Bringing AI to B2B
As most of us within the tech services industry, we’ve been looking into how we can leverage AI for our existing and new clients.
Before diving into the opportunities, let’s make a short summary of the state of things.
AI, or as we prefer to call it, LLMs are primarily focused on the following areas: text, images and in a later stage video & sound. Albeit, at the time of writing this article, video & sound are not yet fully public.
When we talk about AI, we prefer to reference to them as LLMs because, in current AI jargon, we actually have other areas as well. In the past decade, AI has been referred to machine learning, data mining, neural networks, and even some areas of business intelligence. LLMs, are just the newest additions to the field and who knows, what newer technologies will emerge in the near future.
For the purposes of this article, when we talk about AI, we actually refer to LLMs.
So how can LLMs can help your business?
We’ve identified the following main areas.
Applications
Text Processing
LLMs have been proved successful in doing text processing, either in the form of text summarisation, data extraction and data processing. All of these “activities” or “tasks” for the AI, aka the LLM, can be output via various channels. The most obvious one, is in the form of a chatbot, but other channels can be also implemented as well: form-based webpages or UI applications.
Companies that have large set of proprietary data such as prospects, internal documents, product descriptions can use LLMs in the following ways:
Text Summarisation
Self-explanatory however to pin it out, we refer to this task as the ability to compress various texts (and maybe images in the future) in fewer words that highlight key points.
Data Extraction
You can use LLMs to extract data directly into CSV, JSON or Excel format from various data sources such as CVs, prospects, web pages, etc. Basically, all the tasks that in the past were done by web crawlers, UI robots, Python scripts, etc, can now be replaced by tasks done via LLMs.
Data Processing
I refer to data processing to the ability to “comprehend” a text and provide a novelty answer that was not (necessarily) verbose in the input data (“text”). Here, the main difference in the quality is given by the LLM used. The larger the model is, the smarter it is in form of data comprehension. I will explain in a different post, how these models differentiate.
Data Validation
You can incorporate LLMs into any existing (or new) data validation. So, instead of having just the classic validators: is it numeric? Is it a phone number? Is it SSN? You can now use LLMs to inquire about the content of a specific input (eg. Does this text describes an accident? - if your application is processing claims, or eg. Does the text provided include information that could be symptoms of a disease? - if your application is a healt-related one).
All of the above tasks can be easily incorporated into any business applications or you can choose to open a new delivery channel create a chatbot that will deliver this information to your target stakeholders (clients, staff, etc).
Another area where LLMs might prove useful is Text Generation.
Text Generation
Probably the most famous feature of current LLMs was the ability to generate text, or better say, content. This is made possible by the large data set that was feed into the model which based on an “prompt” is able to create answers.
Some applications have already made their way into business realm.
Code Assistants
Tech companies (and developers on individual basis) have embraced LLMs to help them generate code templates, find bugs or improve existing code with great results. This application is definitely here to stay and will likely be mandatory be part of any code development stack and most of the platforms out there (open source or proprietary).
Copywriting
Valid cases and “nefarious” actors are using LLMs to generate content on various topics. Information and desinformation are on the same par, the output of these LLMs. On the business side, LLMs are starting to be used to generate ads, documents, reports, etc. If your business application if full of text fields that require text inputs, mostly likely in the future, there will be a button that will assist you in filling the information in that field.
Pros & Cons
I want to summarise also the pros & cons of delivering LLMs within your business.
Costs
At the moment, we think cost is the most important factor to consider. While the “currency” of an LLM input is rather cheap right now (price per token), when we talk about business applications used by businesses, the costs can easily compound to thousands of dollars when several people do queries in the LLM. Another cost to consider, is the costs for the consultancy services associated with the implementation of LLMs. At the moment, the pool of people that comprehend and cand deliver results with LLMs applications is rather small and hence, expensive.
Private LLMs vs Public LLMs
Not many organizations are willing to allow their proprietary data to be fed into LLMs, not even as input. While most of the LLMs providers do provide assurance that they do not use this data as input to further reinforce their models, one cannot be sure of that due to (obvious) opacity of the code behind running the LLMs. As such, more business will be willing to deploy their own private LLMs running on their own hardware. This will further increase the cost deploying LLMs for your application.
Hallucinations
It is very well documented that, LLMs have hallucinations. When we talk about hallucinations within LLMs, we refer to their “disability” of generating false/fake reference. The fact that they voice their results with a high degree of confidence or without proper references to back the information. This is mostly due to the fact that LLMs are basically compounding numbers and any small “error” that occurs at one stage of the generation process, is multiple by the next steps within the model. However, there are ways to limit the impact of hallucination.
We hope this article provides with enough information to help you generate ideas how LLMs could be useful for your organization.
At Softescu we have already some AI implementations into the B2B arena.
Curious about LLMs can help your organization? Give us a shout out using the following contact form.