The technology for AI bots is here

At this point, we don't need to recite the usual “findings” that AI is great, will change our world and that ChatGPT triggered a GenAI hype almost two years ago. What we can say, however, is that these developments have actually made the full potential of - let's say broadly applicable AI - known to a large audience. In the simplest case, AI bots is the most suitable way for actually implementing this potential in companies. An AI bot is an automated program that uses artificial intelligence to complete tasks independently. Unlike simple bots, an AI bot can hold complex conversations, analyze unstructured data and improve its capabilities through machine learning. In other words, it can offer really useful automation. And best of all, the technology is available, robust and can be used via established service providers:

Essential components of an AI bot:

·      Large language models (LLM) – such as GPT-4, which are available to everyone today and are constantly improving

·      Databases – like established cloud providers that can store and manage data in a secure and organized manner

·      Interfaces to other applications – such as ERPs or CRMs

The basic architecture of AI bots is simple

These components can be easily assembled into a bot. Basically, I first need an interface to an employee, for example a chat window, and then a program that I have developed specifically for a company and this particular use case (I actually have to put a little brainpower into this, but more and more providers are closing this gap with simple no-code or low-code interfaces). This code organizes the interfaces to the LLM (breathes life into the bot) and to my database (the bot's memory). My AI bot is ready. Of course, I can make the right choice for each component. Which LLM capabilities do I need most? How securely does my data need to be stored - and who needs access to it and from where? The architecture also changes if the bot is integrated into processes, for example interfaces to an ERP system need to be established.

From internal assistant to virtual consultant

The promise of AI bots is usually increased efficiency. Leading management consultancies have confirmed that the potential is real. However, the actual increase in productivity depends heavily on the use case. In addition, employee satisfaction is notirrelevant: Everyone is happy to be relieved of annoying processes. But what are the use cases then? Let's take three examples of bots:  

·      Banal: Expense management

·      Not bad: Invoice check for insurance companies

·      Complex: Virtual consultant

Expense Management

Let's say I want to improve the submission, management and accounting of my sales staff's receipts. So I build a bot that 1) uses an intuitive interface, for example a common messenger, so that my employees don't have to install the five hundredth app on their workphone and get used to it; that 2) uses the understanding and also the image recognition capabilities of GPT-4 to read receipts in the best possible way; and that 3) stores the receipts together with other necessary information in the desired format in a table. We recognize the three essential components from above.

You could imagine it like this: Salesman Hans invites customers to a business dinner, takes a picture of the invoice at the end (using the familiar functions of the messenger service) and confirms at the touch of a button what the bot has already stored as additional info, who has paid, who he is traveling with according to the Outlook entry, etc. The image of the bill and the information is then stored and checked to see whether it is valid and whether the price of the meal was reasonable. At the end of the month, there is an automated summary and the data is prepared for the finance crew. Tada!

Invoice check for insurance companies

Let's imagine that I have to check invoices with attached photo documentation (boring!) that tradesmen send me as an insurer. I want to check whether the work corresponds to the service awarded and whether the price is within budget. This calls for an AI bot!

Basically, I'm sticking with my current architecture, but now the bot has the difficult task of understanding photos and evaluating them based on the company's own policies. Firstly, I have to give the bot a lot of knowledge, enable the interface to the employees as a human-in-the-loop and design it intuitively, and let the bot learn, i.e.“fine-tune” the underlying model. This trains the bot to pay attention to the important details. Implementation should be viewed more as a process, with experienced employees “raising” the bot and also monitoring it more closely at the beginning. Once the bot has been set up and adjusted correctly, however, it can quickly check the majority of invoices independently and only send outliers and invoices that it is unsure about to the employees. Much better!

The virtual consultant

We are actually talking about a classic AI use case here: knowledge management. First of all: a bot will not seriously take on creative solution thinking for the time being. But as a sparring partner, analyst and data manager, it is a tool with enormous potential for many service companies.

Two examples: Consulting companies in particular thrive on internal knowledge. Every problem has somehow been dealt with before, and the relevant knowledge is stored somewhere. This is where a bot can close the gap by quickly and intuitively accessing the right knowledge, for example when a consultant asks: “Have we ever done anything on blockchain in the industrial sector, I'm particularly interested in use cases with EBITDA potential”. The bot can quickly find and compile the relevant information from large amounts of data. For the moment, it should only be noted that the database must of course be prepared for this – the data must be “vectorized” so that the bot can quickly find the right target spaces. This preparation is nottrivial and must be considered along with the overall bot solution.

A second example: Analyst work. I can set my bot to scan predefined, available sources constantly and automatically. The bot extracts key data points and saves them in vectorized form.

If I search for certain information at a later date, the bot can prepare this information quickly and intuitively. For example, the bot can compile entire reports, including graphics in the company's own design.

There are no limits here, depending on the complexity and investment.

Companies need to get started now

AI bots are technically available and robust, have clear value propositions in terms of productivity and can be quickly adapted for other use cases. Smaller companies must now also start to put processes that were previously considered non-automatable to the test. The important thing to remember is that it often doesn't require a huge investment or a major transformation - automation can be implemented in small steps and with a quick return.

It is worth picking out the most obvious process, piloting a bot and automating further processes from there. At romtec.ai, we use our AI platform LANDO for internal processes such as CRM and expense management, but also for our important operational work.

There is no longer a flipchart that is not briefly photographed and automatically uploaded to the cloud. The AI understands the content and can also reproduce individual details for me when I ask: “What did we outline again last month on the subject of ‘bot architecture’?” LANDO: “A bot consists of a front end, LLM and databases.Here's the photo of the flipchart.” And that's just the beginning. The virtual consultants are not that far away after all, if you consider that AI can also learn based on the company's own stored data. Either way, the following already applies today: companies need to exploit the potential of GenAI, and bots are the easiest way to do this.