Applied AI for managers - clearly explained

Tip to try out

The company shows what AI can do with language fireflies.ai. Their software-as-a-service solution records video conferences, creates meeting summaries, extracts the next steps and makes the entire conversation searchable. Practical! How will the big providers react?

Back to the topic!

Influencers and influence

When a new technology captivates people, things get scary. In the best puffery, people, most of whom are not tech-savvy, take it upon themselves to spread exaggerated expectations and fantastic predictions about the blessings of this innovation. Social media are the best breeding ground for such exaggerations. Predictions are made, promises are made and lies are told until the beams bend. The more drastic the change is portrayed, the more "views" - or so the hope goes. Attention is a good way to earn money: Thanks to cost-per-clicks or cost-per-view. Many a pretty marketing face can also be brought into the company for a "consultation" in this way. It's hip, progressive - and it's also nice.

The many "influencers", "LinkedIn voices" and "Top 30 under 30" want to be fed.

Competent technicians see through such fireworks of superficial naivety. What counts for them is how the underlying technical "marvel" works, what means it uses, where it has its strengths and weaknesses, under what conditions it works well and badly - and what you can do with it.

Far-sighted entrepreneurs examine whether a new technology has substance or will turn out to be a shooting star, what operational and possibly fundamental effects it can have on their own business, what economic benefits it can develop and what adjustments to their own business model may be necessary. The motto is: What does me no good is worth nothing. What harms me makes me think.

The same problem arises with Generative Pre-Trained Transformer models (GPT), which are on everyone's lips today:

What can we actually do with it?

To answer this question, we must first understand what these models - which are synonymous with AI today - actually do and what they do not do. We will now focus on this.

Performance

What do GPT models and the platforms in which they are embedded actually do? This is important to understand!

GPT models generate ...

• with the help of an input, usually a question or instruction, also known as a "prompt",
• and its model based on vast amounts of previously trained data material - mainly texts from Wikipedia, Internet content or source code platforms such as GitHub, etc. -also known as a "model",
• Predictions about data derived from "input" and "experience" - mostly texts again - also known as "completion".

Short:

The AI platform understands the input and uses an algorithm pre-trained on similar tasks to generate a new output that appears like a human work.

But what is AI not? AI - and here specifically GPT - is not "creative". It also has no understanding of the content that it interprets and generates. Let's not even mention its own consciousness.

AI also has nothing to do with "intelligence" in the human sense. GPT models generate continuations of something on the basis of a pre-trained parameterization of the generating model. Stochastic processes are used, meaning that the result (e.g. the text produced) cannot be predicted from the outside and its creation cannot be traced. The GPT models are too complex.

Although GPT models are modeled on the human brain in their mathematical structure (logical "nodes" as a substitute for neuronal synapses) in order to imitate the graduated functioning of the human brain, they are otherwise far removed from anything human.

Benefit

We now know that GPT models generate an output from an input using a pre-trained algorithm. What can we use this technology for in a business context?

Analysis of large amounts of information

GPT models can summarize, evaluate, explain, put into other words or interpret large amounts of textual - but also visual and acoustic - information in response to a corresponding input ("prompt"). This is their immense strength: they turn data into something else - in this case shorter, differently formulated, segmented texts etc.. They are therefore used for sentiment analyses, summaries, extracting important points and reformulations or translations.

GPT models should not be confused with search engines such as Google, as their primary function is the creation (i.e. continuation) of something "new" on the basis of previously trained relationships of a vast amount of data material to each other (i.e. words, pixels, sounds to each other). For this reason, one and the same search query to a GPT model would always return (slightly) different results. The GPT model is somewhat "idiosyncratic" because it operates with probabilities for the continuation - and these are not always unambiguous.

The same applies to the generation of new information from a search query on the company's own data, such as a contract database with the prompt: "Summarize all contract passages for the customer Mayer GmbH that I have worked on in the last 3 years.". The summary will always look different in detail with the positive feature that the underlying facts are those of the company and therefore provide a stable database. The risk of hallucinations, i.e. the invention of "facts", is therefore significantly lower.

Remember:

GPT models are not a substitute for search engines

In the meantime, some providers have added GPT capabilities to their search engines. These search engines then receive a query (prompt) in order to generate an answer (completion) to the query using a GPT model and a huge amount of previously indexed data, which relates to relevant data that was found by searching for a match to the prompt. Here, the relevance of the data to which the transformation, i.e. the generation of "new" data, is applied is first checked in the background. GPT is applied to a search, but does not replace it.

Sales employees use GPT models to interpret video conferences. What exactly is a customer interested in? What documents did they request, what did they respond to and how? And much more. Since AI is not limited to text, the multimedia field of application should not be overlooked!

Generate source code, make developers more productive

With Copilot X for Visual Studio Code - analogous plug-ins are available from Oracle and others - a use case has become popular for creating software based on textually formulated requirements. The GPT model generates source code in the desired development language from precise requirements and is aware of the course of the requirements description so that it can add selective extensions or changes afterwards.

Furthermore, a GPT model can understand source code, explain it comprehensibly, document it, create test cases, recognize errors, make suggestions for improvement and transfer it from one programming language to another.

We expect these skills to be used in all software development platforms in the near future. The range of tasks and the impact of the software developer will broaden significantly. They will play a key role in requirements, design architectures, specify test cases, carry out large parts of the implementation using written or spoken language, have documentation created automatically and focus on end-to-end functionality. AI will relieve him of much of the development work that he may have grown fond of.

Is the revolution eating its children?

Recognize the signs of the times, is the motto. The revolution will not devour its children, but it will force them to rethink and act. The winners will be the creative and innovative developers who have ideas and want to improve the world. AI has given them a powerful set of tools. Perhaps this will also help to alleviate the shortage of skilled workers.

Create advertising texts and e-mail responses

Marketers use GPT models such as ChatGPT to create outbound email marketing campaigns, copywriting or customized communications.

In the future, GPT plug-ins will answer e-mails, take over customer interactions - and provide a lot of generalities. Because remember:

Human creativity is not a property of GPT models

There is nothing wrong with automatically reworded texts or texts translated into another language, but the clever spammer can generate entire platitudes with a few keywords. The AI does this without complaint.

Intelligent bots instead of annoying voice menus

What works for emails can also be used interactively on websites or in apps. Customer service bots or helpdesk bots are another area of application for GPT in companies. If the available information about the customer is used - e.g. from the CRM - the AI bot can perform relevant and specific tasks. It can also trigger processes - AI should always be used to automate manual work - so that, for example, missing invoices can be searched for and created in a form specified by the customer (e.g. as an X invoice). Furthermore, tickets can be submitted, information about the customer system can be automated via interactive agents and retrieved by the GPT model according to the customer's request, and then updated in the corresponding backend system via the GPT model according to the customer's request. Ultimately, the GPT model performs the relevant action to process the request semi-automatically.

This chain of prompt-response patterns allows extensive automation of business transactions that are controlled by humans but solved by AI bots. The days of simple voice menus are numbered.

Application in the company

GPT platforms can be used in a variety of ways. We would like to highlight two major areas of application that are relevant for every company. Furthermore, it is already foreseeable that AI will find its way into every software product. It is only a matter of time before the individual services described above will be included in most software suites on the market.

Automated document processing

Whether PDF, Excel, e-mails, Word documents, scans or photos: GPT-based solutions understand the content of these documents and can transfer them to other systems. Specific application examples in the company include

• Automatic order entry of heterogeneous PDFs in retail
• Transport order entry in logistics
• Processing and correct posting of incoming invoices in all sectors
• Generation of outgoing invoices in the format requested by the invoice recipient and breakdown
• Verification of certificates of analysis in the cosmetics and life science industry
• Evaluation of delivery bills and packing lists with subsequent posting in the warehouse management system
• Automatic policing of applications in the insurance industry
• Automatic verification of benefit claims from contracts
• Creation of doctor's letters, repair reports, call recordings and transcripts based on recorded conversations.

These are just a few examples. There are no limits to creativity.

What people used to do manually with documents can now be automated using AI.

Enterprise Information Search

What works with selected documents also works with all company information - regardless of where it is located and in what form it is available. In the same way as ChatGPT, existing GPT models such as OpenAI or Cohere can be applied to construction plans, spare parts, stock levels, CRM content, orders and Internet information. This does require a few preparatory steps. However, there are already platforms that enable specialized databases such as vector databases or graph databases to be loaded via interfaces so that the GPT models can then be applied to this proprietary company data.

Concrete example applications are:

• Provide field service staff with step-by-step repair instructions and assembly instructions for specific customer installations.
• Provide lawyers and tax advisors - as well as administrative staff - with concrete answers to a set of facts and have the pleading pre-formulated.
• Create sales offers for specific customers and requirements with all available know-how from the company in order to compile them in a way that is relevant and exhaustive for the customer.
• Search for information on specific problems in seconds instead of laboriously trawling through numerous programs, databases and portals for many hours.
• Interactively and specifically communicate troubleshooting tips to technicians and clerks before work begins and suggest or explain the necessary measures.

All of the company's knowledge and proprietary company data is not just available via verbal interaction. No, AI can search this data within seconds for the specific request and return the selected information or answer in a relevant form. This is where the strength of the generative element comes into its own.

The clerk interacts with their company information as they themselves see the problem - and has it partially automated.

AI helps employees to arrive at a solution step by step or develops this solution in collaboration with them. Laborious compilation of distributed information is a thing of the past.

Components

Technologies alone are not enough. They simply have to be made usable. The question arises: what should a company acquire in order to benefit from all these great AI capabilities?

Answer 1: Nothing for the time being!

You should not buy anything until you have fully understood the functional requirements and broken them down into their technical aspects. Some companies do offer "ready-made" solutions. However, their usefulness is controversial in practice because they usually only cover the technical requirements to a certain extent and the quality of the results is not sufficient for reliable use.

Numerous innovative IT companies from the early days are also faced with the super-rich giants such as Microsoft, Google or Oracle, who simply integrate their products into their own new mass products. Enterprise information search is one such topic, as is automated document processing. It is much more important to adapt these existing, new platforms to the company's situation than to introduce yet another new system from a provider.

Innovators who have not yet achieved a strong market position and introduced a very well-functioning product to many customers without the need for hundreds or thousands of consulting hours will be among the losers of 2023.

AI is not a technological niche but an integral part of all major solutions.

Answer 2: Gain competence

We observe that in the fourth quarter of 2023, companies such as Microsoft - and behind it OpenAI - will provide a powerful basic technology to realize the use cases described above and many more. With the exception of other US tech giants such as Google, Oracle, Meta, Amazon or IBM - and possibly their Chinese and Indian counterparts such as Alibaba - hardly anyone will be able to keep pace functionally and technically. The capital requirements and the necessary know-how are simply too great. This may change again if the open source community gains a foothold. However, this is not foreseeable at present.

Achieving a viable solution that meets expectations requires in-depth expertise in the use of various AI approaches. A superficial understanding of a technology and the use of consumer services such as ChatGPT is not enough. The "battle" will be decided in those problem cases that will occur with 100% probability - i.e. certainty. Consequently, technical expertise in the various AI technologies, models and platforms based on specific use cases is essential. With a little creativity, solutions can then also be found for the technology-related disadvantages of AI, resulting in very successful system launches.

Ultimately, it's not just about AI. Numerous details need to be clarified: Integration in company data, identification of exceptions and errors, embedding in existing company applications (e.g. SAP), addition of auxiliary services such as a vector database, logging and monitoring, security in handling the data, man-in-the-loop for corrections and much more.

Technical expertise is required here to ensure that generative AI does not suddenly become a disaster for a company because the results are incorrect or end up on the internet for everyone to see.

Answer 3: Think and develop modularly

In the cloud, it is worth thinking in terms of functional modules. Due to the accumulation of power among the Internet giants, unfavorable dependency relationships must always be assumed. No company should put all its eggs in one basket. Splitting up the specialist functions across several modules and several providers is a sensible risk mitigation.

Moreover, generative AI is still a relatively young technology. It has only been widely used for a year. Many new offerings are in the starting blocks or will only come onto the market in the next two years.

Consequently, every company should encapsulate its functionalities as well as possible and work towards technical uniformity and reproducibility of results across multiple platforms. In this way, you can benefit from useful innovations on the one hand, but also defend yourself against unfavorable price increases - a bit like switching electricity providers.

A targeted and moderate diversification of the technology stack makes sense from a risk perspective and should be carefully planned. In addition, the creation of the solution should be automated ("loading" with data, etc.) so that it can be transferred to multiple platforms.

Don't put all your eggs in one basket

Influence on earnings and benefits

We now know that the modern, powerful AI platforms are in the hands of a few large US corporations. We also know that there are strategies to at least limit this dependency. And it is to be expected that the open source community will bring even more welcome plurality to this concentration of technology. On Hugging Face you can get an impression of how wide the variety of AI models already is today. So much for the economics.

We return to the technical side: Here we know that GPT models generate an output from an input using a pre-trained algorithm. What does the product of this generation process depend on - and therefore the quality and usefulness of AI?

The following dependencies should be mentioned:

1. The model itself
2. The training data
3. The use of the model
4. The improvement measures

Model selection

There are both proprietary large GenAI models available for use as a cloud service (e.g. OpenAI GPT 3.5 Turbo, GPT 4, Cohere) or open source models that can be used in your own data center or in your company's own private cloud (e.g. Meta Llama 2). There are also a large number of specialized models from various smaller providers, but these are not necessarily better than the large language models.

When selecting a model, the main technical parameters to be checked are. These include

• The number of "nodes" contained in the model (i.e. "parameters", similar to synapses in the brain; the more, the better)
• The number of possible "words" (i.e. "tokens") that the prompt and completion can consist of - important if larger files are to run through an AI model -
• Or the ability to integrate the AI model into a company process (interface features, plug-in capability, etc.).

Public benchmarks for the models can provide orientation. Here the Center for Research on Foundation Models at Standford University. In many cases, a company will not be able to avoid testing the crucial and success-critical point of its use case with the desired data model. This can be done with little effort, but results in a reliable basis for decision-making and an assessment of the quality of the results to be achieved.

Trial makes perfect.

Please note: AI is based on stochastic processes. There is no 100% certainty with regard to the quality of the results. It is therefore important to keep errors to a minimum by selecting models wisely and to be able to control them through clear identification.

Training data quality

In order to make a GPT model usable in the company, relevant training data must be made available. This can be PDF documents, emails, scans, websites and much more. The decisive factor is the relevance of their content to the problem to be solved. If, for example, PDF orders are to be automatically posted to an ERP system, then a representative, heterogeneous selection of these should be presented in order to optimize the AI model accordingly.

Only when the basic quality of the results of an AI model is right should further measures be taken. This applies all the more when classic machine learning (ML) is used. Here, labeling takes place first. The training result is only useful if the training data sample has been well selected. This can then be applied to further, new examples in order to assess the accuracy ("verification").

In GenAI procedures, the training data determines the basic calibration of the model. What exactly does the prompt look like? What additional information is passed to the AI model so that it interprets the training data correctly? Which parameters are set to avoid the usual GenAI problems such as "hallucination"? This is only possible with good training data and its accurate assessment.

Model utilization

Generative AI processes have led to a number of best practices. These in turn result in higher accuracy and therefore higher quality results.

In addition to prompt engineering, which involves the creation of intelligent and effective analysis and creation prompts, two other processes are used: so-called "shots".

By "zero shot", "one shot" and "few shots", we mean procedures in which the correct answer ("zero shot"), an example with a similar data set ("one shot") or several examples ("few shots") are transferred together with the prompt with the aim of the AI model learning from this for the future. This learning takes place in the company's own AI platform account and is not available to the general public. As a result, proprietary data remains proprietary.

"Fine-tuning" refers to the separate transfer of prompt-completion pairs with the same expectation: The AI model should learn from this how to solve a task.

Both methods improve the quality of results. This should be measured and evaluated in the run-up to a broad application of GenAI in the corporate context. The sensitivities should also be evaluated: What effect does it have if I use one method in one way or the other?

Optimization of results

If a GenAI use case has been put into production, errors will naturally occur. These should already have been intercepted automatically before commissioning, so that an error log or a clarification case collection can be used to improve the AI setup. It may also be necessary to combine several methods or models in order to achieve the desired degree of automation.

How concentrated the error cases are is relevant here. If a small error rate of e.g. 2% extends over 70% of the cases, this can quickly make the approach unusable.

Results can be optimized if the aforementioned points are taken into account. They lay the foundation for AI automation.

The usability of AI in the corporate environment depends largely on these four points. It is worth taking them seriously in order to get the most out of the project. 98% accuracy can be achieved in this way, leading to an immense increase in productivity and a reduction in costs.

Data security

As the most powerful and high-performance AI platforms are offered exclusively as a service from the cloud (software-as-a-service), company data always flows into the cloud. This means that it is no longer under the control of the company.

Some AI services can be used in the local data center or in the private cloud (e.g. Microsoft Document Intelligence), but this should not be a substitute for data protection.

For this reason, the leading AI platform manufacturers state that they do not reuse their corporate customers' data for the training of their AI models. According to them, the data is processed in compliance with the EU GDPR and is only accessible to the companies using it themselves.

If a company uses very sensitive data, additional protection mechanisms must be integrated. This ranges from the anonymization of personal data to the precise differentiation of which data is sent to which provider in which environment and in what form.

Before an AI model is used on a broad scale, the data protection conditions of use and operational protective measures must be examined. A certain residual risk will remain, as is currently inherent in all cloud services. However, we ourselves take this issue very seriously and examine it on a case-by-case basis.

AI in the company

How should an AI project be set up in a company? There is a simple methodology here, which appears difficult to follow in practice. It therefore requires good leadership and consistency in project management:

1. Define the task and the problem exhaustively. No empty phrases, no superficial dreams. Clear operational, concrete descriptions.
2. Check whether and, if so, which AI technology(ies) can handle the task and how. Does AI fit the problem? How exactly does AI solve the problem? What is the crucial point that determines success or failure?
3. Try out the jumping off point. Money walks, BS talks. What you haven't seen is a case for the church. There you can believe, with AI you have to validate success by doing when in doubt.
4. Plan for the exceptional cases. No AI solution delivers a 100% result. Errors are inherent in the system. How do you recognize them and how can you remove them for evaluation?
5. Prepare yourself emotionally for difficulties. Rely on your problem-solving skills or get external support. The technology is young, the complexity high, the experience still limited. However, the opportunities are enormous - if you penetrate the situation intellectually.

With this in mind:
Use your human intelligence to get the most out of artificial intelligence for your company!

About Business Automatica GmbH:

Business Automatica reduces process costs by automating manual activities, increases the quality of data exchange in complex system architectures and connects on-premise systems with modern cloud and SaaS architectures. Applied artificial intelligence in the company is an integral part of this. Business Automatica also offers automation solutions from the cloud that are geared towards cyber security.