Garbage in, garbage out? How to improve the output quality of Corporate GPTs

Garbage in, garbage out – this problem also exists in the world of generative AI. If we feed inadequate data into our corporate GPT, we will end up with inadequate information. What can we do about it?

A guest article by CEO & Founder Michael Siegmund and CMO Yagmur Anis, qibri GmbH.

What does GIGO actually mean?

Let’s start with a quick definition of GIGO:

GIGO – Garbage In, Garbage Out – describes the principle that the results of a computer can only be as good as the quality of the data it contains beforehand. If you input incorrect or inaccurate data, the computer is highly likely to produce erroneous or useless outputs.

Thus, GIGO is a particularly hot topic in the age of Corporate GPTs: how can we store our company data in such a way that we can optimally use it at a later time?

Data quality becomes an essential part of every AI strategy.

The importance of Corporate GPTs

ChatGPT has drastically changed the way we interact with information – how we search, create, and modify it. This disruption is directly reflected in the typical CEO agendas: the introduction of artificial intelligence is now high on the priority list.

More and more employees are using ChatGPT to get support in performing various tasks. It’s likely that sensitive questions and information are also being entered into the tool.

At the same time, few companies have successfully implemented Corporate GPTs. Many are still struggling with traditional knowledge and data management problems, such as different repositories, duplicate and outdated documents, and a lack of version control.

As a result, open systems like ChatGPT remain the preferred resource for employees, which, however, poses significant risks to data confidentiality within companies.

So, how can companies create successful and secure Corporate GPTs using their own data and minimize these risks?

Key data quality standards

Implementing data quality standards as a framework can give companies a competitive advantage. After all, the quality of the data is the foundation for reliable and high-quality AI applications.

To assess data quality, a variety of dimensions can be applied. Some of these are fundamental, and we will go into them in more detail here.

Accuracy

Many companies use platforms like Box, SharePoint, or Notion as primary sources for their knowledge management. However, these platforms are often used for everyday work. This means that documents in draft status and without review are often included. Creating content that is intended to be shared across the company, on the other hand, requires multiple revisions and adjustments.

Validity

Things change quickly in business. New trends emerge, and new regulations are constantly introduced. It’s practically impossible to track all changes to existing documents.

However, if old records are kept in the repository for AI initiatives, employees may be presented with information that is no longer valid.

Consistency

We’ve all encountered multiple versions of the same document and had to ask colleagues for help in identifying the correct version.

Additionally, it becomes difficult to trace the origin of the information and verify it with the relevant stakeholders.

Completeness

When different teams maintain different repositories and rely on different systems without a unified source of truth, the Corporate GPT is trained on multiple incomplete elements and produces answers missing specific insights.

Improving the output quality of Corporate GPTs: 5 strategies

Fortunately, we do have some levers available!

Here are 5 strategies that can help companies improve the output quality of Corporate GPTs:

1. Use a framework to reduce hallucinations

While generative AI generally provides impressive responses to user queries, it’s not uncommon for it to generate incorrect answers and present them as facts (so-called AI hallucinations).

To avoid this issue, we can use a framework like RAG (retrieval-augmented generation). This method performs several checks to ensure that the information output is actually valid.

2. Separate past and future-oriented documents

Documents can be separated based on their temporal orientation and purpose. Daily work files can be excluded, thus avoiding inconsistencies.

For example, a Corporate GPT might need to provide two types of information to employees:

• Guidelines on how things are done today: For this, we can create a source of current or future-oriented documents that provide specific instructions in the form of guidelines, SOPs, or best practices.
• Information on how things were done in the past: We could create a base of completed records that describe past efforts in detail.

3. Utilize back-office functions

This means drafts are created and edited in a separate area, and only finalized and approved documents are available in the frontend. This way, multiple stakeholders can collaborate and contribute to the documents without the risk of errors or inconsistencies being visible to the Corporate GPT or the end user.

4. Assign responsibilities

To ensure transparency in the responses from Corporate GPTs, we should create a clear information base that assigns responsibilities for each information unit.

This can help overcome the challenges posed by the black-box nature of AI technology: thus, information can be traced back to its original source.

5. Break down silos

When departments work in isolation, using different formats, methods, and software for creating and organizing information, it becomes increasingly difficult to create a comprehensive and cross-sectional overview.

One way to overcome these silos is to implement a platform that facilitates teamwork and aligns various departments toward a common goal.

This approach ensures efficient collaboration between different organizational units while providing a single source of truth for a corporate GPT.

Summary and key takeaways

GIGO is a hot topic when we talk about Corporate GPTs. Only with well-organized data of sufficient quality can Corporate GPTs truly provide support. If we feed poor data into our corporate GPT, the results can be misleading and detrimental to business success. Accuracy, validity, consistency, and completeness are key criteria for good data quality.

These 5 strategies help improve the output quality of a Corporate GPT:

1. Use a framework to reduce hallucinations: A RAG framework can reduce incorrect AI answers by binding models to a knowledge base.
2. Separate past and future-oriented documents: Separating documents based on their temporal orientation can increase consistency.
3. Utilize back-office functions: Editing documents separately ensures no errors and secures high-quality final outputs.
4. Assign responsibilities: Assigning responsibilities for each piece of information ensures transparency.
5. Break down silos: A common platform fosters collaboration and prevents isolated work.

About the authors

Michael Siegmund is the CEO & founder of qibri. Michael has 25 years of experience in strategic corporate development, including 7 years as a corporate consultant (e.g., Droege & Accenture), 7 years as an executive in an oil & gas company, and 12 years as a 3-time entrepreneur.

Yagmur Anis is the Chief Marketing Officer of qibri. Yagmur has 12 years of experience in B2B marketing with a focus on enterprise software and data analysis, including 7 years at McKinsey & Company.

About qibri

qibri captures and organizes existing know-how for companies, ensuring its continuous improvement and optimal use in daily operations. Additionally, qibri serves as a curated and secured knowledge base for future internal LLMs (Corporate GPTs), providing employees with reliable answers that can be further developed by experts. This helps companies reduce organizational complexity, achieve operational excellence, enhance productivity, ease transformation, and create the foundation for reliable generative AI.