Ground Truths for a company using Generative AI.

Most SMEs and some of the Medium sized businesses I talk to are struggling to grasp the enormous potential for Generative AI to get rid of corporate misery, the shitty parts of our jobs, that if weren't doing, can free us up to do other things that are more productive. The ones who do use these tools though could really benefit from something, I'm not even sure has a good name yet..

"Corporate Ground Truths?"

As businesses increasingly adopt large language models (LLMs) for various applications, establishing a set of "ground truths" about the company and its operations can significantly enhance the value and accuracy of the output Generative AI tools produce. This foundation of factual information could serve as a critical anchor point or input, ensuring that LLM outputs align with the organisation's reality. Not all tasks require this, but think of them as custom instructions and information that can help tune the output to be more aligned.

Key Benefits:

1. Consistency: Ground truths provide a unified base of knowledge, promoting consistent responses across different LLM applications, by different people, that could even include tone of voice adjustments.
2. Accuracy: By constraining the LLM with verified facts, the risk of generating incorrect or outdated information is reduced.
3. Customisation: Company-specific ground truths allow for more tailored and relevant AI interactions and interrogation.

Practical Implications / Application

1. Knowledge Base Creation: Develop a comprehensive, regularly updated database of company facts, policies, and procedures. Make them available to staff, especially new starters. This could be simple to copy and paste at first.
2. LLM Fine-tuning: Use the ground truths to fine-tune LLMs for company-specific tasks, improving performance on internal queries. This is more complicated and depends on your setup.
3. Output Validation: Implement systems to cross-check LLM outputs against the established ground truths, flagging potential discrepancies. "How does this support our xyz initiative or how does this go against our code of practice? Is this accurate? "
4. Employee Training: Educate staff on the importance of ground truths and how to effectively use LLMs in conjunction with this information.
5. Iterative Improvement: Continuously refine the ground truth database based on feedback and changing business dynamics.

For a manufacturing business for example, ground truths would typically cover various aspects of the company's operations, structure, and processes.

1. Organisational structure and key personnel
2. Product lines and specifications
3. Manufacturing processes and equipment.
4. Quality control standards and procedures
5. Supply chain details and key suppliers
6. Inventory management systems and where their data is
7. Safety protocols and compliance requirements
8. Production capacity and lead times
9. Pricing structures and policies
10. Customer base and market segments
11. Distribution channels and logistics
12. Environmental and sustainability practices
13. Company history and milestones
14. Intellectual property and patents
15. Employee policies and benefits

The list is pretty endless and I guess context windows are limited...

so many just limiting it to a few or having a repository to use take the ground truths from would be useful.

• Manufacturing processes and equipment
• Quality control standards and procedures
• Health and Safety / Compliance and Regulatory Procedures
• Product lines and specifications

is the way to go, these directly impact product quality, efficiency, and the ability to meet customer demands. Covering the most operationally critical aspects of the business. As the company becomes more comfortable with using ground truths, they can gradually expand to include other areas from the list.

so the TLDR is: Grounding LLMs in company-specific truths, businesses can harness the power of Generative AI while maintaining accuracy and relevance to their unique operational context. These will become a byword soon with chief AI officers.