Over the last three months, I have spoken with over 50 CIOs, CHROs, and CSOs, all of whom have been inundated by business users in Global 500 companies demanding ChatGPT functionality.

The reality of the enterprise

Enterprise technology leaders are debating the pros and cons. Some, like JP Morgan, have announced an outright ban, and others are still finding a middle ground. The carefully calibrated security architecture, burnished by years of processes, tweaks, and safeguards, is now in utter disarray with the onset of AI. They are all aware that the road to integrating GPT into the enterprise is far from simple.

The challenges we see

At Avaamo, we understand the need for reliability and security, and so we analyzed what could go wrong and how to enable enterprises to unlock the full potential of large language models like GPT-4. We broke this down into the following challenges:

1. Hallucinations:

Large language models (LLMs) are powerful prediction engines but accuracy? Hmm, not always…hallucinations are not a “choice” in the enterprise. 

2. Hate and Violence:

The potential for model bias is high, and therefore the fear of outputs spouting hate speech is high as well. Enterprises need a foolproof system in place to filter or moderate responses.

3. Profanity:

The potential for LLMs to generate profanity in response to irate or frustrated users is real. It’s crucial to address this issue and minimize the risks associated with inappropriate content.

Elicited from ChatGPT, February 2023

4. Enterprise context and content:

Enterprises have secrets and it needs to be protected and controlled. Most LLMs are trained on public data, they still fall short of addressing role-specific context, access and controls.

5. Sensitive data:

A real and present danger for enterprises is customers and employees inadvertently providing PII and sensitive information as part of their queries. In fact this just happened with employees using ChatGPT at Samsung.

6. Identity and Access control:

The notion of an “identity” of the user is extremely critical in an enterprise. The user’s identity defines not only what information and business applications the user has access to but also what specific set of actions the user can perform. This concept, which is the bedrock of enterprise architecture, is missing in LLMs.

7. Compliance and SLA’s:

Enterprises typically require software vendors to comply with (SOC2, SOC3, HIPAA, FINRA, GDPR), SLAs around uptime, response time, and resolution times. Unfortunately, LLMs currently do not offer the capabilities needed to meet these SLAs.

The Avaamo Solution

Avaamo took a hard look at these challenges and is proud to announce the launch of of our first initiative under AvaamoGPT;  We are designing and building a protective software layer that addresses these challenges above to access GPT technology safely and securely. Enterprises can can now deploy conversational AI solutions using LLMs to generate highly accurate and responses to complex language-based queries without worrying about the risk of data breaches or unauthorized access. In addition to AvaamoGPT security features, it integrates seamlessly with existing systems and workflows. This product excites us not only because of how immediately useful it could be but also because we believe this is the most practical and safest path to deploying LLMs in the enterprise to enhance the Employee and Customer Experience.