Enterprise controls and generative AI for business users in real time.
Aible
With solutions for customer acquisition, churn avoidance, demand prediction, preventive maintenance, and more, Aible is a pioneer in producing business impact from AI in less than 30 days. Teams can use AI to extract company value from raw enterprise data. Previously using BigQuery’s serverless architecture to save analytics costs, Aible is now working with Google Cloud to provide users the confidence and security to create, train, and implement generative AI models on their own data.
The following important factors have surfaced as market awareness of generative AI’s potential grows:
Enabling enterprise-grade control
Businesses want to utilize their corporate data to allow new AI experiences, but they also want to make sure they have control over their data to prevent unintentional usage of it to train AI models.
Reducing and preventing hallucinations
The possibility that models may produce illogical or non-factual information is another particular danger associated with general artificial intelligence.
Empowering business users
Enabling and empowering business people to utilize gen AI models with the least amount of hassle is one of the most beneficial use cases, even if gen AI supports many enterprise use cases.
Scaling use cases for gen AI
Businesses need a method for gathering and implementing their most promising use cases at scale, as well as for establishing standardized best practices and controls.
Regarding data privacy, policy, and regulatory compliance, the majority of enterprises have a low risk tolerance. However, given its potential to drive change, they do not see postponing the deployment of Gen AI as a feasible solution to market and competitive challenges. As a consequence, Aible sought an AI strategy that would protect client data while enabling a broad range of corporate users to swiftly adapt to a fast changing environment.
In order to provide clients complete control over how their data is used and accessed while creating, training, or optimizing AI models, Aible chose to utilize Vertex AI, Google Cloud’s AI platform.
Enabling enterprise-grade controls
Because of Google Cloud’s design methodology, users don’t need to take any more steps to ensure that their data is safe from day one. Google Cloud tenant projects immediately benefit from security and privacy thanks to Google AI products and services. For example, protected customer data in Cloud Storage may be accessed and used by Vertex AI Agent Builder, Enterprise Search, and Conversation AI. Customer-managed encryption keys (CMEK) can be used to further safeguard this data.
With Aible‘s Infrastructure as Code methodology, you can quickly incorporate all of Google Cloud’s advantages into your own applications. Whether you choose open models like LLama or Gemma, third-party models like Anthropic and Cohere, or Google gen AI models like Gemini, the whole experience is fully protected in the Vertex AI Model Garden.
In order to create a system that may activate third-party gen AI models without disclosing private data outside of Google Cloud, Aible additionally collaborated with its client advisory council, which consists of Fortune 100 organizations. Aible merely transmits high-level statistics on clusters which may be hidden if necessary instead of raw data to an external model. For instance, rather of transmitting raw sales data, it may communicate counts and averages depending on product or area.
This makes use of k-anonymity, a privacy approach that protects data privacy by never disclosing information about groups of people smaller than k. You may alter the default value of k; the more private the information transmission, the higher the k value. Aible makes the data transmission even more secure by changing the names of variables like “Country” to “Variable A” and values like “Italy” to “Value X” when masking is used.
Mitigating hallucination risk
It’s crucial to use grounding, retrieval augmented generation (RAG), and other strategies to lessen and lower the likelihood of hallucinations while employing gen AI. Aible, a partner of Built with Google Cloud AI, offers automated analysis to support human-in-the-loop review procedures, giving human specialists the right tools that can outperform manual labor.
Using its auto-generated Information Model (IM), an explainable AI that verifies facts based on the context contained in your structured corporate data at scale and double checks gen AI replies to avoid making incorrect conclusions, is one of the main ways Aible helps eliminate hallucinations.
Hallucinations are addressed in two ways by Aible’s Information Model:
- It has been shown that the IM helps lessen hallucinations by grounding gen AI models on a relevant subset of data.
- To verify each fact, Aible parses through the outputs of Gen AI and compares them to millions of responses that the Information Model already knows.
This is comparable to Google Cloud’s Vertex AI grounding features, which let you link models to dependable information sources, like as your company’s papers or the Internet, to base replies in certain data sources. A fact that has been automatically verified is shown in blue with the words “If it’s blue, it’s true.” Additionally, you may examine a matching chart created only by the Information Model and verify a certain pattern or variable.
The graphic below illustrates how Aible and Google Cloud collaborate to provide an end-to-end serverless environment that prioritizes artificial intelligence. Aible can analyze datasets of any size since it leverages BigQuery to efficiently analyze and conduct serverless queries across millions of variable combinations. One Fortune 500 client of Aible and Google Cloud, for instance, was able to automatically analyze over 75 datasets, which included 150 million questions and answers with 100 million rows of data. That assessment only cost $80 in total.
Aible may also access Model Garden, which contains Gemini and other top open-source and third-party models, by using Vertex AI. This implies that Aible may use AI models that are not Google-generated while yet enjoying the advantages of extra security measures like masking and k-anonymity.
All of your feedback, reinforcement learning, and Low-Rank Adaptation (LoRA) data are safely stored in your Google Cloud project and are never accessed by Aible.
