In our previous blog, we explored the concept of risks associated with deploying AI in your product or service, examples of issues that have come up with AI solutions and the impact those failures have had on tech startups. While nothing is foolproof and AI tools such as large language models (LLMs) are still in their infancy, there are steps organizations can take to try to prevent these failures. This post explores strategies mature startups can use to prevent AI product failures before they happen (or mitigate them when they do happen).

Pre-deployment planning

Preventing product failures starts with careful planning before deploying AI functionality. As an organization, you should ask questions such as:

• What is the goal we’re trying to achieve by deploying AI? Or, what problem are we trying to solve?
• How will we define success when it comes to deploying AI? What metrics or key performance indicators (KPIs) will we use?
• Who is this AI-powered solution or functionality for? What will they do with it?
• What are our benchmarks for success?  
• What will trigger corrective action, and what will that action entail?

In addition, it’s important to look out for any vulnerabilities and possible failure points, from data inputs to model outputs, to spot where things could go wrong. This can include assessing security architecture, evaluating data sources for biases, and scrutinizing deployment environments for potential threats.

Finally, evaluate what the impact would be of any potential failure, including the monetary impact and non-monetary effects, such as leaked personal information and damage to reputation.

Using the right data

As is often said, AI is only as good as the data that trains it. You need high-quality data in order for the AI models you’re using to function accurately. Some best practices for data collection and handling include:

• Ensure your data is collected and used in a consistent manner, using the same formats.
• If appropriate, use data from a variety of sources, such as from third parties, to reduce bias.
• Try to use as much data as you can, which will help the models capture more variations and nuances to improve accuracy (as long as the data used is of high quality and fits the formats you’re using).
• Implement comprehensive data governance policies, ensure your solution fits the necessary regulatory requirements (such as HIPAA and GDPR), and check to verify that inputs and outputs are properly masked to avoid sensitive data leakage.
• Train and test your models before deploying them in your solutions. Split the data you use into training, validation, and test datasets to avoid bias and overfitting (or aligning too closely to a limited set of data points).
• Consult with legal experts familiar with the specific data protection laws applicable to your industry and jurisdictions to tailor your data governance policies accordingly.

Note, however, that while using a larger dataset can improve model accuracy, it is also important to consider the relevance and quality of the data. Over-reliance on AI decisions without proper oversight can lead to errors and legal liability.

Putting humans in the loop

Implementing AI into your solutions isn’t “set it and forget it”. Because both AI developers and the companies that use AI can be held accountable for its output, it’s important for organizations using AI to set guidelines for its use and maintenance. The allocation of liability between AI developers and users can be complex and depends on contractual arrangements, the nature of the AI's use, and applicable laws. Here are some ways in which your teams can ensure the AI you implement into your solutions is behaving as it should:

• Conduct regular data audits and ensure you’re using clean data wherever possible, which entails removing duplicates, data with missing values, and outliers, while maintaining cybersecurity best practices.
• Use a feedback loop to mitigate hallucinations, or untruthful content that LLMs sometimes generate. Analyze prompts and responses, and make necessary changes to data and modeling to improve responses.
• Consider using a human-in-the-loop validation workflow, in which a person approves LLM responses to reduce the likelihood (and liability) of an incorrect response.
• Establish clear guidelines for when human intervention is required, and train staff to handle AI-related issues effectively.

Implement a company-wide AI policy

One of the best ways to avoid costly incidents with regard to AI is to establish a comprehensive AI policy for your organization. An AI policy establishes the core principles and guidelines for safe development and adoption of AI technologies within the company. Creating a culture of

responsible AI use helps foster a culture of transparency within your organization that’s beneficial both internally and externally.

Creating a strong AI policy includes:

• Conducting an inventory of your current AI use.
• Identify the potential risks associated with the tools you’re using.
• Encouraging internal entrepreneurship to explore new AI opportunities.
• Establishing clear guidance about tools that are approved to use and how to request approval for new tools.
• Implementing a “stoplight” structure for unrestricted, controlled, and prohibited use cases.

Read more (and create your own AI policy) via Vouch. 

Taking a multifaceted approach to responsible AI use

Mitigating AI failure within your solution and your organization requires a multi-pronged approach that includes risk assessment, utilizing high-quality data, rigorous model and data validation, continuous monitoring, and encouraging transparency within your organization through a comprehensive AI policy. While these strategies can significantly reduce the risk of AI product failures, it’s essential to recognize that some level of risk will always remain. 

Ultimately, mitigating and preventing issues that can arise as a result of AI requires staying in compliance with legal requirements and insuring your organization against AI failure. Stay tuned for our next blog, where we’ll explore navigating the legal and insurance landscape around AI.