Exploring generative AI to maximize experiences, decision-making and business value

In the first part of this three-part series, generative AI and how it works were described.

IBM Consulting sees tangible business value in augmenting existing enterprise AI deployments with generative AI to improve performance and accelerate time to value. There are four categories of dramatically enhanced capabilities these models deliver:

• Summarization as seen in examples like call center interactions, documents such as financial reports, analyst articles, emails, news and media trends.
• Semantic search as seen in examples like reviews, knowledge base and product descriptions.
• Content creation as seen in examples like technical documentation, user stories, test cases, data, generating images, personalized UI, personas and marketing copy.
• Code creation as seen in examples like code co-pilot, pipelines, docker files, terraform scripts, converting user stories to Gherkin format, diagrams as code, architectural artifacts, Threat models and code for applications.

With these improvements, it’s easy to see how every industry can re-imagine their core processes with generative AI.

Leading use cases do more than simply cut costs. They contribute to employee satisfaction, customer trust and business growth. These aren’t forward-looking possibilities because companies are using generative AI today to realize rapid business value including things like improving accuracy and near real-time insights into customer complaints to reduce time-to-insight discovery, reduction in time for internal audits to maintain regulatory compliance and efficiency gains for testing and classification.

While these early cases and the results they’ve delivered are exciting, the work involved in building generative AI solutions must be developed carefully and with critical attention paid to the potential risks involved including:

• Bias: As with any AI model, the training data has an impact on the results the model produces. Foundation Models are trained on large portions of data crawled from the internet. Consequently, the biases that inherently exist in internet data are picked up by the trained models and can show up in the results the models produce. While there are ways to mitigate this effect, enterprises need to have governance mechanisms in place to understand and address this risk.
• Opacity: Foundation models are also not fully auditable or transparent because of the “self-supervised” nature of the algorithm’s training.
• Hallucination: LLMs can produce “hallucinations,” results that satisfy a prompt syntactically but are factually incorrect. Again, enterprises need to have strong governance mechanisms in place to mitigate this risk.
• Intellectual property: There are unanswered questions concerning the legal implications and who may own the rights to content generated by models that are trained on potentially copywritten material.
• Security: These models are susceptible to data and security risk including prompt injection attacks.

When engaging in generative AI projects, business leaders must ensure that they put in place strong AI Ethics & Governance mechanisms to mitigate against the risks involved. Leveraging the IBM Garage methodology, IBM can help business leaders evaluate each generative AI initiative on how risky and how precise the output needs to be. In the first wave, clients can prioritize internal employee-facing use cases where the output is reviewed by humans and don’t require high degree of precision.

Generative AI and LLMs introduce new hazards into the field of AI, and we do not claim to have all the answers to the questions that these new solutions introduce. IBM Consulting is committed to applying measured introspection during engagements with enterprises, governments and society at large and to ensuring a diverse representation of perspectives as we find answers to those questions.

Learn more about this three-part series by reading the first or third installment, and reach out to an expert for start a conversation about your needs.