Fine-tuning AI to deliver business value

Microsoft launched Copilot Tuning a few months ago. It allows its customers to use low-code tools in Microsoft Copilot Studio for highly automated fine tuning “recipes” that are trained on enterprise data.

While GenAI tools are often associated with AI models trained on large amounts of public data on social media and internet platforms, businesses require models that understand their own data and processes. This is the current focus for major AI providers. Commercial AI

These products aim to provide AI systems that are industry-specific, as opposed to the highly generalised models which have been trained using freely available internet data. Theoretically, they should be less susceptible to hallucinations than the more general AI models. They should also better match the way businesses work.

Ranveer Chaudra, vice president and group product manager for experiences and devices at Microsoft In a blog: “AI Tools powered by out-of the box LLMs [large language models] ” Retriever augmented generationdoes not always understand the specific processes, terminology and style of your business. Microsoft’s approach in optimising AI models for businesses is to reduce the complexity that comes with fine-tuning.

Accountancy firm Ernst & Young is one of the customers participating in the Microsoft 365 Copilot Tuning early access program. Marna Ricker said that the firm was integrating its fine-tuned LLM for tax with its enterprise knowledge, and the expertise of their tax advisors, through M365, to deliver an enhanced service to the market. “This synergy increases service quality and advances tax and legal research, with relevant knowledge and information readily available in M365, where people are already employed,” she added.

Gartner’s forecast states that the market for GenAI models specialised to specific industries will more than double by 2026, reaching $2.5bn. Gartner’s forecast of $23bn for general GenAI models is much smaller, but it still shows that businesses are interested in this technology.

Roberta Cozza is a senior director analyst with Gartner. She said that the major AI providers are fine tuning their models because this is where enterprises want to go. She said enterprise buyers value working with trusted technology providers, but also need GenAI tools to respond to something specific to their domain. She said that “what we are seeing is domain-specific modeling.”

Cozza pointed out that many of these models are based on open source models and are deployed as small-language models (SLM). These models offer efficiencies when it comes to resourcing costs but also better control because they can be trained using an enterprise’s data.

GenAI is a valuable tool, but enterprise IT leaders that she has spoken with want it to be trained on the issues, data, and content specific to their industry. Microsoft and major IT consulting firms are increasing their AI offerings to cater to enterprises that now want to deliver business value using GenAI. However, IT leaders need to consider alternative approaches. Cozza said that enterprises should put their proprietary data into the hands of a model builder, or an IT service provider.

The barriers to entry for basic open-source models have decreased a great deal, so smaller AI providers are helping large customers by providing their own small model,” said Cozza. “They can distill a proprietary model, like ChatGPT, or they can start with Meta’s Llama. In Europe, we see Mistral as a beginning point.” They are also evaluating how to stay compliant with the EU AI Act. She said that “AI applications and technologies deemed high-risk will have to be regulated to comply with the EU AI Act.” “But this does not cover frontier models that are based on internet data.”

Cozza explained that models based on internal data of a company and SLMs are less likely to be subjected to regulatory scrutiny. “Training AI and creating something more domain-specific actually helps with general compliance, because you can make the AI comply with policies or regulation,” she added.

Tool like M365 Copilot Tuning ( ) will help lower the barrier of entry for IT leaders that have been tasked to provide GenAI capabilities to add business value. However, SLMs can offer an alternative that is more explainable and can comply with the EU AI Act.

Cliff Saran ()

• Interview – Developing a CIO Strategy for Artificial Intelligence

  Cliff Saran ( Cliff Saran]

• Cliff Saran (
  

  Cliff Saran]

  Cliff Saran (“DigDeeperItem-3” Cliff Saran]

• ]”Cliff Saran”[19][19][19][19]]]]]][19]][19]]]]]]]]]]]]]]]]]]]]]]]]]]]]]][]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]?[19?]?]??[194590]]?]?[194590]]?[194590]]?[194590]?[194590]?[194590]?[194590]?[194590]