Private AI models: redefining data privacy and customisation

AI has hit the mainstream. Public consciousness has been gripped with the release of large language models, like ChatGPT, and the productivity benefits they deliver. Despite how it may feel though, AI isn’t the new kid on the block, it’s just gaining a better name for itself.

From predictive modelling to task automation, organisations have long applied AI to speed processes and improve customer and user experiences. However, organisations across Australia are today being challenged to integrate AI into wider workflows at scale, which has traditionally meant utilising large external, public cloud providers or developing in-house models.

Don’t put sensitive organisational data in the hands of public AI

Facing pressure to do more with fewer resources and lower budgets, organisations feeling pressure to apply AI rapidly, or risk falling behind, may turn to large public cloud providers or utilise public AI models for their AI deployments. Large public cloud providers offer a range of tools and resources to support development, deployment and management of AI-driven applications. These large-scale providers provide vast computing power and resources like virtual machines, containers and serverless computing options that can handle intensive workloads like training AI models. Public cloud providers also have extensive data storage capacity, which can be critical for companies scaling to accommodate rapidly expanding datasets. And they offer access to pre-built services and tools to simplify the development, deployment and management of AI applications, including pre-built AI models, APIs and tools for data preparation and model training.

Alarmingly though, all public cloud providers share one glaring weakness: data privacy. Providing proprietary or sensitive data to these large cloud providers can be a risky proposition, as it’s not uncommon for them to use customer data to train their own algorithms.

It’s worth noting this was enough of a concern that OpenAI changed their policy to explicitly state they won’t train their models on customer data for API-developed applications. However, they still may use your input into the ChatGPT chatbot for training that model unless chat histories are disabled.

Despite its surging popularity, what should not be overlooked is that because AI is a way of accessing and synthesising data, it means that an organisation’s own data is more valuable than it was before. Even for businesses that have strong use cases for AI and are looking to quickly deploy, no right-minded CEO would want their company to upload their own data to train an AI algorithm that they don’t own. Large public cloud providers often build their business models on the premise of gaining access to data. They use their customers’ data to hone their own algorithms. Making matters worse, these algorithms are shared by all their customers, which means an organisation’s own property data could be helping its direct competition.

For many organisations, the lack of privacy makes public cloud providers a non-starter. And organisations in industries under strict data privacy compliance laws will be particularly wary of integrating AI into their workflows for fear of a leak or simply the uncertainty that comes with sharing data without clear parameters around how that data will be used.

The challenges with building in-house AI

Given the very real concerns that Australian organisations have around public AI today, many may look towards building models using an in-house team. While building AI models in-house delivers higher levels of privacy and security, there are long-term costs to maintain the model and the underlying infrastructure, along with significant challenges in building out experienced teams that can execute work to the level and speed required.

Integrating an in-house model into existing workflows can also bring challenges of its own. Building in-house AI models isn’t just a matter of hiring software developers — the complexity and costs depend greatly on the underlying technologies involved. The more complex the IT ecosystem, the harder it will be.

Rapidly and securely deploy with private AI

In response to the challenges around taking on AI development in-house and privacy concerns with public cloud providers and public models, interest is growing in a new private AI approach which allows people who are unwilling to divulge their customer information and risk exposing it to a competitor or the public to still experience the benefits of AI.

Process automation platforms that include native AI capabilities allow organisations to train their own private models without privacy concerns or prohibitive in-house costs. With a private AI model, an organisation is provided with a set of algorithms that exist exclusively for their use. The user controls access to these algorithms and the data used to train them, which protects their  information during model creation, training and use. This means that data is controlled by the organisation that is trying to create, teach, expand and make the algorithm more efficient.

With private AI an organisation creates its own models trained on its own data that’s never used to optimise anyone else’s algorithms. They are creating their own private model, not offering data to train the cloud provider’s larger algorithms or benefiting a rival organisation.

Beyond privacy, private AI models are also just more practical. Each organisation has a unique set of customers, products or services and needs. With private AI, they can tailor their algorithms for their own business, rather than having to use a more generic, publicly available algorithm.

Another advantage of private AI is that it can be accurately audited, unlike a public model. When an organisation receives a result back from AI, it may not know why the AI would provide the answer that it did, or even if it was a mistake. And if the answer needs assessing or correcting, this is near impossible with a public model, because an organisation has nowhere to look for what data informed the answer, because all the data was processed. With private AI though, it is easy to look at the data the AI accessed to develop its output and investigate why something incorrect occurred.

Protecting an organisation’s most valuable asset

Organisations across Australia are grappling with the dual challenge of harnessing AI’s transformative potential while safeguarding their proprietary data against the vulnerabilities inherent in public cloud environments. The emergence of private AI offers a middle road between innovation and information security. By enabling organisations to develop bespoke models that utilise their data exclusively, private AI circumvents the privacy pitfalls associated with public cloud services, while still speeding up deployment compared with in-house approaches.

Private AI signifies a critical step forward towards a more secure, personalised and efficient digital future. By prioritising privacy and customisation, organisations can leverage the full spectrum of AI’s capabilities without compromising their most valuable asset: their data.

*Luke Thomas is the Area Vice President, Asia Pacific & Japan, at Appian Corporation. Prior to his tenure at Appian, Luke has had more than 20 years’ experience across the software and services solutions industry, including previous leadership roles at Oracle Corporation and IBM.

Top image credit: iStock.com/sasha85ru