Data science startup Domino Data Lab Inc. has become the latest company to introduce a platform for automating governance in artificial intelligence development, saying its extensive experience with large organizations and federal government agencies gives it a unique advantage.

Domino Governance is an attempt to automate the largely manual processes that many organizations use to ensure that AI development follows a set of responsible practices. The firm cited its own research that found that 95% of the 278 enterprise AI leaders it surveyed believe their organizations need to rewrite or update their AI governance frameworks to prepare for the broad adoption of generative AI.

AI governance refers to the policies, frameworks and processes that guide the development, deployment, and management of AI systems in a way that is ethical, responsible and safe. It typically includes guidelines for fairness, transparency, accountability, privacy and security.

Numerous recent studies have found that governance is a major sticking point toward broader AI adoption. A Deloitte LLP report released last week found that just 3% of 460 board members and senior executives it surveyed believe their organizations are very ready to deploy AI. Lack of board-level attention to governance was a major reason.

“The biggest problem isn’t so much that organizations don’t know what the policies and processes should be because they’ve been established for a long time,” said Kjell Carlsson, the firm’s head of AI strategy. “The problem is the tools they have are from the Stone Age. Everybody’s using spreadsheets to track everything.”

Quick to build, slow to approve

Lack of coordination between stakeholders involved in AI development worsens the problem by making it difficult for reviewers and auditors to get the information they need, said Ahmet Gyger, Domino’s senior director of product management. “One of our customers took three months to build a model and seven months to get the model through the governance process,” he said. “It can take multiple weeks just to find out if personally identifiable information was used.”

Carlsson said the result is that deployments are delayed and the risk of model drift – or degradation of a model’s performance over time – increases.

“You end up with models that are almost by default out of date before they’re put into production, and they are much less frequently updated because of all of the difficulties and hurdles” to that go into deploying them, he said.

Unpredictable results

The same best practices of risk assessment, access control, versioning and auditability that work in traditional software development still apply, but generative AI has introduced new variables because the factors driving decisions can change in real time. That can lead to biased outputs or hallucinations that can incur harsh regulatory penalties.

The EU’s AI Act, for example, can punish violators with fines of up to €35 million, or 7% of a company’s annual revenue. At least 45 U.S. states have introduced AI bills, and 31 have adopted resolutions or enacted legislation.

Domino Governance includes a set of configurable templates based on frameworks such as the European Union’s AI Act and the National Institute of Standards and Technology’s AI Risk Management Framework. Reviewers can audit data, see work context and review packaged evidence without disrupting production.

Policies are enforced automatically through challenge questions and approvals. A central dashboard provides a single view of policies, compliance status and open actions across all projects and models. These are augmented by a governance maturity assessment and professional governance services.

Integrated model monitoring detects drift and performance issues across all types of AI models and data in a hybrid, multi-loud or on-premises environment.

Carlsson said Domino Governance addresses the significantly greater complexity of machine learning and deep learning models compared to traditional software applications. “You need to be able to track back from the output to the version of the model, the version of the code, the version of the environments and libraries and the version of the data that was used,” he said. “That’s where a lot of organizations struggle, because those factors are often on entirely different systems that aren’t linked.”

Domino Governance will be available later this month.

Image: Pixabay