How Bedrock enables AI governance

We are witnessing an ever increasing dialogue around responsible AI. Even within the first month of 2021, leading analysts such as Forrester and Gartner have already warned about serious ethical and social problems that AI is predicted to cause in the next several years. This year, the attention will be on how governments and organisations are finding the path towards responsible AI deployment

In previous posts, we outlined why enterprises need AI governance and why machine learning operations (MLOps) is so critical for enterprise AI

Despite the wave of ethical AI principles and guidelines published between 2016 and 2019, brand-eroding AI blunders are still commonplace. At BasisAI, our perspective is that the issue with AI oversight and accountability lies not within the governance frameworks, but the chasm that exists between the guidelines and engineering practices. 

At BasisAI, our perspective is that the issue with AI oversight and accountability lies not within the governance frameworks, but the chasm that exists between the guidelines and engineering practices. 

Before we dive into how Bedrock bridges the gap between governance and deployment, let’s refresh our memory on the key elements of AI governance that matter when developing AI solutions.

What is AI governance? 

AI governance is a framework and process for organisations to ensure that their AI systems work as intended, in accordance to customer expectations, organisational goals and societal laws and norms. When integrated with other parts of the organisation, decision tradeoffs can be made in view of overall compliance and risk management perspectives. 

AI governance comprises two broad components: guidelines for action, and systems and processes

  1. Guidelines for action are consistent principles that guide the design, development and deployment of AI in a way that is explainable, transparent and ethical. This leads to responsible use of AI, which ultimately builds trust in AI.
  2. Systems and processes are a set of quantifiable metrics, clear roles, actionable steps and auditable processes such as thresholds and controls on implementation.

At the heart of AI governance is the visibility of machine learning models, as a proxy to the trustworthiness of the AI solution. To achieve this, there are 4 key principles of AI governance that need to be operationalised :

  1. Oversight - To achieve oversight of AI models being used in your organization, as well as ensure visibility into ML performance before deployment and on an ongoing basis.
  2. Maintainability - To diagnose issues, fix errors, roll-back to safe versions quickly, as well as trace the provenance of models and audit measures taken. 
  3. Explainability - To identify key factors that led to decisions, even for complex AI models.
  4. Fairness - To identify unintended biases against minority groups eg. race or gender.

How Bedrock addresses AI governance

As an engineering discipline, MLOps provides a way to practically apply these governance principles as you productionise your AI solution.
Bedrock, our flagship end-to-end MLOps platform, has various features baked into it that are designed to empower data science teams to seamlessly adopt the principles of AI governance into their workflow - something we refer to as “governance by design”. 

Here are 4 common AI governance problems that Bedrock can help to solve:

#1 - Going beyond notebooks for productionising AI models

Notebooks are great for exploratory data analysis and prototyping machine learning models. But the problem is, notebooks are not designed to be the final stop in the journey to creating a robust and maintainable AI model! In a typical AI-mature organization, these models may be handed over to ML engineers, MLOps engineers or even software engineers for deployment, management, and optimization for scalability.

Bedrock distills multiple best practices in software engineering and MLOps, such as version control for models, pipeline schedulers, deployment scalability, and adapts them to the world of data science. This makes it easy for any data scientist to manage and deploy AI models at the level of competency and reliability provided by a dedicated team of ML engineers, MLOps engineers, and software engineers. On top of that, Bedrock reduces the time-to-market of machine learning systems by up to 70%.

Check out these Bedrock features:

Overview of a productionised project with model versioningFigure 1: Overview of a productionised project with model versioning

#2 - Managing collaborations throughout projects and teams

Companies that want to realise the growth potential of AI must have a thriving collaborative environment for developing machine learning models. Having teams of data scientists share their notebooks with other colleagues or other teams manually becomes especially painful as projects grow in complexity and scope.

Bedrock steps in with the ability to manage users, teams, and approval workflows over any combination of projects and organization members. These features allow managers to grant the proper permissions to each team member for each machine learning project, such as permissions for model training, model reviewing, and model deployment.

Check out these Bedrock features:

Role assignment for project collaboratorsFigure 2: Role assignment for project collaborators

#3 - Incorporating explainability and fairness as part of an AI model’s continuous evaluation process

We have seen greater calls for AI transparency across all industries. Businesses that deploy AI models are becoming inundated with requests to explain why some AI models made a particular decision, and those used to treating AI models as a “black box” may find it increasingly untenable to continue subscribing to these models. 

Bedrock brings together the latest interpretability methods for AI models such as SHAP, as well as fairness metrics using AIF360. The Bedrock user interface (UI) also allows users to select a single inference within a trained or deployed model, and quickly conduct a deep dive in the model’s decision-making process to understand the level of explainability and fairness.

Check out these Bedrock features:

Fairness metrics for a protected attribute
Figure 3: Fairness metrics for a protected attribute

#4 - Implementing monitoring and auditability for AI models

A successful, deployed AI model is often supported by robust monitoring and auditability. 

Monitoring comes in many flavours. Two of the most important ones are: tracking the performance of the AI model, and tracking the usage of the AI model. 

Auditability is a must-have in today’s world where AI models are increasingly held accountable for every decision that they make. Along with the fact that AI models do not improvise well in novel situations such as changes in human behaviour, AI model custodians often need to step in at the appropriate time to initiate a re-training or update AI models to keep them fresh.

Bedrock makes it easy for users to track and audit AI models for both short-term and long-term deployments. In the Bedrock UI, users are able to view an audit trail of the historical model performance across multiple model versions. This allows them to keep abreast of feature drift and any changes in the inference distribution of each model. In addition, users are able to view the throughput, response time, and error rate of each deployed model’s endpoints. Users can take action to either scale up or down the deployment of the model based on the usage metrics. 

Check out these Bedrock features:

Monitoring feature distributions between training and production dataFigure 4: Monitoring feature distributions between training and production data


Despite the multitude of benefits an organisation stands to gain from employing AI governance practices, the fact is that many enterprises still use AI as a black box, perhaps out of convenience or an urgency to derive a return on investment. More often than not, complex algorithms are built into software that is trained on synthetic or limited datasets which do not capture the full nature of human behaviour, and pushed into deployment. Even big tech companies that build and deploy complex algorithms on a regular basis are not immune to the occasional unintended or even harmful decisions that may arise from such algorithms.

As a result, enterprises may not realise until it is too late that they have little control over the kind of actual data that the models are ingesting. They have poor visibility into how the models are performing. This compounds into tricky situations where it becomes difficult to detect unintended biases within the ML models, or even in the case when a decision made by an ML model is not at all explainable. Might the algorithm be treating men and women differently? Why did the algorithm decide to do so? Was the data used to train the model not representative enough? In a model designed, built and deployed as a black box, there is no way to really know.

From a business point of view, black box AI systems also do not provide a path for the organisation to be data ready. In certain cases, these black box models can also be detrimental to the business’ reputation if it is found to have made controversial or harmful decisions for its customers. 

We recommend a glass-box approach to building AI; one that uses clear frameworks and empowering platforms which enable governance-by-design. 

If you would like to continue the conversation and find out how Bedrock can help your organisation employ AI governance and scale responsible AI, please get in touch.