Things to consider when managing projects delivering AI solution
AI - causing a seismic change in society. Companies are rolling out AI products and services left right and centre. For PMs, it's providing a new set of challenges
ChatGPT was used to help with this article.
AI is the hottest technology trend at the moment. It’s making a seismic shift in society, and people are working out how to deal with it.
There are courses on prompt engineering, machine learning and data science so people can adapt and take advantage of the new technology.
For project managers, there will be a lot more projects that either deliver AI solutions to customers and businesses or use AI tools as part of the delivery process.
In either of those two scenarios, project managers and their teams will face several challenges. One challenge is the delivery of a robust and ethical AI solution.
We have to stop SkyNet
We all know the movie, one of the all-time classics. And there are jokes about SkyNet when it comes to AI.
What it does show is the problem of an AI product that does not act in the best interest of people. It might not take over the world and kill us all off. But it may make a decision where you don’t qualify for a loan when you actually do. Or tell you that you didn’t make it through the first round of the interview, and you don’t know why.
The challenges for project managers are not necessarily at the technical aspects of AI, but about AI ethics and transparency. The challenge is about building a fair, robust, transparent and secure AI solution that helps people.
With this exponential growth, ethical guidelines, policies and regulations have been struggling to keep up.
As project managers, we have a significant role to play in implementing AI projects and using AI to deliver projects in a transparent and ethical way.
We need to ensure that end users can trust the outcomes and recommendations of AI to be reliable and accurate.
AI Governance is about ensuring that AI is developed responsibly and ethically. It ensures that the AI solution we’re delivering or use of AI in project delivery does not harm people and make them worse off. This may not be obvious to begin with or may not be obvious at all.
Pillars of AI Governance * are:
Accountability - defined roles and responsibilities for the AI systems and their outcomes. This means that the correct stakeholders are assigned to each role and is responsible and accountable for each part of the AI system and the outcomes. For example, finding owners for the following: model selection, data use for training, ensuring that legal, ethical and regulatory obligations are met, etc.
This ensures that nothing falls through the cracks
Decisions can be traced
The right people are making the right decision within their area of expertise
The right people are involved in managing any incidents
Transparency - openness and understanding of the AI models implemented, how data is used and the decision-making process for the AI to reach the outcome. This is achieved through documentation and, use of Explainable AI (XAI) techniques. This helps to build trust with stakeholders and end users. It also allows improvements to be made using the available information.
Fairness - this ensures fair and equitable outcomes and treatment for end users of the AI solution. Without this, you are risking being open to legal consequences, creating mistrust in the end users and harming the people they are meant to help. A lot of it has to do with the data that you’re using to train the AI and also the rules you give it to use.
Reliability and Safety - AI systems are robust, secure, operate as expected and minimise harm. This builds trust with the users, ensures that the AI tool meets ethical criteria so they do not cause harm to people.
Human Oversight - People can review decisions,have control of the systems and can intervene to change the outcome.
How can Project Managers ensure AI projects incorporate these pillars of governance?
Initiation and Planning
During Initiation and Planning, you will need to lead the analysis to see if AI is the right solution for the problem. A review of the ethical framework and impact of society for the AI use and its outcomes needs to be done.
Confirm roles and responsibilities for the development of AI. There needs to be specific accountabilities for development, testing, fairness, ethics, risk management and legal and regulatory requirements
Set up a framework that allows traceability of decisions, model changes and ethics reviews
Evaluate the data and its quality to make sure that the data suits the application to the AI and the desired outcomes.
Identify risk and mitigation actions specific to AI. For example, the risk of poor data quality impacting the model’s training which results in poor outcomes for the end users. A mitigating action would be to ensure data quality is sufficient.
Execution and Monitoring
During the monitoring and execution phase, ensure that there are controls in place to manage data and security to protect critical and sensitive information.
Model training need to be validated and oversight to ensure that the AI outcomes are accurate and reliable.
Need to actively look for bias and mitigate any found during the AI training.
Ongoing tracking of AI performance against agreed metrics and indicators.
Develop a feedback loop so that AI decisions and outcomes can be reviewed and improved, addressing any issues that have been found.
Closure
During closure, ensure that formal sign off has been completed on AI performance that meets ethical, legal and business customer requirements
Have a structured document explaining the AI model that can be shared and updated as the AI develops.
Fairness and Transparency
Use explainability tools throughout the project. This will allow the stakeholders and project team members understands the output and why it came up with that outcome.
Include user-facing disclosure to inform users that AI is being used. This creates trust and offers transparency.
Throughout the project and during model learning, conduct transparency audits on the model, data set and decision processes.
Define fairness criteria and get them agreed at the beginning of the project. Use it as a standard to measure the AI development throughout the project.
Use fairness testing tools like Fairlearn or AIF360 to identify and mitigate any bias in datasets, design and model learning.
Identify and engage a diverse group of stakeholders during requirements elicitation and testing. This helps to ensure different perspective and information is gathered and verified in the AI development.
Monitor fairness after deployment to ensure bias is not introduced into the AI.
Privacy
Data is critical to AI. Project managers need to ensure that user’s privacy are protected. Most organisations would already have controls in place, the project would need to ensure that AI integrates with these existing controls. Below are not AI-specific controls for Privacy; they can be extended to include AI.
If these are not in place, then the project would need to include them as part of the deliverables.
Make sure you have identified the different types of personal data being captured and have a valid reason why you are collecting and storing it.
Ensure that the data is being anonymised and encrypted, and controls are in place to prevent people from having unauthorised access.
Have regular reviews to make sure privacy regs are being met and resolve any issues found.
This regular review process can then be used once the AI is deployed and part of business-as-usual activities. There may be an existing organisational process that the project can use.
Reliability and Robustness
This is to ensure that the AI solution provides consistent outcomes and decisions and is capable of withstanding any attempts to attack or fool the AI. Some of these activities are standard practice within software and technology projects.
Conduct stress and edge case testing.
Include adversarial testing, where testers try to break or fool the AI.
Set up monitoring to identify model drift and a process to remediate it.
Build a failsafe and allow human intervention if the AI model degrades.
Safety and Security
The suggested controls below have been around before AI. They can be extended to AI solutions.
Identify vulnerability and design defences proactively.
Use a red team to attack the AI to identify security and ethical weaknesses.
Ensure there are security protocols around the model and data. (encryption, access, monitoring)
Have an incident response plan that responds quickly to minimise harm.
Human Centric Design
Conducting user research to gather real human needs, concerns and vulnerabilities so that people’s needs are met.
Design of interfaces that clearly explain outcomes so that users can make informed decisions and build trust.
Design human in the loop controls so that human has the ability to review, intervene or override AI actions and outputs.
Assess the psychological and social impact of an AI system on identifying and preventing any unintended negative effects.
The Human Project Manager
As you can see, project managers still have a big role to play in delivering AI solutions.
There are a lot of new and changes to existing activities, tools and techniques that a project manager need to get across.
By developing your knowledge, skills and experience in these areas, you would improve your capabilities to work with this far-reaching technology.
Sources and References
Below are two sources that I used directly for the article.
‘*’ Associate Professor Rob Nicholl, University of Sydney, from his presentation to the Australian Institute of Project Management on AI Governance.
Cornelius Fitchner, AI for Project Manager Course.
Below are some resources that I reviewed for the article:
https://shap.readthedocs.io/en/latest/
https://oecd.ai/en/ai-principles
https://fairlearn.org/
https://www.ibm.com/think/topics/explainable-ai?mhsrc=ibmsearch_a&mhq=explainable%20AI
Definitions
Bias - this is where systemic and unfair discrimination occur in machine learning models due to bias in training data, alogirthms or decision processes.
Explainability - how the outcomes of AI outcome and the process it took to come up with it is understood and can be explained. This helps with building trust in the AI model.
Fairness - outcomes and decisions made by AI are fair, equitable and no discrimitory.
Hallucination - AI generates output that is incorrect or misleading or made up but still sounds coherent and logical.
Model drift - properties of the data being used for training changes over time and leaders to reduce ability of AI to come up with fair outcomes and decisions.
XAI - a collection of tools and techniques that make AI model decisions transparent, traceable and understandable.
My Resources
Below are some of my free resources if you’re new to project management.
Templates
Paid Resources
Simple and Effective Project Risk Management is for those new to project management, the accidental project manager and experienced PMs wanting to learn something new. It will teach you what you need to know and get you started quickly. I’m running a promotion where the first 10 people who buy this get a discount, pay only $19, and can book a 30-minute coaching call with me.