I've previously described the challenges of trying to hire AI product managers and offered some suggestions. Since then my thinking has evolved, and I've come to realize that as the business landscape slopes up toward AI Abundance, we need an additional specialized role to address the unique challenges of AI transformation.
That new role is the AI transformation manager, or ATM. Alternative titles are machine learning transformation manager or analytics transformation manager, depending on how you characterize your AI projects.
Before I describe the role, I’ll first describe why it’s necessary.
Product managers mitigate adoption risk
For the past 30 years, technology initiatives have presented primarily adoption risk.
For startups, the adoption risk is a market risk or a customer risk. Most startups fail because they build a product that never gets the anticipated market traction. The startup runs out of time, and the postmortem report summary is, “We never hit product-market fit.”
For the enterprise, this risk is a business-customer adoption risk. Most enterprise technology initiatives fail (or take years longer than expected) because the business customers don’t adopt the technology. After the CFO shuts down the project, the (former) CTO says, “We just couldn’t get the business customers to use it.”
An entire generation of best practices and methodologies has emerged to mitigate the adoption risk. Solution examples include agile, lean startup, and customer development methodologies.
AI projects have technology risk
AI projects are not immune from adoption risk, and these projects still need product managers to address this risk. But AI projects pose an additional risk that most technology professionals have never confronted: technology risk.
An AI project can have perfect product-market fit and still fail. This has always been the risk in sectors like biotech. There is no market risk in developing a cure for cancer, but there is a massive technology risk in developing a cure.
AI projects can fail because the team can’t get the technology working. For example:
- Research hasn’t developed a methodology to solve the problem.
- Data coverage is insufficient to solve the problem.
- The computational and infrastructure costs might render the solution economically infeasible.
This new class of technology risks require new mitigations. For example, AI projects should not begin by blindly following the agile methodology. Instead, data scientists need time to explore data, run experiments, and solicit feedback about the feasibility of alternative approaches. The project must mitigate these AI technology risks before proceeding to an agile methodology that minimizes technology risk.
ATMs mitigate AI project technology risk
The primary job of an ATM is to mitigate the technology risks associated with AI projects.
An ATM ideally has the following skills and experience:
- A foundational understanding of data science practices such as methodology selection, experimentation, training-data acquisition, feature engineering, performance evaluation, and tool selection. This knowledge is critical for communicating with data scientists and contextualizing their results for customers.
- Infrastructure experience, ideally with deploying and scaling models in modern platforms such as Kubernetes and service meshes.
- A background in data analytics sufficient to anticipate challenges such data coverage, ownership, and usage rights.
- An ability to simplify and explain AI challenges to project stakeholders.
You’ll notice that this list doesn’t include the traditional product management activities related to changing user behavior, understanding needs, or adoption.
The ATM is responsible for identifying and overcoming the technology risks associated with an AI project:
- Providing an initial assessment to stakeholders about what AI can do.
- Advocating for the data, tools, and infrastructure data scientists need to do their jobs.
- Identifying potential methodology patterns (for example, text classification or table extraction) across different customer groups.
- Anticipating and overcoming legal or policy challenges related to data usage. For example, challenges might include the use of sensitive information in model development or contractual constraints on the use of customer or partner data.
- Helping data scientists identify and run the experiments necessary to mitigate methodology risk.
- Critically reviewing the models and systems for ethics considerations.
- Ensuring governance processes are followed.
- Scoping the service architecture necessary to support the application in partnership with the MLOps team.
- Identifying the critical system components that require monitoring and escalation based on the methodology and business risks.
How the ATM relates to other key roles
Like any product team role, the ATM’s responsibilities and skills will overlap with other roles. There are no clear distinctions, and every team will develop their own working model. The following table shows a high-level overview of activity ownership.
Abbreviations for team role
ATM: AI transformation manager
PM: Product manager
DS: Data scientist
DE: MLOps and Data Engineer
Abbreviations for responsibility
P: Primary responsibility
S: Supporting responsibility
Recruiting, hiring, or training an ATM for your team
The ATM is a new job category that satisfies an emerging need. My previous advice holds for creating this capability at your company.
To learn more about how to recruit, hire, or train an ATM, see Winning the War for Data Science Talent.