Getting the Needed Talent Along the AI Value Chain
JF Gagné JF Gagné
October 29 6 min

Getting the Needed Talent Along the AI Value Chain

I’ve just published this year’s Global AI Talent Report on jfgagne.ai, and this year we took a wider view of the talent pool. We estimate that globally, there will be 86,000 authors pre-publishing fundamental and applied AI research papers on arXiv in all of 2020; and that there are approximately 478,000 people globally with the specialized AI technical skills needed for building an AI product. You can read the full report with further breakdowns of the data here.

Below I share why understanding the talent pool can help explain the friction in getting AI to work.


In recent months, I have been asked many times why AI hasn’t played more of a role in fighting the pandemic or the economic recovery. This common question has sparked an important conversation to reset expectations of what AI is and what it can do. AI is a challenging technology to work with, and the communication and media coverage about AI promising a magic solution has inflated expectations and reduced organizations’ willingness to take on that challenge.

It’s important to realize that AI is just software, albeit with new dynamics. I think seeing it as just software helps remind organizations of the familiar experience of rethinking their businesses around new digital capabilities. With traditional software we have seen new business models emerge as the enabling capabilities came online. The sharing economy was not an impossible business model to foresee, but getting onto it early required an ability to work with the key supporting technology of GPS, fast bandwidth, and wide cell coverage. This takes specialized technical talent combined with niche domain expertise.

AI is still coming out of its infancy, and we are only just beginning to get a real grasp of how to work with its new dynamics. Instead of coding with logic and rules, AI is coded with data. This new way of coding alone, not to mention the carry-on effects of its new capabilities, requires whole new processes and tools in order to standardize and scale up for industry. There is an eventual payoff of a smaller software footprint; requiring relatively fewer hours to create, that delivers greater overall performance, but for now the requisite expertise is creating a high-bar of entry for companies who want to innovate with it. What we have as a result is businesses looking to minimize the changes required to apply AI, which cuts off most of the opportunities to create new business models and leaves mostly add-on features that increase speed and efficiency of current processes.

This doesn’t mean we should wait and ignore the big and real potential of AI-enabled technology solutions to accelerate a recovery, whether in medicine discovery, re-designed and robust supply chains, or increased productivity in a digitally-based workplace. The recovery needs to make use of the new capabilities of software, but it should be done to attract both the investment and willingness to take on the challenging journey of AI innovation.

Knowing how to invest today means knowing what talent is needed.

The low success rate (about 10%) of creating AI products that drive large-scale impact is a feature of not having standardized tooling and processes. Without those tools, highly specialized technical talent is needed across the board.

Once researchers make a proof of concept in the lab, the trick is in getting it to deployment in the real world at scale. This takes engineering the AI capabilities for the real world environment, developing software around it to make it into a product, data architecture to feed it the data it needs to work, and domain expertise to ensure there is a feasible business case in the first place. Without standardized tooling and processes, that means the core engineering talent is needed across the value chain to support and even take on some of these roles full time.

This stretches that limited talent, and the resources to pay them, and it makes it difficult to specialize in a domain or focus on building the tooling needed for repeated innovation. It is a poor division of labor. Seeing that, some businesses are putting their qualified talent back into engineering tools, pulling them from solutions development. There is a lot of new talent filling in from the multitude of new professional courses to help support building AI solutions; but, without the full depth of technical talent that combines cutting edge techniques and deep domain expertise, much of this new influx is put to the old problem of building features for old business models.

AI talent avilable to industry

This creates a cyclical problem of talent being pulled back and forth, but it can be resolved with sustaining the right focus. Turning AI product development into a predictable process will help make AI accessible for both domain experts to identify new opportunities and developers upskilling themselves to build out the solutions. Eventually we may see many more specialized and narrow roles emerge.

These more specialized roles will revolve around working with data. Today we see this primarily in data wrangling—collecting, cleaning, and labeling the data needed for training and operating AI. This is a massive part of the workload, representing 65% of the hours of a machine learning project according to Cognilytica. Much of the work is low-skilled, but some areas have shown the need for domain expertise, such as labeling medical images. As the tools develop, I predict we will see more of this mix of domain expertise and data work, including business unit leaders and even front line workers able to contribute to the ongoing training and development of AI solutions, without the need for a data scientist to help them directly.

Until these tools emerge, the friction of spreading around engineering talent will continue. Developing tools for governance and monitoring AI models will be another significant area of growth. Adaptive systems that learn and change with inputs cannot be left alone and these tools will be critical for effective deployment. Better understanding what skill sets are needed where, and tracking the changes along the value chain, will help with the balancing act of talent development, and also help ensure we don’t lose track of investing in the research that keeps pushing what’s possible.

Education leaders should work together with industry to incorporate into their curriculums these new processes and tool development for getting a proof of concept into real-world deployment and under effective governance. We have seen this happen with Software 1.0, and so the roadmap should make clear how we need to re-standardize for this new way of building software.

Read the full talent report

Contact us to explore how we can help you implement AI, using our expert team.