Machine Learning Teams Have a People Problem

“I manage products, not people,” is a common quip from those supervising technical teams. It’s also dead wrong.

If you’re struggling to build great products and things are falling short, you may not have a product problem. You might have a people problem — 65% of failures of VC-backed startups happen because of people issues.

In this article, we’ll share the three biggest challenges startups face on their first year of building an AI/ML team. These insights, and many practical solutions, come from conversations with founders who successfully built productive ML teams in healthcare, agriculture, and finance across the world — from New York to Cape Town to Jakarta to São Paulo.

We’ll also share a few things we learned at Google, while building strong data science/research teams.

No matter what, growing AI/ML teams is hard work (possibly harder than your business or operations teams). So to help get you started, here are the three main areas where we often see teams struggle.

If this in-depth educational content on implementing AI in the business setting is useful for you, subscribe to our Enterprise AI mailing list to be alerted when we release new material. 

Getting the culture right

Depending on how core AI/ML is to your product, you’re probably coming from either end of this spectrum:

Owkin, a healthcare startup based in New York and Paris doing remarkable work across biomedical, genomics and clinical data, launched in 2016 and began on the left side of this spectrum. They had a strong research orientation, hiring some of the sharpest minds from top academic institutions. But working through deep and difficult research questions was taxing work, and incubation time was long. Their researchers were deeply motivated by building accurate models and looking across data modalities to discover new knowledge, with the goal of publishing their discoveries. However, their operational discipline of creating goals/OKRs and tracking against them, or productizing work done for specific customer cases, was not a strong point

Meanwhile, the rest of the team were more motivated to build features that were just fast enough, reliable enough, and scalable enough for production.

This is a common starting point for teams that are building products with AI at the core (as opposed to products that can succeed without any AI application). It also tends to happen when the founders themselves are from academia, as was the case at Owkin.

A balance must be maintained. It’s important to preserve a part of the culture that’s incubation-oriented, allocating enough time and energy to work through the complexities of a problem to get the answer right. At the same time, a company needs to produce products. They need to launch and experiment with users. They need to show forward motion and traction.

In contrast, other startups begin at the right end of the spectrum. Often, the founders have a strong commercial background and a desire to get something out in the market and learn from the experience. Launch and iterate. In startups like these, product management dominates and the company rallies around an urgency to ship new features. Some would even argue that you should question the need for features that rely too deeply on ML expertise. Take for instance Google Search Ads. A decently profitable business for Google and a product that needs almost no AI to be excellent.

Whichever side of this spectrum you come from, whether research or product driven, you’ll want to strike a balance between incubation and production. You need to get your product to market if you want to succeed. But you also want to keep your data scientists engaged and motivated. (You’ll want to make sure you actually do need data scientists. You might not.)

How do you balance that out?

1. Hire people who obsess over the problem, not the solution.

• When assessing interviewees, see how much they care about real-world problems (vs. research problems). Consider a question like: “If you were to design product X with almost no AI, how would you go about it?” Or “What user adoption challenges would you expect from sophisticated, AI-powered product X?”.

2. Help academics make a mental shift by exposing them to real and complex adoption challenges.

• Douglas Eck, a Principal Scientist at Google AI, made the transition from academia in 2010. Since then, he has talked about an important mental shift that needs to occur during the leap — one that took him several years to achieve.

The main mental flip is to put in a lot of energy to understand the problem without assuming the research skills you have will matter at all. Some people lead with a research solution. And sometimes an honest solution doesn’t really need AI.

• To help someone do this, show them all the real challenges of landing (not just launching) a product. We see this a lot in healthcare. Building models to recognize cancer cells is one of the easier parts of the work. The real challenges are sociological and political — doctors’ aversion to robots replacing their expertise, regulatory hurdles, etc. Give everyone in your team, including the data scientists, exposure to these parts of the problem.

3. Amplify the work with project managers.

• Your researchers should focus on the highest and best use of their time. Do this by having disciplined, action-oriented project managers who can run multiple projects, help everyone stay on track with deadlines, and manage the relationships with different stakeholders — including the client (in the B2B scenario). A word of caution: a project manager can quickly feel deflated by researchers who feel a high sense of self-importance, or who feel like the project manager doesn’t understand the work well enough, pushes for unreasonable deadlines, or poorly represents the team’s work. Find a project manager who is skilled at leading without being overbearing and has a genuine interest in the technology.

4. Formalize incubation time.

• “We run a journal club, where our scientists have time and space to pick up recent research papers and figure out how to implement the ideas they read about into our product,” says Benji Meltzer, co-founder of Aerobotics, a Cape Town-based agritech startup that uses drone imagery and AI to boost crop yield. He proudly talks about how they’ve been able to keep their data scientists engaged as they work on interesting problems and bring in new thinking.

Keeping your data scientists

You either grow them or lose them. Data scientists are in high demand and you risk losing them if you don’t give them a compelling vision of professional growth in your startup.

Data scientists tend to have different career aspirations compared to typical software engineers or product managers. While product teams may be motivated by product launches, researchers and data scientists love finding elegant technical solutions for tough problems, with the prospect of publishing their discoveries. How do you fight intellectual boredom and keep these folks engaged?

Several ways, including the following:

1. Take care of your people managers — they matter a lot.

• People managers have a disproportionate impact on growing people in their teams. “People don’t leave companies, they leave bosses,” is a proven truism at Google and everywhere else. Is that exacerbated with AI/ML research teams? We believe it is. As you might guess, Googlers in our ML/AI research teams are often the target of other companies, big and small, trying to bring AI into their products. So our people managers have an important task: they must understand a Googler’s career goals and connect them to relevant work or mentors. We invest a lot in our learning and development resources at Google, and people managers receive a big share of it. They are critical to every company, and so your investment should reflect it.
• One important point: beware of “outsourcing” the work of a people manager to programs or processes. Startups make the mistake of launching people initiatives, like new performance management processes or career ladder frameworks, without getting their people managers behind it. When done well, people managers create a culture where team members are excited to do their work and see a future in the company. When done poorly, well-intentioned people initiatives are undermined.

2. Find mentors.

• This is the most underutilized way to help people grow. Find strong mentors from your own network, or from academic institutions you work with. Some founders have tapped the network of their investors — they’re likely investing in similar types of startups in other markets. And don’t underestimate the power of a cold email. You’ll be pleasantly surprised that most people are willing to help out and give some of their time.

3. Explore academic partnerships.

• Consider partnering with academics and have your team supervise the work. You get the double benefit of expanding your research team fairly inexpensively and can keep your team engaged in new problem sets. Watch out for some of the intangible costs here — who owns the intellectual property, how much control you’ll have, and how you’ll hold them to output. This work is usually longer-term in nature and work you won’t mind sharing control over.

Leading a data science team when you’re not one yourself

This last point is specific to founders who don’t have technical backgrounds and must engage closely with data scientists.

Frustration can quickly build up on both sides with a setup like this. Here are some considerations to think about.

1. Don’t micromanage.

• Manage for specific outcomes. When you’re not as familiar with the work, the temptation to micromanage is high. It’s natural to feel nervous about timelines, especially when your data team is asking for a lot of time and no involvement from you. “When you’re tempted to micromanage, find ways to be clear about the output and a monitoring system,” a CTO tells us. “Chunk the work and have agreed on check-in points to provide feedback.” Focus on output and give your team the space and trust to figure out how to get there.

2. Involve data scientists early.

• Involvement creates commitment. There are always important data pipeline issues to be sorted and by giving someone a seat at the table, you welcome their influence over other parts of the work, communicate that you trust their expertise, and build commitment to the work. Don’t give them a designed product and expect magic.

3. Keep pace with expertise.

• Rafa Figueroa was an economics major, a banker, and now the CEO of a healthcare company based in São Paulo called Portal Telemedica. Even though he employs data scientists, he eventually realized that he had to build some technical depth to lead his company well. This meant viewing every YouTube video possible (some 100 hours worth!), digging into academic archives, and reading five peer-reviewed research papers a week. To this day, he consistently seeks out mentors and meets people working on deeply technical problems.

Now, with these three challenges in mind, let’s think ahead

There are many other important people challenges, particularly as you get into year two or three of building your team. Formalized career paths, growing people into “head of data science” roles, org design (align to engineering vs. product, for instance) will all play a part.

Let us know what people challenges you’re facing. We can help you sort them out. Let’s continue our journey to startup success together.

The article is prepared in collaboration with Josh Yellin. Josh is based in San Francisco and is the Global Lead for Google’s Launchpad Accelerator program, which to-date has partnered with 300 of the world’s foremost growth-stage startups. Josh has mentored many of these startups in people and operations topics.

Special thanks to Bar Vinograd for his insights.

This article was originally published on The Launchpad and re-published to TOPBOTS with permission from the author.

Enjoy this article? Sign up for more updates on AI business applications.

We’ll let you know when we release more technical education.