How to make your AI team faster and more productive

AI teams often move slower than the business wants. In this post, I share the habits I’ve used to speed teams up while keeping them motivated: shipping simple models early, making deployment easy, fixing how data is stored, and keeping the team close to the people it serves.
Business moves faster than science
Quick, which is faster: business or science? Business, usually. Picture a team of brilliant data scientists in a back room while a marketing manager knocks on the door asking for the ranked list for the latest campaign.
Manager: “This needs to ship in two days.”
Data scientist: “But it isn’t perfect yet.”
My favorite engineering saying fits here: if it works, it doesn’t have enough features.
I can’t remember working with a company where this tension didn’t come up. Data scientists don’t fit neatly into two-week sprints. We get stuck on too little data or too much, and investigations take longer than planned, because we’re often working out how to solve something that hasn’t been solved before.
Imagine building a system that personalizes emails for a toy company. Depending on the data and how accurate the early models are, the team might need weeks of feature engineering (combining and reshaping the data so the model has more to learn from), or more data, or a new approach. Meanwhile, marketing wants to know when it can use the thing.
These are the seven fixes I’ve seen work.
1. Release simple models early
Convince the team to release simple models early, and reward them for it. A simple rule usually beats doing nothing. A recommender that shows everyone the most popular products isn’t glamorous, but it’s much better than nothing, and it gets data flowing through the whole system. Breaking the habit of polishing a model for months has another benefit: the team gets more done, sees its work used, and gets credit for it.
2. Make deployment fast
When data scientists can’t get their models into production, their work goes nowhere. Build a good relationship between the AI team and the engineers who run your systems. Make sure those engineers understand exactly what they’re deploying and how often it needs to be refreshed. Agree on the boundaries, so the engineers aren’t recoding models and the data scientists aren’t configuring servers. When deployment is slow, data scientists start to wonder whether their work matters, and before long they leave for a company where it does. I cover the whole process in putting AI into production.
3. Store data the way the team uses it
Storing data in the right shape saves a lot of time. If pulling basic customer data takes hours of joins across ten or more tables, everyone slows down and errors creep in. Build a summary table of the customer data the team uses every day, and every later request takes seconds. Find out what data your team needs daily or weekly and make it easy to reach. I’ve seen this one change speed up whole teams, with the bonus that everyone works from the same figures. There’s more on this in why so much of an AI project goes to cleaning data.
4. Work in pairs or groups
Group projects move faster, produce better results, and make for a happier team. If a project has a single technical person on it, consider adding a second. In my experience, it has always paid off.
5. Show the work to the rest of the company
I call this data science sales. On a project for a large client with nearly a dozen internal stakeholders, we made sure every major stakeholder saw and understood the main AI projects. Two things helped. We gave the team a simple web page where people could try out its work and see the results. And we had a designer spend time with the team each week, turning the work into visuals the whole company could follow. When people across the company understand the work, projects start sooner and finish with fewer surprises.
6. Put a data scientist in the weekly business meetings
Invite someone from the AI team to the major marketing, sales, strategy, product and engineering meetings. I like to set these up weekly. Within a few months, I’ve watched productivity and new ideas grow on both sides, and each side learns to explain itself more simply. AI that sits far from the business is one of the five reasons AI projects fail to pay off.
7. Hold regular one-on-one reviews
Reviews take real time from managers, but they’re worth it. Data scientists, like everyone else, want to know where they stand and what it takes to reach the next level. Quarterly one-on-one reviews work well. Cover:
- Major accomplishments
- Upcoming projects and goals
- Recognition for good work
- Guidance on upcoming work
- Where they stand on the team and in the company
Mentorship matters here too. I wrote about why in mentorship is not optional.
I’ve used these techniques for years and seen them work. If your team needs a hand, our engineers work inside client teams every day. You can read how we work with clients.