How to Stand Up an AI Team

One of the questions I get most often is what roles you need to hire to build an AI team that can deliver tangible value. This mostly regards companies that are still beginning to leverage AI (and yes, you should! AI is more than large language models and can significantly improve your operational results, whether you work in manufacturing, services, transportation, finance, healthcare, retail, or energy). However, even companies further down the road can be confused.

So here is my take:

1. We can coarsely categorize the science of AI into the fields

Perception | Representation | Forecasting | Planning | Action

When setting up an AI team for your business, you need to identify the domain expertise you need your data scientists to have.

- Do you need to make sense of images? You need a vision expert (perception).

- Do you have a lot of unstructured and semi-structured data? You need an expert on semantic linking and possibly natural language processing (representation).

- Do you need to anticipate the future? You need a machine-learning expert. Note that they come in very different flavors, so you will want to consider whether you need a time series expert or an expert in classification, etc. To make things more complicated, machine learning has revolutionized the work in all five fields of AI, so just because someone is an ML expert does not mean that they work on forecasting. For example, an expert on deep learning and auto-encoders may work primarily on representation in practice (forecasting).

- Do you need to make data-based decisions? You need an expert in optimization and possibly, if the number of decisions to make is not clear upfront, an expert in AI planning (planning).

- Or do you need AI to execute decisions quickly? You may need an expert in reinforcement learning, AI-based controls, or robotics (action).

You may think of these five categories of data science as different parts of an engine, which contains sensors, a carburetor, cylinders, and an exhaust.

2. Just because you have an engine (your data scientists, be they optimizers, knowledge graph experts, or forecasters) does not mean your car moves. You need a gas tank, a gearbox, and a driver. The same is true with an AI team. You need data engineers (the gas tank) to make sure your data scientists (the engine) have data to work with, ML engineers (the gear box) to deploy your analytics, and analysts (the driver) to sniff out trends and problems your team should go after in the first place.

3. With that background, here are your steps:

a. When you start out, you first need a data science leader—someone who understands the science of AI as well as the adjacent engineering roles. Crucially, that person needs to have business acumen and a practical mind. If you hire the next professor out of a university, they will, in general (note that I am talking about the distribution here and not individuals), love problems and the intellectual challenges that they pose. You want someone who gets the science and can go deep, but who will favor simple, easy, and robust solutions that deliver business value quickly.

b. This leader's first task, then, is to identify the AI entitlement in your business. That is to say, before you hire ten people, find out the tasks where AI can benefit your business. This means checking three things:

i. [Do we need a car?] Are there decisions that we execute regularly that could be improved by automating them and/or basing them on more or other data than the human decision-makers we currently use are considering? Note the word "regularly." With exceptions, the biggest value you will get from AI will come from the same set of decisions that need to be made over and over: pricing, inventory relocation, routing, spam filtering, ad placement, and production planning are all good examples. Mergers and acquisitions? AI can help, but you should not set up an AI team to mull over one-of problems. Hire outside help instead.

ii. [Is there gasoline?] Do we have data? It does not have to be big data, but it has to contain enough signals to base robust decisions on them. If your business still uses a paper billing system, you are simply not ready to use AI to identify customers with upselling potential.

iii. [Are there any roads?] Can we integrate AI into our business processes? If there is no way to execute the AI-based decisions effectively, then there is no point in digitizing them. Do not under-estimate this last point: Your AI solutions will not be adopted unless it is easy to integrate them. If you need to overhaul your entire business operations first, don't bother, unless it is existential.

c. The first hire should then either be a data engineer or an ML engineer, depending on the state of your business. Just as you do not build a car around the engine, aside from the leader, do not hire the data scientists first. They will be frustrated, and so will you.

I am curious to learn about your thoughts. Please leave a comment.

PS: Image taken from this interesting article of roles on an AI team: https://www.datacaptains.com/blog/guide-to-data-roles