The critical roles you need for your AI Project

Day by day we see more and more AI solutions appearing around us. Or rather, some of them can actually be seen (like customer service bots or biometric facial recognition) while others remain in the back-end systems where only the highest level of specialists can “see” and appreciate their invaluable input. But they are there nevertheless so one can’t help but wonder –who do you actually need to make an AI project successful? You can’t just mail-order them or download from a disk. You need a specific skillset from actual human beings to make it work.

It goes without saying that an AI project can’t do anything for your business if it’s not well thought through from business perspective in the first place. This process is described here in a bit more detail but for the sake of this article let’s assume we have specific and clear process in mind and we have decided to automate it using AI. Another thing we should set straight from the beginning is that every project (regardless of its aim) is actually a result of people working together and in order to make this project successful — we must bring out the best from those people and let them do what they do best. Making communication between them smooth and relationships healthy and productive.

The 4 critical roles

Dependent on the complicity of the process you are automating, the number of actual human beings behind those roles may vary so let’s not think of those as FTE-s but specific jobs that need to be handled. The minimum number of a living and breathing soles behind each role is 1 but dependent on the task at hand — it could be more.

It should always start with a business problem owner — someone who feels that he / she is in trouble. I cannot emphasize this enough — it has to be an actual human being with thoughts and emotions and feelings because companies don’t have problems — the people who fill specific tasks for those companies do. A company is just a number in some registry, everything that is actually creating value is produced by the people who make this company work.
If the company is run poorly then it’s usually the 
chiefs of something or shareholders who have those problems. But in a team that is well managed (usually having leaders instead of managers) those problem owners are top specialists of a specific field or middle management members because they have the in depth understanding of the company’s processes and actually feel it’s their responsibility to make them better.
The cool thing about an AI project is that it cannot and shouldn’t ever start from IT but from the person (from the business side) who has a problem. And it really makes no difference whether this person is from accountancy, law or mechanics department.

This person will now hold first of the critical 4 roles and it will be up to him/ her to:

- Define the problem we are solving and explain why it is important (to whomever it may concern)
- Make the GAP analysis and bring out the (objective) differences between the situation today and how it should ideally look like in the future
- Be ready to put a significant amount of time, effort and energy into whatever is next to come, thus understanding that this is not just another IT development project where you describe your demands and wait for the brighter future while the IT wizards do their magic.
- Pick the correct metrics and KPI-s that should be measured in order to assess how successful your AI project actually is and also sets targets for the whole team based on that
- Will constantly (and throughout the project) be the one to evaluate how the team is doing (business vise), providing the feedback and guiding them toward the business results desired.
- Will set up the necessary processes to keep the AI project in line with business after the development phase has ended.

It is critical that all this is done by the actual person who owns that business process. Even though all this seems like a full time job you should never hire a separate project manager to hold that role. The minute this thought enters your mind — you have chosen a wrong project.
If all this seems like too much time and effort for your specialist / manager then you shouldn’t solve that problem with an AI project. There are many alternatives that will end up cheaper and more easily implemented — and if they work for you — choose those! An AI project should never be brought to life just because it’s AI and seemed like a cool idea.
But in case your problem owner nodded enthusiastically to every single point above — you are all set to move on and 
take a look at data.

The technical term that now needs to be addressed is data engineering and to put it bluntly — nothing good can ever come from a crappy data set. AI will not be the miracle tool that turns water into wine — in order to get wine you have to provide your team with proper grapes and someone has to cultivate, harvest and prepare them. That’s basically what a data engineerdoes.

The quality of the dataset sets the ground for a successful AI Project

First task will be to understand what sort of (if any?) digital data the process components provide. If the process you are trying to automate does not generate any meaningful digital data, then it’s quite hopeless to push forward with AI. It would be much smarter to take a few steps back and re-create the process so it would start generating necessary data sets.

But should you already have digital data — it’s now time to evaluate the actual quality it withholds. Every piece of digital information that is stored somewhere cannot be considered AI-worthy material and a good way to determine what is worth using (and what should be ignored) is turning toward your business problem owner. He or she should be able to explain how a human being uses those data points and what conclusions are drawn based in which information. If that comes easily and seems logical to your data team — you might be ready for your harvesting.

But it is important to keep in mind that more often than we might think — things don’t work out that easily. Even if someone claims that “all our IT systems have been communicating with each other for years — we have everything under control” — it doesn’t mean that it will work for your AI solution. The requirements are just so much higher that it is completely normal to spend most (!) of your AI development phase on data engineering.

Data engineer will be responsible for:

- Creating datasets from data sources
- Developing and managing the infrastructure to transfer the data
- Building the pipelines from (sometimes several different sources) and preparing the data for a data scientist, which usually means a multi-step process to construct, test and maintain architectures, such as databases and large-scale processing systems
- Will be “forever” responsible for the quality and usability of the datasets used by your AI

It makes no difference whether all those tasks are handled by one or several different people. It is perfectly normal to have (a group of) people who have in-depth knowledge of your data sources and the links between IT systems and can guarantee that those systems will keep on providing the necessary data. And then another group of people who know how to make this all useful for a data scientist — to build the pipelines and handle the technical infrastructure for those purposes.

The last part can actually be outsourced quite easily since it is a specific skill and especially with your first AI projects — it might not be the smartest idea to invent all the wheels by yourself.

Now that we have data and it is “harvested” properly we are approaching the third role — the data scientist (well the vintner to keep the wine analogy alive).
Just like with wine making, it takes so much more than just good grapes to produce an excellent wine, and therefore this is the part where the actual magic and mystery happens. Although no data scientist allows me to call it either — this is what it looks like and how it should be respected by us others, mere mortals. You should never allow any amateurs near your data who “have just googled data science, read 5 blog posts and now are super-exited to try it out”.

The data scientist will:

- Clean the data so it could be used for building a model
- Find the patterns and trends that are hidden in the data and are actually relevant for solving this particular business problem (not just odd coincidences)
- Create and train the machine learning models that will be the heart and brains of your AI
- Evaluate the results and correlation with the business KPI-s
- Visualize and communicate the results for all involved parties.

All this will be a mixture of mathematics, statistics, programming and even a little touch of biochemistry — so in short — not for layman. Just like owning a calculator does not make you an accountant, you can’t just wake up one day and decide that you are a data scientist. 
If you are serious with your intent to actually solve that business problem, then you should now decide between two options.

You could hire your own team of data scientists or find a partner that will provide this as a service. Both have their pro’s and con’s but from my personal experience — the first option is very rarely justified.
Main reason being that data science is a hot topic all over the world and the better part of those hot (young) data scientist really don’t fancy the idea of working for a dull old corporation (and all corporations are immediately by definition old and dull for them). Since it’s the data scientist’s interest to continuously improve and develop herself/himself with all the trends and new solutions that are constantly emerging — they just wouldn’t feel fulfilled by the projects of just one company. On the other hand, a company that provides that service for multiple different customers is a whole different story — so using such a service provider actually gives you the brainpower and skills that you really need, but wouldn’t be able to attract by yourself.

Elisa’s partner in data science (and data engineering) has been MindTitan for a couple of years now. And it’s no secret that this Estonian machine learning and data science company has been the key for all our successful AI projects so far (and hopefully in the future as well).

It is my belief that hiring one’s own data science team might make sense only in the case where your main product is actually ML based by itself or achieving that is your main goal.

And even though it now might seem that we have all our critical roles in place (we have the business problem, the people who understand where data comes from and know how to use it, plus we also have cool young data scientists with their magic powers) — this setup still doesn’t work. Not by themselves anyway. They need one more person — someone to make them all play nice together.

As I said before — companies don’t have problems, people do. A company is still just a number in some boring registry. Everything that is actually happening in those companies can be thought of as human relationships going wrong or right. And since people cannot be considered rational creatures, on the contrary, they are charmingly emotional and unique beings — every single one of the 7 billion of them — a lot can go wrong when people don’t feel good while interacting with each other.

Projects usually succeed when all the participants like what’s going on and even feel happy being a part of the process. They respect each other and value others’ opinion. There is mutual trust and no unnecessary overlapping or micromanagement. Sounds ideal?

Well. If you take an enthusiastic and emotional business representative that is driven by the customer centricity and actual problems that clients have with your products. Then add a steady and analytical DW specialist who likes to describe system parameters and only talk about facts and logical correlations. And add a third person — a data scientist (probably from another company), who really doesn’t get who should know what or answer to which questions — then you will pretty soon end up in a situation that is really uncomfortable for everyone involved. The cooperation really just doesn’t work, a lot is lost in translation, deadlines are never met and it all just seems too much hassle with very few happy moments. Like trying to open a wine bottle with a fork. It can be done — but the result is messy and the process definitely not enjoyable.

The role that is missing is a bridge between all the others — a translator. The best suited title in my opinion would be machine learning product manager. We could call her or him a project manager as well, but that has a temporary whiff to it and this position is definitely anything but temporary.

This role will be the one to:

- Understand and explain how the business units operate and how business value is created
- Knows what is “regular IT” and is familiar with the common programming languages but also understands what is machine learning and how it differentiates from regular IT
- Has good people skills and knows how to communicate to both sides (and explain/ translate between them)
- Owns the toolbox of leading people without legal authority
- Keeps the team productive, within lines and clearly explains the expectations to each participant

Once all those boxes are ticked — you are all set and ready to go!

All the key ingredients are already here today for you to build your first (or next) AI solution. All you need is a good team to bring the results home. Sure, the science will keep on progressing but there is no good reason to stand still and wait while that happens.

If you wish to see your company as prosperous as it is today in 5 years’ time, or to be 5 times more successful by then — then today is the day to get started.

Define the business problem, find the players for each of the 4 roles mentioned above and start experimenting!
Just keep in mind that the last role is the key to it all. Otherwise you’ll end up with a fork, a messed-up wine bottle and cork crumbles all over the place.

The story was first published in Estonian, IT uudised 11.05.2019