How much does it cost to build artificial intelligence?

Let’s say you are going to do an AI project for the first time. You have been doing IT projects before. Maybe you even have a lot of experience with IT projects. But now you have to build your first solution based on artificial intelligence, and now what? How do you know what it actually costs to build an AI solution?

The short answer is that one cannot know it in advance. I have previously written about the unique challenges in AI that explains a lot of the reason why. Among other things, you does not know in advance when an AI is good enough. It is also more difficult to estimate the development because the process is more experimental, as you typically do not know the final solution until you are well underway.

In a world where the “business case” is the universal sacred measurement, it is not a particularly useful answer to say “I don’t know”. So what do you do if you want to start building artificial intelligence?

No more funny years

First, I just want to put things in perspective. IT used to be a party for everyone. It was a bit of a casino, where the lowest blinds were pennies and the expensive tables where piles of cash. The entry barrier of getting started in IT was and the possible returns were great. A computer in hand and then you were off. Even the set-in-stone Indian caste system was shaken by the introduction of IT, which became a casteless industry.

That’s not how it is with AI, and in general I think we are moving in a different direction now. GDPR and a lot of focus on proper handling of data - which is essential for AI - has created a larger buy-in than before. I myself have experienced how a few years ago you could get data without problems, but now you have to sign data processor agreements, involve the legal department and maybe even have an IT audit. These are initiatives that require a certain investment to be able to deliver on.

At the same time, AI requires more different experts in development than before. Data scientist, Machine learning engineers, backend engineers and DevOps are pretty much all needed to build a working with AI. It was both easier and cheaper when we just had backenders and frontenders.

Turn your budget upside down

If I had to put together a budget for an AI project five years ago, I would have estimated algorithm and machine learning development to be the most expensive. Implementation would come second, and eventually I would have set aside a bit to find and clean up data that should probably be ready within the first few weeks of the project.

When I budget AI today, it's almost completely upside down. Data is most expensive. Not just in startup, but it is typically an ongoing cost that is an important part of maintaining the AI.

Next, I will budget with implementation, as implementation is typically not just installing some software. AI often comes with a need to rethink the whole business process it is a part of. The people involved also often have to get used to the fact that AI is very different from traditional IT and it requires a cultural change. If you have ever consciously tried to change a culture, then you know that it requires elbow grease and persistence. 

The algorithm development is also expensive and should not be neglected, but if you set aside the entire budget here, you will be missing out on the other items.

Milestone funding

There are some tricks to make your life a little easier and give yourself a chance to build a business case.

First, AI projects should get some inspiration from research projects. Here you often have milestone funding rather than project funding. In other words, money is only set aside for part of the project at a time, and new funding is only triggered if a number of success criteria are met.

In AI, I think there are some natural milestones that you can be inspired by:

Domain knowledge: Can you create a clear vision for the AI ​​and achieve enough domain knowledge to be able to continue work?

 Data collection: In this step you try to collect a reasonable amount of data of a reasonable quality and a natural criterion is what that collection costs. If it's expensive now, it's probably also expensive as an ongoing activity. 

 Modeling: This is where you have to work with algorithms and models to get the first quick prototype, which can show some results. In an AI company I have visited, the rule of thumb here is getting results in one day. If you can not get just reasonable results in one day, then you should be aware. It’s not going to be easier as you dig deeper.

Be also aware - You may risk having to go back to data collection if you do not get good results. 

 Deployment to users: See the AI ​​in a production environment and make it ready for evaluation by users.

 User evaluation: Get the AI ​​evaluated by users and get early feedback.

Once you are through all of this, you can start over and iteratively improve your solution. Of course, not all AI projects are the same, so it is not a universal form. So you should try to make your own milestones for each project.