Costs Of AI Development - What Do You Need To Take Into Account?

Artificial Intelligence (AI) is one of the hottest business trends of the early XXI century. The machine learning technology behind modern AI-powered solutions is frequently compared to the steam machine when thinking about the impact on business and society. But is technology affordable when the company is not a GAFA-sized player? 

Artificial neural networks that power up the dominant share of modern Artificial Intelligence are not a new concept at all. The first theoretical approach was done during World War II by Warren McCulloch and Walter Pitts, who created the computational model of a neural network. Yet during that time, the subject was clearly academic research with no viable perspectives to apply in daily life. 

There were two barriers to overcome in the early days of artificial neural networks. 

The computing power - the great idea behind building an artificial neural network was to emulate the working of the human brain. And that’s a nearly impossible task. Various estimations point that that a human brain - and not an extremely brilliant one - operates on the computing power reaching 1 exaFLOP that equates to a billion operations per second. The strongest computer currently existing, a Summit computer being currently the fastest supercomputer in the world is expected to operate on 200 petaFLOPS. But it didn’t reach this power yet. 

So every human skull carries a computing power worthy of five or more strongest computing machines existing on this planet. Wachowski sisters - maybe that’s a good hint for the upcoming Matrix?

The data - to be effective, every Machine Learning solution needs the mind-boggling amount of data. Again, the natural neural network within the head learns from years of data and connects multiple channels to make all the system work. Developers expect the neural network to reach and accuracy of an expert not in a lifetime but in some viable time. Depending on the Machine Learning project it can be an hour, a week or a month, but not a dozen years for sure. 

With the big data and cloud computing explosion, both challenges were overcome. Gathering data is easier than ever before and the cost of warehousing it systematically get reduced. Also, the cost of computing power in cloud drops 

By that, companies got access to the modern AI-based solutions and all the scientific concepts found their way into the daily business operations.

What can you gain with AI software development?

Modern AI can bring significant benefits for every company, be that an international giant with thousands of workers or a smart and lean SMB willing to expand the operations. There are multiple uses of Artificial Intelligence:

• Decreasing fixed costs - there are multiple costs considered frozen, that the business needs to deal with without complaining or a way to reduce. These are usually utilities or renting costs. But the AI can bring significant savings on that. By using a Deemind-developed AI model, Google was able to reduce a cooling bill in the data center by 40%. And when it comes to computations, cooling is one of the most significant costs. 
• Automatization and increasing the quality of processes - Artificial Intelligence can augment any process that requires a tedious and repetitive job, be that responding to customer inquiries or reading the same documents on and on. Amtrak, the US railway company, saved over $1 million by using chatbot instead of delivering traditional customer service. A chatbot named Julia responds to customers at first line and providing them with accurate information on the schedule and available tickets. 
• Anomaly detection - an anomaly in business is usually a sign of something suspicious and unwanted. That can apply to fraud detection or quality control in the company, depending on the type of business run.  
• Saving on the expert's cost (long term) - artificial neural networks already surpass human performance in many fields. There are solutions in development that support doctors in comparing human body scans and thus deliver accurate information on the patient’s health. According to various estimations, a computer can process an image up to 1000 times faster without losing accuracy and focus.  

The examples above are only the tip of an iceberg. More information about the ways of using AI in business and general perspectives of augmenting the processes with machine learning can be found in our complete guide to AI in Business. 

How much AI development costs?

Examples mentioned above undoubtedly sound tempting. Yet building the AI-based solutions, despite being affordable, is far from being cheap. Before jumping on board, the company needs to take several unavoidable costs into account. 

Business consulting and feasibility studies

Every company is different and there are no one-size-fits-them-all solutions. There are of course SaaS or out-of-the-box software solutions that deliver results, yet to deliver a tailored requires a good amount of consulting and studies. And when it comes to AI, the connection with existing processes is tight. So tight in fact, that in the ideal situation, be that auto-completing the Google query or automatic suggestions in Netflix - the process and the AI solution are inseparable. 

Data scientists and R&D

Data scientist, a guy or girl who develops machine learning in business models and algorithms powering AI-based solutions, is a rare beast. Despite the increasing amount of them in the market, the demand grows even faster. That’s why data scientist is considered the best US job of 2019. With the increasing complexity of the solutions and the rising impact of AI in human daily life, the cost of data science talents will only increase. 

And that’s only a human-side of the research and development process. The next step is data gathering and labeling. And that can be tricky. The data needs to be standardized and labeled properly to teach the model to deliver the desired outcome. What’s even more tricky, there can be hidden biases within the dataset that can stealthily hamper the company’s performance in the long-term. Or cause an outrage - Amazon had recently been criticized for developing a gender-biased model that automatically rejected all female engineers applying for the job. 

Computing power 

Another cost that can grow. Running the model usually isn’t computational heavy. But training it is a whole different story. To effectively train neural networks, the company (using or developing AI) needs access to a tremendous amount of computing power. The challenge is usually solved by running the computation in the cloud, but that generates costs. High costs to be honest. 

MVP

A model performing task is one thing - but model performing the task for a business benefit is a whole different story. Before the implementation, companies need to build a Minimum Viable Product, that shows that all this Machine Learning for business is not a gimmick and can bring significant improvements. And that’s a cost either, but a required one - trying to build an AI solution without the MVP is a clear way to the disaster, as there are multiple flaws and mistakes that go out during this phase.

Implementation

That’s the hard part and the implementation can deliver an unexpectedly high cost. It is uneasy when it comes to any software or solution, be that an ERP, CRM or other enterprise software. 

The cost itself is hard even to estimate, as it varies depending on the size and the industry. But to have a glimpse of it, it is wise to consider: 

• The cost of automatic preparing the data taken from other systems
• Delivering the infrastructure to power the process seamlessly
• The processing itself
• Making the effect actionable for other systems and employees 

Maintenance and further learning

The machine learning model is not a refrigerator that serves unchanged for years until it gots broken. The solution needs to be maintained and updated accordingly to the changes in the business model and the environment. 

A good example comes from image recognition software - the quality of images rises over time and users get access to devices with better cameras. The model needs to handle the bigger images or get an update that resizes them to the easily-chunked size. 

Summary

Building an AI Startup or AI-powered solution is neither easy nor cheap. But the savings it brings and business effects it delivers is worth the effort. In the long term, it will bring a lot of money, both as increased income and reduced expenditure. 

Wondering if AI solutions will be profitable solutions? Contact us for a free consultation.