Artificial intelligence has been exciting the imagination for many years, but real prospects for transforming the business have appeared relatively recently. However, so far 85% of artificial intelligence projects do not bring the promised results to the business (Pactera Technologies data, 2019). Which, however, does not prevent most experts and market players from optimism about AI potential.
The risk is high, but the benefits can be huge. Often, AI projects are unpredictable - you can achieve your goals, you can return to the starting point without a result, and you can find treasures where you did not expect. Let's talk about how you can manage risk in AI projects, but get the most out of it.
Develop Projects Porfolios
All AI projects carry a risk. The best of them bring a huge return on investment and it is easy to find a lot of evidence of this. But many projects fail or cost more than they bring value. A good AI approach is a structured portfolio of projects that allows you to explore many strategies before choosing the best.
To achieve the same business goals, you can explore many ways, and assessing their success at key points, performed on the same metrics, will make it possible to transfer efforts and resources to more promising ones. If one project shows a double return on investment, and another five-fold, it is not difficult to make the right choice.
However, failure can also be useful. From unpromising projects, valuable lessons can still be learned for the projects that you will develop.
As in any innovative technology, the first hint of success often inspires and forces you to continue not the most promising project. This can happen, for example, due to experts (marketers or engineers) who clearly understand their narrow goal and know what data affects the result, but do not have a fully informed business perspective. Which, of course, does not detract from the vital role of the field experts, the importance for the formulation of the problem and approaches.
Often it will be more profitable to take a step back and continue on the different path. Useful data sources may be found that we did not think of before, because they do not correspond to their initial models. Or researching the expected correlation (for example, changing the frequency of newsletters improves sales) will eat all the resources, while in practice other factors have a more significant effect.
Of course, any decisions should rely on business goals, and not on the pursuit of technical excellence. If you can reduce the error rate of the model from 2% to 1%, it is not necessary to do this, if 2% is acceptable.
The best decisions can be made by constant analysis, identification of potential traps and discovery of new and new opportunities. Do not be afraid to abandon projects and switch to more promising routes.
Focus on Data Quality
Machine learning has been extremely successful in solving problems of classification, recognition, ranking, and diagnostics. However, for its effective adoption you need large volumes of labeled data, which in real life is a big problem.
There is evidence that most data scientists spend only 20% of their time on analysis and 80% on finding, cleaning and reorganizing huge amounts of data. Obviously, this is an inefficient data processing strategy.
Contacting various departments to obtain data, waiting, determining whether they contain the necessary information, solving problems with data quality - all this takes a lot of time and effort. And so that data warehouses do not turn into landfills, datasets need to be systematized and classified. Compromises lead to poor-quality training and models that do not work well in practice.
The solution may be a well-thought-out data management policy across the organization and the use of modern cloud tools to automate the tedious processes associated with finding and cleaning data.
However, for a number of tasks there is obviously no data for training. If you want to use AI to predict the failure of a multi-million dollar gas turbine, you are unlikely to find much data on the failure of such equipment. In such cases, the use of trained models, obviously, will not work.
Now we have only touched the tip of the iceberg. As in any new and complex technology, successful application requires taking into account a huge number of factors and nuances. Softline employs experienced and skilled specialists in the implementation of AI solutions to solve problems in various industries. I invite you to review your tasks together in terms of the application of AI solutions.