Preparing for and implementing AI projects can be a multi-year journey. According to the latest figures, only 28% of respondents reported getting past the AI planning stage in the first year. This is due to several factors including the relative maturity of the technology (at least in the ever-expanding set of industry use cases), the level of complexity involved such as extensive integration requirements, limited enterprise experience and lack of internal skill sets, concerns with AI bias as well as governance, risk and compliance concerns, extensive change management requirements and more.\nWith so much emphasis on demonstrating quick wins, whether as part of corporate innovation programs or digital transformation initiatives, over-long AI projects can potentially impact the reputations of much larger initiatives than just their own. As CIOs move from \u201cprojects to products\u201d in their approach to product management, these lengthy AI projects can delay innovative new internal or external product releases as well.\nTo gain quick wins around this important enabling technology and further the business case for more investment in broader digital transformation and innovation initiatives, here are five ways CIOs can fast-track their AI implementations:\n(While we focus on AI machine-learning (ML) initiatives with examples relating to loan decisioning in financial services, the recommendations are applicable to many other AI initiatives and industries.)\n1. Build or buy based on whether AI is going to become a core competency for your organization\nOne of the first decisions to make is whether to build or buy. While we hear a lot about the various platforms, infrastructures and frameworks for build-it-yourself AI, the unsung heroes are often the more niche, specialty AI vendors who offer cloud-based AI services which can be quickly trained and deployed for your particular use case. The decision to build or buy is really based upon how critical AI is to your organization as a future core competency.\nAs an example, while every financial services company should be concerned about the looming digital and financial divide between the AI \u201chaves\u201d and \u201chave-nots\u201d (see \u201cTaking a counterintuitive approach to business strategy and technology deployment\u201d), not every company needs to build their own algorithms in-house. Smaller shops can quite effectively focus more on the business benefits and outcomes of incorporating third-party AI technologies into their core workflows, such as loan underwriting, without having to build their own internal AI\/ML expertise.\u00a0\u00a0\u00a0\n2. When it comes to data, \u201cmore is more\u201d and quality is key\nIt was once said that success is 10% inspiration and 90% perspiration. When it comes to AI, a successful implementation is often 10% AI and 90% data. Any datasets that are used to train AI\/ML algorithms to mirror human decision making need to be as large as possible and as cleansed as possible.\nIn simple terms, this means that 10,000 rows of data with 1,000 attributes per row is far more useful to an ML algorithm than 1,000 rows of data with 100 attributes per row. According to Marc Stein, CEO of Underwrite.AI, a company focused on applying advances in AI to provide lenders with non-linear, dynamic models of credit risk, however, it\u2019s not quite as simple as \u201cmore is better\u201d. The data type and quantity has to be matched to the algorithm type. Deep learning requires a massive number of records to be effective, whereas statistically-based algorithms deal better with smaller data sets.\nIf you\u2019re using AI to model human decision making, get as much data as you can, ensure every data field has a value, and place a premium on data quality and consistency. This can be time-consuming, especially if drawing from multiple disparate sources, but if done thoroughly early on, it can avoid a lot of costly rework.\u00a0\n3. Spend time on change management and training on how to best interpret the results\nWhile it\u2019s technically straight-forward to call an AI API to pass in a new dataset and receive a score, what\u2019s far harder is the change management and training that\u2019s required to enable business analysts to best interpret these scores and incorporate the new process into their daily workflow.\nWhile some forms of AI may produce automated decisioning, such as a \u201cYes\u201d or \u201cNo\u201d decision on a new loan based on credit history, it\u2019s often the case that ML algorithms provide a more subtle response as well. This response may need to be used in conjunction with existing human processes to best decision the loan. As an example, the AI \u201cscore\u201d may be a grade from \u201cA\u201d through \u201cD\u201d and \u201cF\u201d. \u201cA\u201d and \u201cF\u201d may be the clear-cut \u201cYes\u201d or \u201cNo\u201d decisions which can be fully automated for real-time decisioning, but grades \u201cB\u201d through \u201cD\u201d may still require a human underwriter in the loop.\nJust as you spend time training analysts to use a new financial model and how best to interpret the results of the model, the same is true for AI-based results. Business analysts may need to spend several weeks or even a month just observing the results coming back from the ML algorithms, so they have a baseline in terms of how to best interpret the scores. If you\u2019re working with an AI vendor, this vendor can provide guidance in terms of how to interpret the results and how to train employees to get the most out of the new system.\nAccording to Mr. Stein, it\u2019s critical to understand that AI isn\u2019t magic. It\u2019s just a process for discerning patterns in past behaviors that allow for more accurate future predictions. It can only succeed where a business has a clearly defined problem and a readily understandable metric for success. For example, \u201cWe need to reduce loan defaults measured by a loss rate\u201d or \u201cWe need to increase the conversion rate from its current 32.5% rate\u201d and so on. If you don\u2019t fully understand the problem, you won\u2019t understand the solution either.\n4. Take a hypothesize and test approach as opposed to success or failure\nSince every AI implementation is unique, it\u2019s important to go into every project with a \u201chypothesize and test\u201d mindset as opposed to viewing projects as either outright successes or failures. By making hypotheses at each step and taking the learnings from each step into the next iteration, you can quickly refine your AI deployment until it becomes a workable solution that can deliver meaningful results.\nWhile the hypothesize and test approach will lengthen project deployment times, the benefit is that you\u2019re continually fine-tuning the solution to incorporate real-life lessons learned, to align with customer and employee requirements, and to continually pivot to the most compelling business case that will make your solution sustainable.\u00a0\n5. Incorporate and integrate all forms of automation into your future vision\nAs you embark on initial AI pilots, proof-of-concepts or MVPs, bear in mind that the future vision for your organization with regard to enterprise-wide AI is likely a fusion of multiple types of automation all the way from completely manual processes, to those employing robotic process automation (RPA) to more sophisticated AI. It\u2019s often a case of re-inventing business processes from the ground up and then applying the best tool for the job at each new step. Simply inserting RPA or AI into unchanged existing business processes may well miss the art of the possible. \u00a0\nAnother important factor is the handoffs that occur between each tool. This could be human-to-machine or machine-to-machine. By optimizing the handoffs and making them quick, seamless and reliable you can further enhance your future business processes to be as cost-effective and competitive as your business goals and the market dictates.\nThe good news is that AI implementations can be fast-tracked, but it\u2019s not necessarily about getting smarter about AI. It\u2019s about making the right choices such as build versus buy, becoming obsessed when it comes to data quality (as well as customers), spending sufficient time on change management and getting the business involved early on, taking a \u201chypothesize and test\u201d approach, and ultimately combining multiple automation techniques into your future vision.\nIf your AI project is taking considerable time, be patient and stay the course. You may also be able to leverage some of the recommendations here to help fast-track your race to the finish line. Of course, just like digital transformation, this race is never over.