Why do AI projects fail?

Every business wants to leverage big data and artificial intelligence (AI) initiatives. Lately, many businesses have tried to dive into big data and AI, but only a few have truly reaped its benefits. Though they had the right intentions, the failure occurred far too often.

Here, we categorise common failures and provide guidelines for solving or avoiding them. For context, we present each failure type as a scenario and place these scenarios under three themes: Organisational predicament, Accustomed to doing things, and New ways create new problems.

Originally published at LinkedIn Pulse on 10 November 2019.

Photo credit: www.pexels.com

Organisational predicament

The never-ending queue of relentless requests

A lead data scientist becomes the Chief Analytics Officer (CAO) of a large enterprise. These are some of the first requests to reach his office:

• The Chief Marketing Officer wants deeper customer understanding to inform marketing spends and drive revenue through personalisation.

• The Chief Operations Officer wants to benchmark operations against competitors and build dashboards to optimise efficiency.

• The Chief Information Officer wants to reduce costs by leveraging artificial intelligence for production monitoring and technology operations.

• The Chief Risk Officer wants big data to address regulatory and risk management gaps.

• Trading Managers want solutions for portfolio optimisation, automated research, and more.

The solution: As big data analytics can address a range of problems across the enterprise, they should create a steering group, headed by the CAO. This group can channel, prioritise, filter and sequence requests.

Functional silos

The IT team at a retail company is horizontally structured into tribes for program management, business analysis, user experience, deployment, and database management. Each tribe has a leader and its own culture. Awareness and interest in big data lead to three additional tribes: data science, data engineering, and data visualisation.

When they develop a high-profile big data project, they’ve assembled a Centre of Excellence (CoE) with a representative from each IT tribe. But the members of CoE did not get out of their functional siloes, delivering a suboptimal product.

To avoid this: Cross-functional teams should be given collective responsibility for goals, with strong alignment between data scientists, data engineers, data visualisers, and business analysts on products being developed.

A disconnection between business and IT

The IT arm of a retail company sponsors a data science group that quickly becomes an independent subunit. But business continues to view it as an ‘IT project’ and does not intervene. Six months go into building sophisticated models for demand forecasting and inventory management.

During a demo, business users ask, “Is the new model really necessary? The old model does what we want, at least for now.” But given the investment had already been made, the IT lead produced the proposed model. A year later, at launch, business users refused the product.

To avoid this: Leaders need to understand that technology can be a business enabler only by taking a continuous approach that meets business needs.

Careless positioning statements

The same retail company was probably dealing with ill-informed assumptions. Quite often, big data analytics tools are expected to provide insights on trading recommendations, process, asset recommendations, etc. This crosses paths with what senior business analysts and functional experts work on.

To avoid this: Organisations should establish Data Science as a force multiplier rather than a threat among their employees. This helps create the right human-machine symbiosis and interactions.

Accustomed to doing things

Huge upfront investment

The Chief Technology Officer (CTO) of a product company allocates 50% of his hardware budget to commercial technologies. In a year, the company onboards 10 new customers, and the data volume increases by a factor of 100. However, the company’s infrastructure supports only 10% of their use cases, and the CTO is caught between a rock and a hard place.

To avoid this: Technology choices are often made without keeping in mind realistic business use cases and time for sufficient experimentation. The big data space is also fragmented, with multiple tools and platforms to carefully choose from. Bugs also need to be ironed out before these platforms are ready for production.

New ways create new problems

Using the wrong tools for the job

An airline company wants to use predictive analytics for effective fleet management. The airline wants to aggregate flight data, part specifications, and flight schedules. The information will predict which parts need servicing or replacement before each flight, reducing maintenance time.

The analytics team uses brute force to integrate the accumulated data into the legacy data warehousing and business intelligence infrastructure. The job takes several days to run and is obviously not scalable.

To avoid this: Leaders need to understand that it is not enough to re-label an organisation’s data warehouse and business intelligence investments to tools meant to solve a specific set of problems. A deeper data mindset is needed when expanding organisational capabilities, such as predictive analytics.

Building organisational trust in data products and algorithms

A gas company wants big data analytics to provide insights into and optimise its pipeline and storage assets. Their data scientists develop an innovative algorithm that pulls in all pipeline and storage capacity, consumption details, and demand and supply scenarios. Using this information, the algorithm derives recommendations. But the business leaders do not trust the algorithm.

To avoid this: Analytics leaders need to build trust. Building trust is important because the initial time should be invested in helping leadership and key teams understand the scope, potential investments, and running costs of big data products. The analytics team needs to build trust through detailed, creative data visualisations. This requires pre-planning and proper instrumentation of the code and pipeline to capture the correct intermediate snapshots. This will ensure less time being defensive and more time being innovative.

Concluding remarks

AI and analytics are a high-impact area. But there are risks that can be smartly mitigated as businesses embark on their big data journeys. It all starts with answering a few pertinent questions and visualising a few forthcoming scenarios to get the best out of high-stakes big data analytics initiatives.