Enterprise machine intelligence (MI) transformation: five big pitfalls, and how to avoid them

Many insurers use the phrase "digital transformation" to describe their use of new technologies to increase their scalability and agility, personalise customer experience, increase ease of regulatory compliance, and remove internal organisational silos and frictional costs.

Our latest sigma research examines the transformational opportunity that machine learning (ML) and artificial intelligence (AI) offer to achieve such goals today.

However, delivering this transformation is not always straightforward. Deployment can often fail – frequently due to organisational constraints such as legacy systems and poor data rather than difficulties with the models used.

These roadblocks can lead to repeated cycles of investment that fail to deliver the change required. In an overview of survey data, we found that less than 10% of firms in all sectors have managed to scale MI pilots for rollout across multiple processes.

What are the biggest pitfalls that lead to failure in enterprise-wide MI rollout? We look at our top five, below.

Pitfall 1: Costs

In a recent survey, 93% of respondents at US insurers going through MI transformation expressed concern around the costs of implementation and return on investment (ROI). The reason is usually that cost/benefit analysis focuses on direct expenses (such as the price of the vendor solution) and misses significant, often recurring, indirect costs.

An honest assessment of ROI must set the benefits of workflow transformation, measured in lower costs, higher revenue and new business opportunities, against both direct and indirect costs. Integrating a new MI-enabled system into an organisation will require workflow process re-engineering, and this will almost always constitute most of system deployment costs. Maintaining the integrity, security and privacy of a new system will also require large budget at first, though it may decrease over time, and the pay-off would be increased efficiencies and decreased business costs in the long-run.

Pitfall 2: Absence of a firm-wide data strategy

A successful enterprise-wide MI implementation requires a holistic, fully enterprise-wide data strategy and architecture. Yet most companies do not develop one before attempting to roll out digital transformation. In one recent survey, less than 10% of Chief Data Officers (CDOs) across industries said they were able to measure the financial value of their information and data assets. Another survey found that as many as 75% of insurers lack the ability to harmonise different types of data.

In addition, we've observed that in many companies there is no centralised curation process, leading to a duplication of data engineering by several teams.

Pitfall 3: Too few data engineers

It may seem like a simple error, but often deployment fails due to insufficient hiring of data engineers. Typically firms often start by developing an algorithm, but they then make the mistake of under-investing in data engineering.

Notably, in an end-to-end enterprise process, a sophisticated algorithm backed by poor-quality data tends to actually underperform a weaker algorithm supported by high-quality data. Therefore, these firms are making a crucial investment error.  International Data Corporation research found that data professionals spend 67% of their time searching for and curating data.1Investing in engineers to improve data quality can reap significant long-term benefits.

Pitfall 4: Lack of alignment of MI to the business need

Crucially, business use case and data availability should determine what sort of system, and algorithm, is selected. Much of the technology is so new there is little consensus over which methods or models best serve which use cases, but we expect consensus to emerge over time. For now, even with best efforts to match the technique to use cases, a trial and error process is normal to determine which works best. But the process should be led by business need –many failed implementations arise from a mismatch of algorithm to use case.

Pitfall 5: Lack of senior-level sponsorship

Absence of management buy-in can be deadly for large transformative deployments of MI, particularly in times of leadership change. The model/algorithm development itself is only one element of success: equally critical is a detailed, cross-functional strategy that has both senior executive sponsorship and buy-in from business units.

Reasons that a project may be pulled can include senior management being not sufficiently briefed on the proposed MI deployment; poor implementation by frontline staff; poor coordination among business units; and shallow.


 1. End-User Survey Results: Deployment and Data Intelligence in 2019, IDC, November 2019.


drivingdigitalinsurancesolutions insurancesolutions insurance sigma ai bigdata digitalecosystems


May also be of interest to you

  • Blog ​Behavioural economics: Predicting the irrational in a data-driven world

    Christoph Nabholz Head Life & Behaviour R&D