The cost of building AI: Why partnering is the better choice
In the past, tech investments depreciated in about 10 years. For today's AI solutions, it's more like a year or two.
As organizations rush to take full advantage of AI technology, many are tempted to custom build these applications to address unique business requirements and create their own “secret sauce” in the process.
For large companies with healthy IT budgets and top-notch people, this doesn’t seem like a stretch. But there are substantial risks, several of which are unique to the AI domain.
Related: AI integration propels insurtech sector despite funding slump
As with most IT projects, building AI applications is harder than it might seem initially. In this case, though, some of the biggest risks will not be apparent until fairly late in the process, when a company is already heavily invested in a sizable, expensive AI initiative. Even for large, well-funded organizations, buying AI solutions is generally a far better option than building them in-house.
Here’s why.
Insufficient data
The first significant barrier to building effective AI models in-house is the lack of sufficient data. This is true even for most large, established organizations.
A top-ten insurance carrier, for example, might have a wealth of information on specific types of claims or litigation, but they could still be lacking in data for certain categories of injuries or for specific geographies.
AI models need that level of detail in order to train properly. Most companies simply don’t have enough data to develop models that can deliver meaningful results.
Moreover, maintaining accurate models requires continuous data updates. Consider how social inflation has impacted verdicts throughout the United States. Inflated claims are a widespread problem, but juries in some areas are more likely to award extravagant sums of money. Machine learning models cannot draw accurate conclusions with piecemeal or stale data. If your data is both fragmented and outdated, your results will be further distorted.
Buying third-party data is an option, but it’s an incredibly expensive one. The data itself costs a great deal; then it takes a small army of people to cleanse, harmonize and refresh that data. Total costs using this approach will far exceed those of an all-inclusive commercial AI platform.
Building AI solutions from scratch is resource-intensive and a significant drain on finances. The initial investment combined with the costs of maintaining, updating and securing the models generally makes the in-house approach prohibitively expensive.
Scalability, bias and security
When it comes to AI, there are several other factors that are often overlooked: scalability, safeguarding AI models from bias, and securing the data.
AI is extraordinarily computationally intensive. While cloud computing offers substantial benefits in this respect, there is still a broad tendency to underestimate the challenges of scaling AI models effectively.
Ensuring that models remain free from bias is also a complex, ongoing task. The legal and regulatory environment governing bias in AI is still evolving, but this topic will undoubtedly get more attention as real-world examples emerge. We have already seen several examples in the headlines, including a gender bias case in which Apple and Goldman Sachs were credibly accused of offering lower credit limits to women due to a flawed approach to AI.
Finally, companies must secure the data they’re using to drive their AI models. That includes careful vetting and ongoing monitoring of any third-party vendors that provide components for your homegrown AI solution. This means continuing costs as well as new risks for any organization that chooses to do it themselves. Established AI service providers are equipped to handle security, scalability and bias more efficiently.
AI’s exponential growth trajectory
This brings us to a third category of challenges for companies that take the DIY approach to AI. By the time they have brought their project to fruition, it’s already obsolete.
AI technology is progressing at an unprecedented rate. Advances in large language models (LLMs) and other AI technologies are happening so fast that building an in-house solution actually puts an organization behind the innovation curve. In the past, technology investments had a depreciation period of about 10 years. For today’s AI solutions, that number has shrunk to just one or two years.
This rapid pace of innovation makes it nearly impossible for companies to maintain cutting-edge AI models internally. Renewing that initial investment every two years is cost-prohibitive, extraordinarily disruptive, and it introduces perpetual project risk to an IT component that will soon be critical for maintaining competitive parity.
The market evolution: AI service providers
Even AI companies themselves do not build everything from scratch. They leverage open-source technologies and contributory databases to stay ahead of the game. For insurance companies and other businesses, partnering with AI service providers who have access to vast datasets and advanced technologies ensures better outcomes.
Look for a provider that has 10-times or more access to data than what you currently maintain internally. If an AI company is relatively new to the game and not well-established in the market, it’s likely that their datasets will fall short of the mark. This won’t necessarily be apparent at the outset, but it has significant ramifications for the overall effectiveness of your AI initiative.
How fast do you need it?
AI is here, and for most organizations there is a sense of urgency to move forward. Today, AI presents an opportunity to forge ahead of the competition by implementing better technology, but that window of opportunity will eventually close. For insurers, AI delivers more accurate risk assessment and more efficient claims management.
Do you want that today, or can it wait two years?
With respect to competitive strength, the build vs. buy decision has significant implications, especially over the next three to five years. Companies that choose to wait, or which take 12 to 24 months to build their own solution, will necessarily defer those benefits.
The build vs. buy debate is not new, but when it comes to AI, the calculus has changed. The unique characteristics of AI make this a fundamentally different decision. The need for comprehensive datasets, the risks associated with bias and data security, and the lightning-fast pace of innovation make building AI solutions in-house a daunting and often impractical task. Working with established AI service providers is generally a superior choice. Partnering not only ensures access to cutting-edge technology and vast datasets but also provides scalable, secure and bias-free solutions.
Chad Langford is vice president of Data Science at CLARA Analytics.
Opinions are the author’s own. This article is published with permission and may not be reproduced.
See also: