A significant decision along the path from interesting to useful is where AI technology’s application hits the industry verticalization challenge. Lawyers want AI specific to their legal practice; healthcare insurance providers need AI that understands the complexity and details of healthcare insurance etc. For a technology vendor, the need to meet the requirements of a specific industry is weighed against both the difficulty and expense of taking that vertical route.
For emerging technologies, the early-stage heavy lifting is usually shared between the software vendor and an early adopter. The trade-off is simple; it allows a vendor to understand whether this or that industry vertical is suitable for productization without making a massive commitment. For the early adopting organization, the value of a potential competitive advantage is often worth the cost of the exercise. It’s an open secret in the software industry that the cost of delivering any significant piece of functionality is borne by the very first customer that sees it as a “must have,” and in this respect, industry verticalization is no different. Yet AI itself is different from traditional enterprise business application technologies. It faces unique challenges when it comes to industry verticalization.
Challenge Number 1 – New Data
There’s a point in developing any new business application where it makes contact with the reality of everyday work, and that’s no different when considering AI. Even a well-constructed corpus of test data used during the development of an AI-driven application will hit the real world’s messy and messed up reality.
That’s true whether it’s built from well-established, open source datasets (perhaps for benchmarking) or supplied from the archives of a friendly partner (such as a keen early adopting customer). In the first case, the data is likely both “wide and shallow” to provide a wide range of test scenarios but lacks the precision required. In the second case, the dataset may be more representative of the real world but will still heavily skew toward the specifics of one customer in one language (further skewed by organizational lexicography). It’s why the Eiffel Tower still appears in almost every computer vision demonstration, but not hundreds of the same piece of product packaging; a more representative “real world” use case. While both sorts of datasets are essential, the type, volume, and velocity of change of the data used are all vitally important to planning and defending your verticalization approach.
Challenge Number 2 – Regulation
All industry vertical AI business applications will need to adhere to regulations – if not regulatory approval and formal certification – for customers to be able to use them at all. Regulatory constraints are not unique to AI; the same barriers to entry exist for any information management technology; however, with AI-driven applications, the impact of regulations can be more keenly felt.
Whatever industry vertical, regardless of whether it’s traditionally thought of as governed by stringent regulatory bodies, meeting regulation is a minimum cost of entry for any AI-driven business application. Depending on the regulatory frameworks involved, a proven failure of adherence can result in penalties per instance rather than per incident. The worst-case scenarios are potentially horrendous for AI systems that work without supervision, processing significant volumes of information. Such potential horror shows are not reserved for traditionally highly-regulated industries; for example, GDPR fines are notoriously punitive. Though not every vendor is all that concerned about regulatory compliance, as in almost every case, the regulatory requirements are held by the end customer, not the technology supplier; there’s no real way to offshore regulatory requirements to 3rd parties.
Challenge Number 3: Localization
While it’s not a subject exclusive to AI-driven business applications, localization – ensuring that such systems meet the specific needs of geographies or sub-geographies – cannot be avoided. Basic entry costs such as user interface translation are a given – if not always easily achieved – however, many more subtle cultural practices need to be adjusted and addressed. For language-based AI systems, words within the same language groups might have distinctly different meanings depending on their use. This is the case with the English language family and its regional variants, but it is the case with all languages. This means that generic “out-of-the-box” configurations must be reviewed with local knowledge before use, as these are likely to be a higher risk factor than those created within local markets. Back on regulation again, while many regulatory environments operate on a pan-national basis (for example, GDPR). Such regulations also often act in addition to, or as a minimum standard, at the national level.
Challenge Number 4: The Language of Work
Suppose your AI business application is designed to deal with language processing. In that case, it must be appropriately configured for the tasks it is applied to in an industry verticalized operating environment. Taking Natural Language Processing (NLP) as an example, ensuring it runs successfully will likely require a serial set of adjustments. Making sure that the “Part of Speech Tagger” which breaks down the language into its constituent parts, can deal with the shape of the material – the language construction rather than format – consistently. Medical patient notes, for example, are likely structurally different from a similarly sized film review. As such, the most salient features might be missed downstream – for example, using a summarization processor – if the upstream processes miss the nuances.
Language shape is also allied to the specific lexicon of the industry (and the language, plus its language variant), creating a complex set of circumstances that need addressing; the challenges here are substantial.
Translating value to business returns
AI is extremely complex and requires highly skilled workers to get it off the ground, but that gets obfuscated in marketing pitches that shout out, “Ta Da! – isn’t AI can do everything, it’s like magic!”. At a high level, the challenges posed by verticalization to meet a specific set of industry needs are no different for AI than for any other business application. However, because of the twin elements of complexity and overhyped expectations, AI has the potential to suffer more greatly from those challenges.
AI Challenges & Opportunities?
AI-driven business applications are the future, and at Deep Analysis, we research and advise our clients on their use. Still, they are not magic and face considerable challenges to adoption. Less cynically, we can say that AI-driven business applications promise to meet the challenges at ‘moments of action’ to bring clarity and immediacy when essential business decisions need making. Yet no organization can use a generic tool, be it a business or productivity application, in a generic way. In reality, the organization overlays its particular requirements through direct human interaction with the system. There is a harsh limit to business activities that can automate to 100% of efficiency and accuracy. Building these applications to be specific to vertical industry needs is essential.
Still, one should never underestimate the scale of the work involved in taking a generic AI product that can impress in a demo to one that can genuinely deliver on its promise in an authentic and messy working environment.
If you are a buyer or user of Business Applications and wondering how to navigate these difficult times, we have a free – yup, free – confidential advisory service for you. A service that means all you have to do is find a time on our calendar.
If you are a seller of Business Applications; then, like many others. In that case, you may reassess your current business, pricing, or go-to-market strategies to adapt to these turbulent times. Again feel free to reach out and chat with us as we may be able to help by providing some critical but constructive advice and support.