The Increasing Accessibility of AI in Business (AI Zeitgeist 4 of 7)

This article is part 4 of a 7-part series called “AI Zeitgeist,” where we’ll be mapping out the details of AI adoption over the next 10 years and explore the critical changes in the AI ecosystem that business leaders need to understand.

In this installment of the Zeitgeist series, we’ll be talking about how artificial intelligence technologies will become more accessible to businesses and non-technical employees.

Today, if you don’t have a strong data science background, you are very unlikely to leverage powerful AI vendor tools. In five years, many vendor tools will require almost no formal data science training, and that transition is something that very few business leaders consider.

AI accessibility will evolve as we move from the “Emergence” phase, where AI remains a wizard-skill that’s inaccessible to many companies, to the “Dispersion” phase, where AI becomes ubiquitous and simple to use across almost all business sectors.

(Note: If you haven’t read the first article in the AI Zeitgeist series, you should familiarize yourself with it first. In that article, I explain what the 3 phases of “Emergence,” “Adoption,” and “Dispersion” actually mean).

AI tools come in many varieties of complexity, so in each phase of the AI Zeitgeist (“Emergence,” “Adoption,” and “Dispersion”), I’ll be referring to the following three rough categories of AI solutions:

1. In-house Solutions: Building AI or machine learning solutions from scratch. I refer to this colloquially as “doing artificial intelligence.” Companies who build their own solutions need to understand the business context of their problem, as well as the right combination of data assets, hardware, software, and algorithms to make the solution work. Most companies have nowhere near the amount of in-house talent to do this.
2. Vendor Solutions: Artificial intelligence vendor companies generally develop solutions specific to particular verticals (such as banking or pharma) or horizontal applications (such as customer service or marketing). These pre-built solutions often require some integration with a client’s own software systems and data assets, and they also generally require upkeep. 
3. Tools and Scripts: Simple lines of code or APIs licensed from large AI providers (Microsoft, Amazon). While it is not possible to hire a data science-savvy freelancer to build a robust AI solution from scratch or to properly manage or integrate a complex AI vendor solution, it is possible for remote contracted workers to set up a simple sentiment analysis API from Amazon.

We’ll begin by exploring the accessibility of AI in the current phase of AI adoption: the “Emergence” phase:

Phase 1: Emergence

1 – In-house Solutions

Companies that aim to build AI solutions in-house (for their own teams, or for a product which clients will use) require resources and capabilities that are rare indeed, including:

• Massive reams of proprietary, valuable data on which to train algorithms
• Data infrastructure, data standards, and data storage that will permit staff to access and use said data
• Subject-matter experts to work with the data science talent to help them define, refine, and strike directly at the business problem (for example: Improving email marketing revenue, automating complex paperwork, whatever)
• Enough data science talent to potentially build a solution and translate a business problem into an AI approach (with the courage to admit what can’t be done)
• The ability to endure the risk of an AI project not yielding results, or costing much more than estimated
• (Arguably most importantly) A corporate culture that accepts the long periods of data handling, iteration, and research required to bring AI to life in a company

The number of companies with those ingredients is remarkably small, and while many other companies may deceive themselves into believing that they should be innovating and developing their own in-house AI applications, it will remain a small upper echelon that is able to genuinely develop novel applications.

While machine learning algorithms and statistical methods have been around for well over a decade, turning those algorithms into business value is still (mostly) the Wild West. Applying machine learning to business problems is hard; building solutions from scratch is often just as hard.

Our own poll of 30 AI researchers points out the biggest barriers to getting an ROI on AI in business:

Make no mistake about it, in order to build in-house AI solutions, all three of these issues will need to be addressed, and most companies (including public companies) have neither the right data nor the right talent to build solutions in-house. I’ve written much more about this in our enterprise AI adoption article from last year.

2 – Vendor Solutions

Warning – we generally find that 66% of companies who claim to be offering an AI solution are in fact not “doing AI” at all, and many such companies have no in-house data science talent whatsoever. Our article on “avoiding AI hype” will help you get a sense of how to tell the real from the fake.

We’ve never been advocates for buying from companies using AI simply because AI is interesting. Ultimately a B2B purchase should be about achieving company goals, and if that gets done with boring “old school” means, so be it. That being said, we dislike it when companies use the AI buzz to get more attention than they deserve, and we’ve previously shared some of our tips for avoiding companies who are lying about AI use.

While there is a gradient of difficulty for vendor solutions, most of today’s present solutions require the following resources on the part of the buying company:

• Tens of thousands or hundreds of thousands of dollars (often even pilot programs require sums of this kind)
• Some vague semblance of data infrastructure or at least a robust way to harmonize existing data streams to be useable for training algorithms
• Some degree of in-house data science staff to understand the data and integration needs of the vendor company
• In-house subject-matter experts who can understand some of the lingo of data science, and operationalize a specific problem
• Months and months of effort for data cleaning and iteration to ensure that the model works and application is delivering meaningful results
• Some vendors make data collection somewhat simple (for example, by working only with established APIs, such as the API of Zendesk or some other established enterprise vendor), but for many vendors, it’s a grind to just get their hands on the data to begin work (with no guaranteed timeframe for producing results, as machine learning requires experimentation and iteration over time)

Does that mean that AI vendor solutions aren’t worth working with?

Of course not, but it’s no surprise that the majority of AI vendors are targeting enterprise clients, who are likely to have the budget and data volumes to at least get started. Those behemoths will be the petri dishes in which AI solutions will be tested and worked out before they ever make it en masse to the mid-market.

If you just read press releases and AI vendor company communication (they’d sure like it if those were the only sources you paid attention to!), you’d believe that most vendor solutions are “push-button” east, and integration is a snap. Not so. Not so at all.

Because many businesses have no idea of the data requirements, talent requirements, and time requirements of using AI vendor solutions, they shop around when in fact they shouldn’t. Often they realize that they’re in over their head before buying. Sometimes they buy because AI is “cool” and they want to feel hip and modern (we call these “toy applications”, and we generally frown on this kind of decision-making).

It will be years before even enterprises realize the demands that working with AI vendors requires, and so the majority of vendors will have to educate enterprise buyers and teams from scratch – not just about their solution – but about what artificial intelligence is in the first place. This ain’t changing any time soon.

Also, it should be understood that many vendor companies, even those who have raised $30MM-$80MM, are still figuring out their product, their ideal customer, and their value proposition. They’ve raised money by convincing investors they’ve figured it out, but in many cases, they’re still “feeling out” what problem they solve and for who.

What this means for buyers is that they are often the guinea pig for a solution that’s only ever been integrated at 2-4 other companies ever, with no real robust evidence of ROI as of yet. Again, this doesn’t mean that vendors of this kind shouldn’t be worked with at all; it’s just important to understand the situation you’re getting involved in, and the risks and rewards involved.

3 – Tools and Scripts

There are an increasingly large number of AI tools which are available for free, or for a relatively low cost – without requiring data science staff (at all!), or existing data assets at all.

Google offers APIs for image recognition, speech recognition, and more via Google Cloud.

Microsoft does the same with their Azure Cognitive Services.

Amazon does the same with AWS Machine Learning.

There are basic tools and scripts of this kind that simply require a relatively average level of development talent to simply set up the system itself. The

Currently, simple APIs from providers like Google or Amazon (or a whole ecosystem of smaller players) offer relatively braod AI capabilities. These capabilities might not be hyper-tailored to a specific business use-case, but they apply widely across a variety of businesses and so might be useful.

For example, today’s simple computer vision APIs can help with the following broad tasks:

• Create a description of an image
• Search for similar images
• Identify celebrities or famous places
• Identify faces and their genders/ages
• Extract text from images (signs, handwritten letters, etc.)
• Find similar images on the web
• Spam filtering/content moderation (example: adult content)
• Logo detection

Today’s natural language processing APIs can help with the following broad tasks:

• Find similar text on the web (plagiarism/copying)
• Spam filtering / content moderation
• Sentiment analysis
• Text summarization
• Entity extraction (people, location, companies – Google has done this via API)
• Search (eCommerce or content)

In the future, we might expect more custom capabilities to develop, allowing businesses to tailor specific AI applications just to their own particular needs.

Phase 2: Adoption

1 – In-house Solutions

In the “Adoption” phase, “doing” artificial intelligence (building a robust AI solution in-house) will be significantly less burdensome than it is today.

We can look at the various required factors of “Doing AI” in the “Emergence” phase, and see how those factors might change in the “Adoption” phase:

• Massive reams of proprietary, valuable data upon which to train algorithms: For many applications, machine learning will require less data in order to still garner results. While this might take years of development in the computer science, there will also be more efficient algorithms for discrete tasks (NLP, vision applications, anomaly detection, etc), which will likely require less data than today’s applications. This isn’t to say that companies won’t need large volumes of data, but in some cases, that burden will be lessened.
• Data infrastructure, data standards, and data storage that will permit staff to access and use said data: This will certainly still be an issue, but over time (as the value and treatment of “data” become common knowledge in enterprises and mid-market companies), data will naturally be more and more accessible as the years go on – particularly in industries where the value of data is relatively well understood, and industry pressure exists to modernize data systems (pharma, insurance, eCommerce, etc).
• Subject-matter experts to work with the data science talent to help them define, refine, and strike directly at the business problem (for example: Improving email marketing revenue, automating complex paperwork, whatever): Over the next 2-3 years, it will be easier to find different subject-matter experts who can “talk data” and understand the value of data and the basic functions of AI. When the business problem experts understand the basics of how AI works, and what problems it can be applied to – entire teams can move much faster than they often do today (when data scientists must operate as the only “wizards” of data and AI knowledge in their entire department).
• Enough data science talent to potentially build a solution, and translate a business problem into an AI approach (with the courage to admit what can’t be done): In-house data science talent, and
• The ability to endure the risk of an AI project not yielding results, or costing much more than estimated: (See the “experimentation” point below.)
• (Arguably most importantly) A corporate culture that accepts the long periods of data handling, iteration, and research required to bring AI to life in a company: (See the “experimentation” point below.)

Experimentation: I believe that the most important factor in making AI easier to “do” is the number of actual applications that are developed, and the evolutionary process of trial and error.

As more companies build AI solutions, they’ll be forced to think seriously about cleaning their data – and maybe even about storing it in more effective ways to future AI applications. They’ll learn how many team members of what kinds need to be involved in various kinds of AI applications.

They’ll have a sense of how long these projects take – and once they see some tangible “wins.” They’ll be willing to endure the time and investment in iteratively improving an AI application.

In addition, companies who develop AI solutions will be providing more and more of their staff (programmers, data scientists, subject-matter experts) to the process of building AI solutions, and as these employees leave to join or start other companies, they’ll be able to pollinate those firms with a better sense of “best practices.”

This experimental process takes time, but even in just two years we’ll be drastically farther ahead than we are today, and building solutions will be easier.

2 – Vendor Solutions

Good news: In the “Adoption” phase, vendor solutions will not only be easier to integrate and easier to use, but also more accessible by mid-market clients with less data, less data science talent and less budget to work with.

Why?

Because capitalism, that’s why.

Vendor companies are fighting vigorously in the market in order to win market share and win deals. “Winning” in the market implies (among other things):

• Sharpening their value proposition, making it easier to understand, and more able to achieve
• Bringing down the barrier to entry to deliver results for clients, including:
• Reducing the complexity of integration
• Reducing the requirements and burden of client companies
• Reducing the technical complexity of upkeep for clients

As more buyer companies understand data and AI, and as more vendors understand their value proposition and their exact offering, more “wins” will come about from vendor solutions. These “wins” (positive case studies) will help to solve the chicken and egg problem:

Without case studies, companies don’t feel good about adopting a product. Without selling a product, no good case studies can be developed.

We’ve seen this dynamic most clearly in the intersection of machine learning in healthcare, but it exists in essentially all sectors. Some fields – like eCommerce – will have an easier time with this case study traction than other stodgy industry – like healthcare – but they’ll all be chipped away over the years ahead.

In the “Adoption” phase, a huge bulk of the AI solutions in any given sector (from banking to mining and beyond) will have established case studies and a clear onboarding process – and buyers will have a more intuitive sense of AI and data, and a better sense of expectations about what a vendor solution can do.

3 – Tools and Scripts

Tools and scripts will expand to more and more domains of AI capability and functionality.

While intricately customized needs (i.e. A lead scoring system customized for one specific business, or a recommendation engine for a very novel online jewelry store catering to women in India) will not have a “plug and play” solution to their problems, there will be an increasingly broad set of services available at low prices.

In addition, we can expect a wider and wider pool of API providers for AI capabilites (like NLP or computer vision) and wider and wider use of simple scripts (like logistical regression) in various coding languages for simple applications.

Phase 3: Dispersion

1 – In-house Solutions

Any and all established enterprises in the “Dispersion” phase (that aren’t on the verge of dying) will have a strong understanding of:

• The value of data, and best practices for storing and organizing it
• The general low-hanging fruit opportunities for AI and machine learning
• Any and all established enterprises in the “Dispersion” phase (that aren’t on the verge of dying) will have the following capabilities:
• Reasonable well established data infrastructures, making most useful data available for important business applications and training algorithms
• Reasonably capable data science staff (including ML engineers) who are able to communicate reasonably well with the business unit they usually work with (example: Customer service, or marketing)
• The non-technical staff which understands basic concepts of AI, and can speak in the language of data with data science staff and AI vendors

Think about “the internet” today.

Not every company has a masterful internet strategy – but any global company that’s alive today at least has enough capability to get by. AI will be much the same in the “Dispersion” phase. “Doing AI” will be synonymous with “doing business” for most large and even mid-market firms, and certainly any capable venture-backed company.

While the “Dispersion” phase may not bring about the ability for provincial small businesses to innovate in AI – these smaller and less technical players will be able to use vendor solutions.

2 – Vendor Solutions

Think about it this way.

At some point in the past, say 1996, the idea of “marketing automation” was extremely complicated and foreign. Sending email messages based on actions and segments of an email list, and automatically tagging, organizing, and communicating with contacts based on their actions – that would have sounded like magic.

In 1996, it probably was magic, it was “wizard skills” to be able to do something of that kind. Only enterprises could purchase and use such complicated, novel solutions.

Now, any Joe or Mary who decides to open a business can open a MailChimp account for free and set up email automation and email segments in minutes, in a simple, intuitive interface.

What happened? Time and capitalism.

Companies fight to find easier interfaces, simpler functionality, lower prices, lower barriers of use – and eventually “wizard skills” become accessible even to small businesses.

By the “Dispersion” phase, AI will be synonymous with almost all enterprise software offerings, and many small and mid-size business software offerings, too. It’ll just be software. Users will think: “All software is smart – what’s the big deal?”

3 – Tools and Scripts

Just as online contractors can set up WordPress Plugins or other simple HTML scripts from the internet, eventually all commoditized AI capabilities (text, speech, vision, and more) will be reasonably accessible to anyone with basic technical skills, no formal data science training needed.

For example:

• Businesses looking to programmatically change the look and feel of their website on the fly may be able to adjust their content and product offerings based on the activity of online users.
• Companies looking to solve unique computer vision problems (say, detecting errors in fabrics or carpet) will be able to train algorithms to detect errors using a simple interface where they can upload “good” and “bad” examples of the fabric in order to train the vision system quickly.

Concluding Thoughts

This has been week 4 of 7 from the “AI Zeitgeist” article series.

In the coming weeks, we’ll be exploring the following Zeitgeist topics, in order:

1. The 3 Phases of AI in the Enterprise: Emergence, Adoption, and Dispersion (AI Zeitgeist 1)
2. How “AI” Will be Discussed in the Future (AI Zeitgeist 2)
3. The Evolution of AI Talent and Training (AI Zeitgeist 3)
4. The Increased Accessibility of AI in Business (AI Zeitgeist 4) <— You are here
5. Buying and Adoption Readiness for AI (AI Zeitgeist 5)
6. The Changing Landscape of AI Priorities of Business Leaders (AI Zeitgeist 6)
7. The Competitive Dynamics of AI – Now and in the Future (AI Zeitgeist 7)

Next week we’ll explore how the purchasing and selling of AI solutions will change in the years ahead (Preview: The way AI is sold and marketed today is quite different from how it will be in the next five years – and in general this will make AI more accessible to business leaders, and easier to understand).

Stay tuned and see you next week.