Reading time: 5:30

Core recommendation

Focus primarily on enterprise business applications that address a strategic business need, built and supported by a leading supplier of that class of application. Three details:

  1. If there isn’t a commercial offering available, rethink your investment timing. Act when you find — in production and thoroughly vetted — real-life commercial examples of what your business needs in uses cases very similar to your own.
  2. Unless you’re in the enterprise business applications business, don’t develop your own or do-it-yourself.
  3. Today’s biggest, most successful application providers may have succeeded by disrupting the previous generation of winners but that justifies looking beyond today’s leaders.


A majority of enterprise AI projects consist of pilots, experiments, and proof-of-concept demonstrations. Enterprises are learning. In turn, that should lead to production investments, but they have been slow to materialize. Many factors are holding back production investment. In this piece, my analysis and advice zero in on technical abstraction levels discussed below.

Most enterprises should focus on AI Technologies at the highest practical level of abstraction and generally stay out of lower-level areas such as developing their own:

  • Specialized hardware
  • AI system services and middleware
  • Deep learning models, tools and frameworks

In some cases, it makes sense for enterprises to add AI application functions (such as language translation and speech-to-text services) into existing enterprise systems. Do it if the interfaces are simple, and the risks and potential level of long-term technical debt are low.

Abstraction in more detail

Einstein famously observed

“Everything Should Be Made as Simple as Possible, But Not Simpler”

We use abstractions to simplify technology. For example, we

    • Separate logic (as in a flow chart or a set of business rules) from the underlying technology implementing the logic.
    • Write programs in higher-level languages that can run on a variety of processors with different machine language instruction sets.
    • Buy network access without specifying the communication medium used by the provider – but we do specify, in our contract, bandwidth and uptime minimums.

To recast Einstein’s words, abstractions make technology as simple as possible, but not simpler.

Here is an AI-related Abstraction Level Model to use. It has six different levels, ranging from hardware at level 1, the lowest level, to enterprise business applications at the top.

Each layer hides much of the complexity of the layers beneath it. Layer 5 is a partial exception because it can apply to all layers, hence the vertical arrow on the right side of the figure.

Start at the top, not the bottom.

If you’re seeking a particular business outcome that exploits AI technology, start at the highest feasible level of abstraction. Every level in the model is essential, but it would be foolish for most enterprises to spread their focus across all levels at once! It would also be prudent to make sure someone else is sweating the small stuff at lower levels of abstraction.

Also examine the model, layer by layer, from the bottom up:
  1. Don’t focus on hardware development – that’s usually someone else’s job. Do know enough to make a judgment on whether someone else is appropriately addressing hardware issues. There are, of course, exceptions to this and all of these rules. Sometimes you have to do-it-yourself (DIY) but that’s generally inappropriate at any level below the highest feasible level of abstraction. DIY is probably inappropriate at all levels for most enterprise needs.
  2. Don’t focus on system services. Same logic. Don’t ignore differences between cloud-based services providers, for example. But do make sure that the system service providers are paying attention to and effectively servicing your needs while delivering on their own specific and unique business needs.
  3. The closer you get to the top of the model, the more attentive you need to be. If there are critical new, disruptive technologies that can make or break your solution, study what your primary and alternative solution providers are using. At this level, look at the frameworks, models, and tools in use and the alternatives that may be available.
  4. At times, application functions from layer 4 might be the highest feasible level of abstraction. For example, your business might benefit from services that automatically make your e-commerce website available in additional languages. In Does Machine Translation Affect International Trade? Evidence from a Large Digital Platform, researchers reported on the impact of applying automatic machine translation on eBay’s web commerce platform. They found large, statistically significant improvements in international trade sales volumes from 17 to 20 percent. Look to adopt application functions that cost little, pose little risk and add real business value.
  5. Level 5, application engineering, development, and management, merits significant investment if you’re constructing your DIY level 6 enterprise application. Be wary of inheriting responsibility for engineering, development, and management for the system across its entire life cycle and watch out for growing technical debt.
  6. Enterprise application investments are justified based on the delivery of significant business value. They are the stuff of major change in how businesses work. They reflect (or transform) the process models of the enterprise. AI technology offers the promise of creating entirely new enterprise application categories. That hasn’t happened yet, at least not at a scale visible in the market.
There is a vacuum in new enterprise application categories driven by advances in ai

Yes, the world now has many virtual assistant packages, chatbots, and programmable smart speaker services. And some of them are quite good (but we’re a long way away from stunningly simple approaches to dealing with complex dialog and user intents.) These aren’t game-changers that deliver significant new business value.

Decades ago, major technology transitions enabled entirely new enterprise application categories to emerge. RDBMS and more robust, inexpensive servers and networking enabled

    • CRM to replace SFA (sales force automation) tools and
    • ERP to replace MRP and similar earlier application categories.

Most of today’s AI is still in the horseless carriage era, applying old analogies (like Business Intelligence and Decision Support; process automation and improved customer experience) to sell new technology. Where’s the vision? Where is this going from an enterprise application categories point-of-view?

Secondary implication

The lack of new enterprise business application categories exploiting AI-technology-driven is exacerbating a shortage of people with AI Technology skills. The skills shortage is, at least in part, an artifact of the failure of a new AI-technology exploiting enterprise application category to emerge. It’s also a product of excess demand stoked by excess hype but that’s the subject of a future research piece.

Models Matter Big Time

Enterprise applications like ERP, CRM, Supply Chain, and HRMS have been embraced because they’re not technologies. They provide model-focused (domain-specific) solutions. Each one has been designed with a model for how their core functions are supposed to operate for businesses and from a business point of view.

There have been major breakthroughs in AI technologies but, outside of what the dozen or so big AI firms (such as Google, Amazon, and Alibaba) are doing for themselves, there are no new model-focused AI-based solutions.

The AI-Platform-as-a-Service (AI-PaaS) providers are competing heavily in Layer 1 through 4. Don’t get confused by these offerings. Learn, experiment, pilot, and exploit where appropriate Layer 4 Application Functions. But most organizations should also seek out the new, AI-enriched enterprise application categories that will emerge.

Three sources to consider

These three haven’t “cracked the code” on an entirely new class of enterprise business applications spawned by the AI technology revolution but they can get you on your way.

  1. Big consulting organizations such as PwC, EY, and Deloitte are eager to help with bespoke solutions but bespoke solution models don’t necessarily scale and I haven’t seen a compelling vision of the next great application category here.
  2. Salesforce, SAP, Oracle, and other major application leaders are adding AI features and functions to their suites. They are enhancing their suites’ capabilities and giving others — like you — the opportunity to build new ground-breaking applications. Tactically exploit some of these capabilities but strategically, avoid any greater DIY temptations.
  3. Visionary entrepreneurs like and are building “Enterprise AI Platforms” — arguably, these firms have more vision but are tiny compared to the other Enterprise application providers and the big consulting organizations. The visionary entrepreneurs have yet to define as products the new, model-centric business-transforming application categories.

Which one is your favorite to bet on?

Next piece of research

My next research piece is going to expose the strengths, weaknesses, opportunities, and threats (SWOT) associated with each of these three alternative sources. It will also include the age-old alternatives “do-it-yourself” and “do nothing now”.

Reminder — prior work

This research piece ties back to my previous post on AI DIY. Most of us did not build our own RDBMS or email systems. Being way ahead of the market is usually a source of long-term failure. Are you still using MySpace? I didn’t think so. Focus on enterprise business applications that deliver the business results you need.

Disclosure: The views and opinions in this analysis are my own and do not represent positions or opinions of The Analyst Syndicate.  Read more on the Disclosure Policy.