AI Sobering

A recent Tweet / LinkedIn post where I used the term AI sobering caught some attention and a request for more details on my thoughts and research. The term ‘sobering’ makes a direct link to the fact that there is apparently a discrepancy between the desired outcome of AI and the actual/perceived outcome.

I started using this term based on the research I’ve done in the last 18 months within Smartfin Capital. Having been in dialog with 150+ international AI companies, accumulated with a lot of desk research, a couple of facts are striking and revealed to me why some AI companies are successful and others not (yet).

The AI market

State of the AI market spend — 2020

Let’s start with some general market dynamics. According to IDC, the global AI spending is growing at a compound annual growth rate (CAGR) of 38% and will reach the $80 billion level by 2022. The total spending on artificial intelligence is expected to hit $35.8 billion this year, up 44% over last year. A couple of interesting facts:

• 2/3rd of the spending is currently happening in the US
• 1/3rd of the spend is currently in Retail and Finance alone
• The number of AI/ML projects per company doubles annually
• ~40% will be spent on applications and software platforms
• >30% will be hardware, mainly compute power

Personally, I don’t really see those last 30% as growth in AI spending itself as it is purely infrastructure growth and not adding any inherent value. I’ll come back to this later when I talk about gross margins.

The AI technology ecosystem

According to Crunchbase, there are over 15.000 AI companies, and more than 20.000 VC’s active in the AI market. You read that right; there are more VC‘s than companies in this domain. Some more detailed info from Pitchbook shows that with one quarter to go in 2020, we’ve already observed 2.665 AI-related deals. That’s 33% less than in 2019 overall but still good for $5 billion capital investment in total, which is already more than last year.

Internal evolution is showing within that absolute stagnation of the last two years. It’s good to see the M&A and IPO part picking up pace. Median deal size has increased heavily since 2017 and this is correlated with growing post-money valuations. It is however still significantly lower than general valuations cross-industry ($7M angel/seed, $20M series-A, $56M series-B).

It’s not all artificial intelligence

It’s striking to learn that merely 40% of these so-called AI companies actually use real AI according to a recent study of the MMC Ventures, featured in the Financial Times. AI is a means to an end, not a goal. If you don’t need AI, don’t use it as it only adds complexity (and cost). That being said, it’s hard to think of an industry that will not be affected by AI in the years to come, much like electricity transformed almost everything 100 years ago.

For me, there is a distinction between companies that use AI and AI companies. It’s not because an ISV writes a wrapper on top of python that you can call it an AI company. In an AI company, it is its DNA, beyond products and services. “It’s a continuously looking for opportunities within the organization to create a better company with higher margins.” (Stijn Meganck, Chief Research Officer Timeseer.AI ).

Expectations vs Reality

The customer

There is a huge difference between companies who are still experimenting with AI versus the ones who are evolving towards a more mature AI practice (O’Reilly — AI adoption in the Enterprise). The main inhibitors shift over time and we find different challenges comparing these stages.

A couple of things that have come to my attention:

• While 15% is still not doing anything with AI, 50% of the companies do claim to evolve towards mature AI practices.
• According to IDC, 50% of all AI projects fail mostly due to data challenges. It has to be said that those with mature AI practices cited “lack of (augmented) data or data quality issues”. This is confirmed by Industry analysts such as Michele Goetz (Principal Analyst AI at Forrester) and Jim Hare (Research VP at Garner). The cost and hassle of preparing and cleaning data come as a shock for a lot of companies, according to Arvind Krishna, SVP at IBM. These related challenges are an important reason for IBM clients halting or canceling their analytics projects.
• Another main inhibitor is that the lack of skilled people remains a major issue. If 80% of your time spent as a data scientist is collecting and cleansing data, you’ll need a lot more people to get a decent amount of challenges solved.
• Today 80% of the AI projects are project-based and are stuck at the Proof of Concept (PoC) level so turn out as not scalable (yet). This will improve over time however Gartner indicates that by 2023 half of application leaders will still struggle to move their AI, ML projects past proof of concept to a production level of maturity.

Failing projects due to data quality, spending way too much time NOT solving actual problems making these projects so expensive in man-days, and lack of scalability are causing what I am referring to as that current state of #AIsobering.

I also notice a #ToolSobering as an aside. Many organizations have invested heavily in AI applications and platforms, but without true ROI, new entrants are being blocked by corporate buyers. Companies today start thinking about scalability before they start new initiatives. It is promising to hear second-generation AI companies will fail forward and embrace all these learnings. But closing the door for them is just not the right strategy. Don’t dwell in a past of disappointing opportunities or you’ll be surpassed by those who are still moving forward!

The startup

In all types of technology companies, there are three main categories: application, platform, or services oriented organizations. All three have a very distinct reasons for existence (usually built in the investment terms and business plan). They do mix and match sometimes, but it is very rare to see a healthy combination before product-market fit! What I have noticed is that a majority of AI companies out there end up as a services-oriented organization, not necessarily by their own choice. A couple of assumptions for reasons are:

• One-off projects based on insights that are not repeatable
• AI efforts that unexpectedly end up not being scalable or even feasible
• Time/Cost to build breaks the balance between TCO and savings
• Not achieving >80% productization.

Two other issues I regularly see are:

• In SaaS you’d typically expect to see 60% to 80% gross margins but In AI this often drops to 40%-60%. This can be attributed to heavier dependency on (cloud) infrastructure costs for storage/compute (data cleaning + [re]training + running) and ongoing support.
• Too many companies focus on serving too many different use cases. By expanding verticals too soon they want to be both the application where the value is and the platform where the scale is. What they fail to understand is that as stated here before, these two types of organizations are very different in funding strategy and returns over time. Applications can become a monopoly faster in a small market and grow organically, while platforms require much more (usually external) funding.

The VC

The point for VC’s is short: Focus enough on real product-market fit and the scalability of your AI investments. Also, be aware of overvaluation. Nobody benefits from overvalued AI companies where there is a huge mismatch between the fair market value (likely an ARR multiple) and the current post-money valuation. Valuations on-promise are only a valid strategy as long as the music doesn’t stop...

Closing

Are we in a bubble? I don’t think so. What we observe today in the AI world is what Gartner describes as the ‘trough of disillusionment’.

All sides (supplier-customer-VC) have their learnings regarding AI and this will eventually lead to the plateau of productivity:

• 2nd generation of scalable product AI companies
• More successful AI investments
• Many more successful AI projects and achieving a full AI maturity

About me: I claim to have a dual-sided view of the analytics market. After my Master's Degree in Computer Science, I worked a couple of years for Covestro and Lanxess and then became an entrepreneur that successfully sold our first AI company in 2018, turned Venture capitalist as a partner at Smartfin Capital.