Artificial Intelligence as a Product — The new SaaS of Software?
It is impossible to productize Artificial Intelligence at the current state of art. Instead the industry must focus on standardizing the process of configuring and training AI software for each client’s needs.
In the world of for-enterprise software, AI services are everywhere — and AI products in short supply. For professionals trying to procure intelligent automation software, finding a product that fits their CIO’s requirements can be a nightmare. There are three reasons why:
- Artificial Intelligence models are only as good as the data on which they are trained. An AI trained for your competitor, for instance, will not work for you.
- Accessing and preparing data is difficult and requires a different approach depending on the security and data infrastructure of each enterprise.
- There are high levels of divergence in both the problems enterprises are trying to solve and the ways in which each wants to solve them.
These are major roadblocks to a workable AI-for-enterprise automation product. But any AI company that really cares about its customers should still be attempting to overcome them.
As we have seen with IBM Watson, AI services often don’t live up to the hype with which they are sold. Vendors often end up committing enterprises to prohibitively-expensive deployment phases of consultation, data preparation and human hand-holding. The promised transformative benefit of artificial intelligence is its ability to more efficiently do things humans do, at scale, without human assistance. If an enterprise must first risk racking up millions in costs gathering, curating and treating data, this core benefit collapses, and one can understand why it would make many CIOs balk at investing in an AI service.
Products, on the other hand, offer far greater accessibility. They put less strain on budgets, because the same code can be sold thousands — if not millions — of times. They are more reliable, because they are already proven and working at thousands of other companies. Unlike a service, their performance is known. A product’s strengths, bugs, pitfalls and fixes are all already cataloged and easily ingested.
Ideally, CIOs want an AI software product that can be configured to the specific needs of their organization. As AI is analytical and data-driven, CIOs want it to customize to their particular context without much effort. Also ideally, they want it to talk seamlessly to other software products on the enterprise shelf.
These requirements underline why there are so few out-of-the-box AI-for-enterprise products in the market. Making any software product that meets all of the above criteria is difficult enough. For AI, the challenges are seriously compounded.
The best AI solutions are context-specific — like an intelligent human, they learn about the environment or system they are attempting to improve, and then use their learnings to make impacts that have a high probability of improving overall efficiency or uncovering new gains.
As such, at the technology’s current state of art, it is near impossible to productize a high-performing AI solution. Products that do exist are trained on a small sample of enterprises and applied to an entire market, with unreliable results.
But just because we can’t productize the final product, it doesn’t mean we can’t bring the cost-effectiveness and reliability of productization to AI software.
In general, while the data for each context varies tremendously, the core code it takes to train, configure and customize an AI is often the same. And the training, configuring and customizing of an AI is often the most costly and time-intensive phase of enterprise implementation.
So it makes perfect sense to productize an AI’s training, configuration and customization.
Like a newborn baby who knows nothing of the world and yet comes into it with all the cognitive tools it needs to make sense of its surroundings, so can an AI be created that contains all the code it needs to adapt to multiple enterprise environments and purposes.
For example, almost every implementation of our natural language search solution goes through the following process:
- Create a new instance
- Upload Search Corpus
- Upload/ Enter Contextual Data
- Upload Known Questions and Answers
- Configure various options
- Launch Training APIs
- Run
After following these steps, it is possible to return an almost-trained AI, a gap analysis identifying key deficient features, and APIs that further train the system to more than 95% accuracy within weeks.
Sometimes we get lucky, and get a good enough AI by training on generic data.
The perfect emulation of human cognitive process comes only when an AI is trained very specifically for the context. However, in some cases training on generic data gives a solution that is a great entry point for an AI technology. It becomes a great way for our clients to start benefiting from power of cognitive automation, while they make up their minds on a complete deployment.
An enterprise must be able to call on specially-trained employees to ensure the implementation of such AI goes smoothly. But in the age of big data and cognitive analytics, the value of investing in an internal team of talented data scientists should already be patently obvious.
The bottom line is CIOs prefer software products to software services. And getting products out the smart way not only satisfies the CIOs — it enables our partners and clients take more control of the process. With AI, however, they must make an exception, because AI as a product is truly a myth.
Adapted from Coseer Blog.
