What type of applied AI product should be built?

While building a semantic search engine, I understood that there are no narrow AI standalone products. Instead, AI is always a tool to do something.

A feature.

With this in mind, I researched what companies do when they conceive implementing AI-based features into their products.

As a result, I believe that there is a need for a new low-code ‘construction kit for the AI’ type of product.

What is the current state of applied AI?

According to a data scientist, Peter Skomoroch, machine learning transition will be around 100x more challenging than mobile. But every enterprise company needs to do it.

Simultaneously, Deloitte’s state of AI in the enterprise survey stated that there is currently a steep learning curve with most companies reporting skill gaps in this transition.

At the same time, Algorithmia’s 2020 state of enterprise machine learning report described how at this point, there is a too-long road to deployment of any machine learning related features due to it taking 31–90 days per model, on average. They also noted that with lots of computational resources required, it is too expensive for most enterprises to train and host their own models. Meanwhile, managers have recognised that the problems only become worse with scale, so it is a hard sell even to start developing anything in-house.

On the ground of all this, I understood that

• 55% of companies surveyed have not deployed any ML models at their companies
• 68% of executives report moderate-to-extreme skills gap to integrate AI
• 70% of data scientists time spent on deploying new ML models
• And 47% of deployed ML models fail to go into production

In short, the current state of integrating AI-based features by enterprise companies is terrible, with only up to a quarter of all expenses made reaching an MVP state.

Based on the current trajectory, according to Cognilytica, the estimated market size for ML model operationalisation management will still reach $4 billion by 2025. And Marketwatch forecasts that by 2025 the global AI market size would reach $210 billion, already in 2019.

The question I have is how to simplify the transition and expand the market further?

AI has a complexity problem

As I grasp it, AI is unreachable for most companies today because the whole topic is too complex for most engineers (if the companies even have engineers).

Even if people understand the concept, then the whole process is still a nightmare — especially for beginners.

First, one needs to structure the data. Then they need to choose or train relevant models. Only to start fine-tuning them based on the existing data to achieve suitable results. And eventually, they need to scale the whole thing.

Yes, there are a wide range of approaches to implement AI within organisations. But most of them only suit those with lots of resources, or in case of needing to solve simple tasks.

The options are as follows:

1. In-house — teams with their own AI experts get to create fully customised solutions that address all complexity needs with the lowest error rate. However, this option only suits large organisations with a high budget and large developer teams due to the considerable resources needed.
2. AI consultancies — when lacking in-house expertise, outsourcing to someone like Accenture or McKinsey Digital enables the development of tailored solutions that are almost as complex and customised as those built within the team. But again, this option suits only the large enterprises due to the high fees and length of implementing such projects.
3. Intelligent RPA solutions — traditional RPA players, like Atos and Nintex, have started introducing AI capabilities to their process automation solutions that address complex multi-stage processes with unstructured data. Still, this option requires RPA developers for implementation. Thus this is again suitable for only medium and large enterprises with sufficient expertise in their team.
4. Vertical-specific AI automation solutions — there are various highly specialised automation tools for specific verticals, like 6estates for document processing. While they allow handling complex processes with unstructured data, their area of application is limited. And they are usually more expensive than other solutions due to the high cost of preparing the underlying data to train these models.
5. AutoML solutions — these tools allow people to create machine learning models from scratch, with no or low coding knowledge required. For example, there are products like Google’s AutoML, Microsoft’s Lobe.ai and Clarifai. But while they can help to give structure to unstructured data, they fail to automate end-to-end workflows since they lack the workflow builder feature.
6. Self-service tools — these, mostly no-code or low-code solutions like Levity and Noogata, allow building customised workflows and machine learning algorithms through a friendly UI. They come with lower performance for highly vertical tasks. But they are more suitable for most companies as they can handle end-to-end multi-stage processes and are affordable.

From all these, I see the most potential in the self-service tools.

These solutions can remove the steep learning curve for companies while speeding up the deployment time, from days to hours. Plus, with everything hosted in the cloud, they also drastically reduce computational resources requirements for most companies, even at scale.

That said, one of the main problems with existing self-service solutions is that users with higher expectations have to change their toolset as their needs expand. Or at least they need to start mix-and-matching between different tools, increasing the likely error-rate.

Thus there is room for new players.

What can be the different targetable markets?

As I see it, there are three different targetable markets for new self-service solutions.

First, there is a market for companies with more resources and prior experience with AI model deployment.

These firms can employ talent to prepare their data according to their machine learning needs whilst also training and deploying relatively fine models that reach production. At this point, they tend to mix-and-match a variety of tools to get the work done. For example, they can currently be using things like Gradient by Paperspace.

Of course, all this can be made easier for them via an end-to-end solution. But as these teams consist of people that know what they want to get done, they will have high expectations towards any solution they use.

Second, there is a market for companies with resources and understanding what AI can do for them but who have no prior experience and thus would likely outsource the work.

Without sufficient knowledge, the first step of cleaning data is already usually way too hard to comprehend. They might get suitable models, but there is only little they can do without the clean data.

Thus they either need someone to clean the data, buy clean data from someone else or a tool to do it for them. Here there is an opportunity to build such a product that starts by cleaning the data or integrating with existing data sources and continues to take care of choosing relevant models, fine-tuning them if needed and supporting the scaling of their needs.

Third, there is also a market for hobbyists and students trying to learn how AI works and who are simply looking for tools that enable them to test the models created by them quickly.

But even if they do use some tool, the main questions is how would this tool keep them around for the long term.

If they do graduate to the professional level, there is a need to offer a much more complicated service. Or if they understand that AI is too much for them, then there is still a need to provide a low-code alternative tool.

Either way, this segment is the worst as it tends not to have money or a strong enough moat to keep such users with them for the long-term.

In short, I would target the less-experienced businesses.

For them, we would need to build a simple infrastructure layer tool that either helps them with cleaning their data or integrates with their other products, so they can just plug-n-play — an easy-to-use construction kit for AI.

On the one hand, such a product would need to be simple enough for non-engineers to get started quickly. Enabling businesses outside the tech scene to categorise their email’s contents, summarise the main points of contracts and create marketing materials on the fly.

On the other hand, it would have to be powerful enough for more experienced users to do custom projects while being in complete control of the experience.

My question, however, is whether it’s doable?

If you have any ideas, let’s talk.