Blogs

Why Services+SaaS works for Enterprise AI

Software Ate the World, Now AI Is Eating Software. Traditionally, enterprise software was mostly sold as a service – where service providers used to charge for the time it took the software developers to first develop and then maintain a custom solution for a client.
Future of AI Saas or Services

Future of AI SaaS or Services

Software Ate the World, Now AI Is Eating Software.

Traditionally, enterprise software was mostly sold as a service – where service providers used to charge for the time it took the software developers to first develop and then maintain a custom solution for a client.

The SaaS subscription revolution powered by the cloud changed the dynamics as now the solution could be built once and sold an infinite number of times. As is evident from the latest trends, one by one every industry and function is being eaten by SaaS software as predicted by Marc Andreessen. Salesforce, Slack, Zoom are just a few of the many such examples.

But the AI revolution that started in 2012, shook up things a bit.

  • Data: AI solutions depend a great deal on the quality and quantity of data Sets. The belief was that SaaS companies offering similar solutions across the industry would be able to leverage data of all their clients to offer a highly accurate solution for the entire industry. The ‘data network effect’ will lead the market leader to become better and better with every new client and thus dominating the industry.
  • GPU Processing: Similarly, requirements of GPUs for processing meant that it made sense to use the scale of the cloud to be leveraged for training models as well as for inference.

This cocktail led to a rush of AI start-ups and VCs backing several such start-ups which promised to create the next big SaaS company for enterprise leveraging AI.

Two Act Tragedy

Act 1 - Long-tail of edge cases

It turns out, that real-world use cases often have a really long-tail of edge cases. This makes it really difficult to use one machine learning model across clients and use-cases.

This makes it impossible to leverage the supposed ‘data network effect’. Rather, what makes more sense is, either to fine-tune models on every client's data or create client specific models to handle edge cases.

Though this works well – it breaks the standard mould of ‘SaaS’. Often, AI providers end up working almost as a software service provider rather than offering a SaaS product across industry.

Act 2 – Heavy data often makes using cloud inefficient

AI adds the best value for the data which is in the form of videos, images and audio files. The challenge here is that all these formats are heavy data files. It is expensive to upload this data onto the cloud.

Even if you are happy to bear the cost, processing the data in real-time then becomes a challenge – as internet bandwidth becomes a major constraint.

So, the better choice is to do processing on on-premise GPU hardware – which makes it cheaper to deploy and also enables faster response time for the models.

As the world is moving to cloud-based services, AI is pushing solutions back to on-premise deployments. Again, this breaks the classic mold of cloud-based SaaS models

But this does not mean that we can simply go over to the old days of services projects. There are three aspects that still mandate SaaS like subscription model

Content drift

Using past data to predict future works - but often only up to a point. A model trained on last year’s data may work this year but is not really bound to work the next year. This happens because data distribution changes with time.

This necessitates that the data is captured, annotated and trained on repeatedly. This then makes it critical that the AI solution provider is engaged throughout. So, having a SaaS like subscription aligns incentives of both the client and the AI solution provider.

Evolving AI models and approaches

The field of AI is rapidly evolving with new algorithms, approaches and hardware are being introduced at breakneck frequency. Enterprises have no option but to make sure that they engage an AI solution provider who keeps them ahead of the curve.

This again provides a huge incentive to firms to have a SaaS partner on board which ensures that the AI stack they are using is the best.

SaaS makes sense

All the remaining benefits of SaaS do apply to AI SaaS as well. AI is but only a small part of the entire solution. The complete solution involves – UI, UX, integrations, scalability etc.- all of which greatly benefit from the scale that SaaS offers.

So, though the AI models may be running on the local servers it may still be better to run the application on a cloud which is continuously talking to the on-premise server.

What does this mean for us?

Though we at Silversparro have realized these challenges in offering AI solutions as pure SaaS or pure Service quite painfully over the last 4 years, we are not the only ones. Awesome folks at Andreessen Horowitz have written a near-perfect piece describing the new business models that are emerging with AI.

A.     ‘Services’ like initial model development: This consists of following aspects which are similar to consulting or services engagement:
  • Understand problem statements, available data, and edge cases
  • Create data capture strategy to continuously create training data
  • Provide server specifications for on-premise deployment
  • Collect and annotate training data
  • Train, test and deploy custom models
  • Configure reports, alerts & dashboards
B.      ‘SaaS’ like subscription: This consists of all the features that ensure that the solution delivers business value throughout the engagement period.
  • Maintain accuracy of all models
  • Maintain solution uptime
  • Continuous AI algorithms upgrade pushed remotely
  • Ensure availability of reports, alerts & dashboards
  • Remote troubleshooting
  • Continuous feature updates just like regular SaaS

This combination of starting with “services like” model training followed by the “SaaS like” subscription period allows both partners to leverage the best of both worlds within the constraints of the AI technology.

As the article by a16z says that “AI is still early in the transition from research topic to production technology.” We, too, are eager to learn and understand how other start-ups are solving these challenges.