To Build or Not to Build? 5 Reasons to Develop Your AI Stack In-House

Micha Y. Breakstone

Micha Y. Breakstone

Using AI, or rather saying you use AI, has become extremely popular with startups these days. It seems no matter what your startup does, adjoining the descriptor “with AI” is a sure bet, and examples abound: create your new company logo with AI; write more engaging presentations with AI; Identify security vulnerabilities with AI, and heck, even script a new car advert – you got it – with AI.

And that makes sense. According to Forbes, as of 2018, there’s been a 14X increase in the number of active AI startups, with venture capital investments increasing 6X in AI companies since 2000. At the same time, all the top tech companies which set the trend for the industry in general, and for the M&A scene in particular, are focusing more and more on AI. FAMGA (Facebook, Apple, Microsoft,  Google & Amazon) have all built their own AI Virtual Personal Assistant (VPA), they’ve also all built, or are currently building their own AI hardware, and Google’s CEO went so far as to declare that Google is an AI-first company.

So it’s no wonder entrepreneurs are drawn to build their companies with at least a hint of AI.

That being said, building your own AI in-house is extremely costly and requires expertise and focus, and at the same time, many AI-driven tasks can be simply and effectively outsourced through APIs or built by non-experts with easy-to-use libraries. Examples include Google Speech API and IBM Watson for Speech Recognition, TextRazor and spaCy for a wide variety of Natural Language Processing tasks, and Google’s cloud vision API, Microsoft’s Computer Vision API, and Amazon Rekognition for image analysis.

So should you invest the time and money in hiring an AI team in-house to build out your own proprietary AI stack? If AI is an inherent part of your value proposition – and not only an add-on or afterthought –  you probably should. Here are 5 important reasons to consider.


Precision – External APIs are either not customizable or only minimally customizable, and out-of-the-box solutions will not account for the noise and oddities of the “real world”. Consider for example a company building a Virtual Personal Assistant that is geared at children. Using an external API for Speech will get you nowhere, because to achieve good accuracy one needs an Acoustic Model that is trained on children (higher-pitched voices). In the same vein, a company building a tool to help lawyers or doctors take notes will want to augment standard English vocabulary with all the relevant technical terms (legalese/medicalese), and names of relevant companies, drugs, or other entities. The bottom line is that vertical-specific precision is impossible with off-the-shelf APIs, and getting the fundamentals as precise as possible is critical for any downstream processing.


Costs – While using APIs isn’t costly on a one-off basis, using them at scale can incur significant costs. For example Google Speech API costs around $0.006/15 seconds, or around $1.5/hour, and Amazon Rekognition Video API is only $0.1/minute or $6/hour (non real-time), which may sound cheap, but if you’re processing thousands of hours of voice or video a day, that becomes a huge expense. While developing your own AI – including hiring and building a team of experts – is an expensive initial effort, it ultimately allows you to build a product with much higher profit margins as you scale.


Privacy – Privacy is becoming top of mind globally. For example GDPR has been enforced since May 2018, not to mention all the recent scandals with misappropriation of user data by tech giants. There are many cases in which you don’t want to send data to a third party: if you’re analyzing sensitive customer conversations, if you’re building a business-geared VPA, or if you’re analyzing images in secure locations. Building your own AI means you don’t have to send your data to any 3rd party, and can maintain maximal privacy on your cloud.


Real-time – While using APIs can be quick and some near-real-time services exist, any application that needs to respond in true real-time will have to account for network latency, 3rd party downtime, and delegating matters of buffering and segmentation to a 3rd party. For the last point, consider an app that processes conversations using an external API for speech recognition, with an objective to guide a speaker during the conversation. Even an API that processes conversations in near real-time, would leave the app developer with no control over when utterances begin or end, which may in turn result in delays of 10 seconds or more as the transcription results wait for the end of the utterance to be determined.


Owning your destiny – While many standard AI tasks are supported by external APIs and/or highly sophisticated packages that allow non-experts to deliver expert-level AI results, this is not the case for many specialized use cases. If one truly want to own their destiny, and be able to go where others haven’t gone before, they’ll need to build their AI in-house. Here are a few real examples (all startups):

  • Analyzing and understanding the meaning of legal contracts (
  • Analyzing and understanding best shelf ordering at retail shops (
  • Diagnostics through medical imaging (
  • Analysis of DNA and extraction of insights (
  • Anomaly detection in data (
  • Analysis and understanding of sales conversations (, of which the author is a co-founder)

In sum, if you’re deliberating whether or not to build your own AI stack in-house, first look in the mirror and ask yourself: is AI an inherent part of the company’s value prop. If it is, consider the reasons detailed above: Precision, Cost, Privacy, Real-time, and Owning your destiny. Most likely at least one of these reasons will apply to you, and while most you will not need to build everything from scratch, you’ll definitely want to build at least part of your AI stack in-house.

Related posts