15 Questions Every AI Startup Should Have Answers To
An excellent approach to debunk hype
Artificial intelligence is the new cool in tech, with companies popping up every day that differentiate by applying a smart layer to an existing process. The trends are everywhere — instead of pitching a mobile/social/local spin on an old idea, startups are increasingly using AI to build excitement.
There are many good reasons that AI is particularly exciting and projected to revolutionize technology. Importantly, you no longer need a Ph.D. to apply game-changing techniques like deep learning. Along with increasingly fast and cheap hardware, the public research community is making great strides. Progress has manifested in open-source libraries and lower barriers to entry.
Still, many artificial intelligence startups are riding a wave of excitement without understanding the challenges, limitations and possibilities of AI. Here are fifteen questions that can be asked to understand how a company gauges the importance of artificial intelligence in their business.
Disclaimer: This post is written from the point of view of someone who practically applies AI in specific verticals, for others who are interested in the scope required to do so. This post represents my views as of one year ago, when it was written. This is a fast-moving field, for a more updated view reach out to me or watch my Stanford guest lecture on the subject here.
Data Questions
1. What data are you training on?
Data is a basic requirement for AI to perform well. Whether making predictions on objects in images, classifying a user’s intent from sentences, or creating recommendations from business metrics, an accurate data set is a must. This data is used to train the AI, directly informing the decisions that it makes. A clean data set should rank high in a machine learning priority list.
2. How are you collecting, cleaning and labeling your data?
Many deep learning techniques require a supervised training set of data, which basically means that you point the AI to the decisions that you would like it to make. These questions are important because oftentimes a labelled set of data is not readily available. Further, public sources of data like Twitter and Reddit often require cleaning and review, so that unintentional results can be avoided.
For examples of unintentional AI predictions, such as a racist crime-fighting AI, see: http://www.techrepublic.com/article/top-10-ai-failures-of-2016/
To illustrate this, a company might want to use AI to predict how to best educate a student. Machine learning can be used to do this, along with a vast set of relevant data. The data that is required would need to represent a few things very well:
- A student’s qualities & educational journey
- Their understanding of course material
Even if you are able to aggregate a data set, accurately representing the above attributes could be complex. How do you properly encode and quantify what a high quality educational journey is? What’s commonly measured today, test taking ability, may not be the proper variable to optimize for a realistic company.
3. How much data do you need to make initial predictions?
Many AI companies collect their data through their product, promising better predictions over time. It becomes a chicken-and-egg problem: it’s hard for the AI to make good predictions without lots of usage data, and its hard to get lots of usage without making good predictions. A technique to get around this is to inject domain knowledge into the system — a set of rules that can be used to help make predictions. Here is an example of a paper by Yang and Agrawal that uses domain knowledge injection to improve their deep neural network.
At Fireflies.ai we use Deep NLP to predict tasks from text. When we first started out, we generated our initial data set by collecting and manually labelling 20,000 of our own data points. Krish and I pulled an all-nighter, hacking together this interface, then labeling each data point twice. It was painful.
A general rule of thumb is that you need 1,000 data points per class to start making sensible predictions; at 10,000 data points per class, predictions are decent. A class is the number of different results that the AI predicts. As research advances, this number may change — OpenAI has achieved 99% accuracy on recognizing handwritten digits with only 10 labelled data points per class.
A sentiment analysis task has two classes (positive and negative), whereas predicting the next word in a given sentence may have 80,000 classes (potential next words). Generally, the more labelled data you have, the better your AI will perform.
4. How much data do you need to make personalized predictions?
If the goal is to personalize predictions based on the behavior of each user, AI startups should be aware that they will need data from each user in similar quantities discussed above (1,000–10,000 per class).
However, there are other techniques to reduce the amount of data needed to start making accurate and personalized predictions. One such technique is categorization — classifying the user into a bucket of other users for which data already exists. Quickly likening the user to other users through metadata or an explicit survey is a quick way to provide personalized results.
5. What biases are embedded in your data set?
Every AI startup should have a clear sense of the assumptions implicitly and explicitly contained in their data. Data and the predictions that it provides can be different depending on how the data was collected, and where it is being used for prediction. For example, data on company size and activity from AngelList may bias for more young companies than data from LinkedIn would. If this data is used to predict business metrics, it’s important to provide users with a full disclosure of the context of the suggestions that are made. These biases will affect your offering, as the AI (read: math) will predict along the lines of what it has been trained on.
6. How is your data at scale different from your data starting up?
This is particularly interesting if your AI is continually learning. What benefits, if any, are reached when you have a lot of clean, labelled data? What happens when the data is personalized? SwiftKey is a good example of a company with a very mature data set; they are able to quickly learn your typing style by looking at your email, messages, and social media posts. At scale, SwiftKey’s data set permits us to type lazily and spend less cognitive effort when composing messages. It seems like their AI magically knows what I’m typing even before it happens.
Each Fireflies.ai user begins with common meeting note predictions, but as users provide context by sending more meetings and reacting to the AI’s suggestions, data is fed back into the RNN. This trains the AI to intimately understand each team that is using Fireflies. For example, if one team’s client is Burger King, messages involving it may tend to be tasks whereas in another company those same messages are just chitchat about food. With continued usage, Fireflies.ai becomes a team member that adopts your team’s culture and workflow.
7. What are your data security policies?
Data security and privacy should be a priority for every AI startup. It is important to provide clarity on data policies to users of your service, especially when sensitive and/or personally identifiable information is exchanged. Is there a legal requirement in how your data should be handled?
In all cases, user data should be encrypted during transit and protected by a firewall. It should also be clear where data is going — is the machine learning happening locally or is it delegated to a third party cloud service? Finally, some AI companies need to supplement the AI’s suggestions with human verification. If this is indeed the case, the user has a right to know.
Business Questions
8. Is AI actually core to your business offering?
Companies should understand how critical artificial intelligence is to the problem that they are solving. Is being accurate the driver of the value that you are providing? Is it speed? What behavior is your AI actually replacing? Answers to these questions will vary depending on whether the business is consumer or enterprise focused.
Being smart or fast is not always the most critical piece of a company. For email assistants like Amy/Andrew from x.ai and Clara from Clara Labs, the critical piece is that 1) the meeting actually gets scheduled and 2) neither participant gets annoyed. A “smart and fast” agent is not as reliable as a human at the end of the day, and a mis-scheduled meeting is much more frustrating than a few extra minutes. Understanding the core drivers of the business is critical to gauge where AI can be the most helpful.
9. What predictions or process does your AI help with?
After understanding what drives value in your business, this is a straightforward question to answer. However, this must be included in this list as far too many companies just list the words “AI and Machine Learning” on their websites.
An understanding of the explicit predictions that create value or efficiency is a basic building block of a legitimate AI company. A satellite imaging company can estimate the change in traffic to stores by tracking what parking lots look like over time. An AI layer on top of this could predict consumer spending habits — something that is very hard and tiresome to do by a human alone.
10. How does your AI impact important customer metrics?
At the end of the day, a for-profit company needs to generate value for the users, employees and stakeholders involved. Measuring the impact of your AI implementation on important metrics like retention and churn can be challenging. Is it even possible to separate out the AI’s performance from things like product-market fit, product design and customer support?
Improvements to the performance of the AI can be understood through your test data set (which should be kept separate from the training set) and with a benchmark analysis that makes basic predictions on the data. Your benchmark analysis can use a set of rules to gauge your performance without machine learning. These improvements should constantly be compared to customer metrics to understand where business resources should be allocated.
11. What are the stakes of the predictions?
Or in other words, how accurate does the AI’s predictions need to be to reach your success metric? Businesses that have high stakes, such as making traffic decisions for a self-driving car, need to be much more accurate than the human standard from day one. On the other hand, keyboard predictions do not need to be accurate 100% of the time as long as the guessing is sufficiently accurate and useful. Understanding the stakes of each prediction being made can inform decisions about how to collect data, how to allocate data science development resources and when to deploy the AI in production.
12. Is there a human-in-the-loop?
A human-in-the-loop is an employee or third-party that aids the AI in making its decision. The human can label data points that a user was not able to, or the human can verify that the AI made a correct prediction before the user receives it.
Whether yes or no, both answers pose challenges from a business perspective. If there is a human-in-the-loop, how much does it cost to manually label each data point? Does the value derived from the AI’s improvement offset this cost? The hope for many companies is that the cost of data labeling goes down once a large enough set is created. Thankfully, companies like Crowdflower and Scale exist to help create a great data set with an on-demand work force.
If there isn’t a human-in-the loop, how do you verify that your data is clean and predictions are not diverging away from where they were intended to be? Do you actually understand why the AI is making its decisions? Seek out targeted customer feedback in this case; they will bring to light anomalies in the AI’s performance that you are unable to anticipate.
Societal Questions
13. Did you engender your AI? Why?
Intelligent conversational assistants in particular run into the issue of naming very early on. These agents can take on names that could form psychological bonds with its user. Humans are social creatures, and although AI is just a computer program, it can be comforting to talk to screens that listen.
The Amazon Alexa vs Google Home case study comes to mind when giving your AI a name. Google Home is an ungendered name, and invokes a business mindset every time you say “Ok Google.” On the flip side, Alexa’s engineers chose her name due to its familiarity “in everyday life” and for a “somewhat geeky Star Trek-ish reason” (source). Names that work for both genders are available, but innate biases may kick in anyway; if your AI has a voice, things get even more complicated. We chose Fred Fireflies as the name of our AI out of simplicity, and as an attempt to challenge the pre-existing stereotypes of note takers in corporate meetings.
14. Will AI replace humans?
People love asking this question. Although not relevant in day-to-day business operations, startups should be prepared for this dialogue. I think its important to form a solid opinion for this, as working in the early AI industry provides opportunities to shape where it goes. To many, successful AI employees are viewed as “experts” and could be referenced in unanticipated contexts.
I recently joined a panel of other AI startups in an event titled The Rise of Artificial Intelligence, where the general sentiment was that it feels like the early years of AI — perhaps similar to the Internet in 1995. It will get easier and easier for people to use AI and to understand how it works. Eventually, AI will empower each individual with a personal company of analysts. It is hard to project where AI will be in 10 or 20 years. I personally believe that it is a powerful tool that simply reflects the intents of its creators.
15. How will AI affect equality among society?
If AI becomes prevalent, it may replace jobs in a range of different industries. McKinsey released a report that demonstrated that AI can automate 45% of the activities that people are paid to perform. What will happen to the people working in these jobs?
Both aged and recent examples from history can lend some insight. During the Industrial Revolution, the power loom also threatened jobs when it automated traditional hand weaving performed by millions of people. In reality, the new technology allowed these people to focus on a set of new skills. Markets responded; they changed to demand more of these new skills, and the pay of the affected workers class actually rose to match the new needs.
More recently, the rise of computers and mobile phones automated many processes away (like manually storing files), at the same time creating new markets for services like Dropbox. In the same way, continued advancements in AI may increase our demand for efficiency and make us larger consumers.
Closing Remarks
These questions should not scare passionate founders away, but rather excite them. There are a large host of challenges to be solved.
I am eager to learn more about the problems that AI startups are facing and how to continue solving them. Ping me if this excites you and you want to brainstorm more 🤔
Sam co-founded Fireflies.ai to capture and automate tasks as they occur during calls and meetings. Tweet @ him!
