Enterprise AI is a set of products and services offered to enterprise companies, based on building custom machine learning models, which can take over various tasks, and be integrated at scale, within the workflow, of a large corporation.
The Cambrian explosion of AI companies
Since 2019, I’ve been selling AI services at an enterprise level.
The real turning point, came, in early 2019, when GPT-2, was released. I still remember, when the founder of a tech start-up which I helped develop the business side, demoed me the model.
In the previous demos, I had been disappointed, and actually, I had thought the whole thing was laughable, but that time, I didn’t, I was impressed!
The text generated by the machine, not only made sense but was also extremely good.
GPT-2 and later on, GPT-3 are language models, developed by OpenAI.
They spurred a whole new AI industry, based on generating content at scale.
I thought in my mind how that would change the whole publishing industry.
However, it’s also important to point out the various drawbacks that those language models have.
For instance, a few days back, I asked GPT-3 to tell me something about “business validation” and that is what it produced:
Business validation is the process of verifying that a business idea is sound and that there is a market for the product or service. validation can be done through market research, surveys, interviews, and focus groups.
Good, isn’t it?
It does make a lot of sense.
However, it’s wrong!
I’ve been devoting my whole professional life to business validation, and I developed my own way to look at it.
As such, this generic, mediocre, definition, which GPT-3 provided, does make sense. But it’s worthless, and actually very bad.
Indeed, my argument is that the last thing you want to do, to validate a business is to use market research, surveys, interviews, and focus groups.
Instead, you want to use other approaches.
In short, this doesn’t want to be a debate against the machine, but rather, a way to understand where the machine is useful and where it’s not.
For instance, when it comes to defining “business validation” the machine is not only a mediocre consultant but its definition is misleading.
Does it mean it’s worthless?
As a business person, I think the value of what I do stands, in the personal experience developed over the years, the consistency in the message provided, and the fact, that here, you find stuff you can’t find anywhere else.
Thus, the whole value you might get from me is given by my personal style, a combination of expertise, experience, understanding, and a lot of mistakes, which I’ve made, that I want you to either avoid or leverage to speed up your journey.
That is how trust is built.
This means that if I were to use this stuff, I’d kill my business and credibility, quite fast!
But if the machine can’t be good at having its own style, what can it do effectively?
Where does the AI make sense from a business standpoint?
Based on my experience in the last three years, refining enterprise services, in the AI marketing space, these are the applications that I’ve seen as successful ones:
Humans in the loop
One critical element is that, for now, the human must be in the loop. Whatever application, for the enterprise, you choose, it’s critical to have a team of experienced people that are able to understand how to handle these models at scale.
When building AI models, those custom models (built for specific tasks of the company) might become more and more relevant over time.
However, it’s also critical to have the human in the loop, both at the input and output levels.
On the input side, it’s critical to have people spend time enriching the data available to the model and improve it (imagine if you want to make a language model that describes your products, the more information you give it about these products, details, and features of the products, the better the model).
On the output side, as the model generates things that go off, those must be corrected and audited, which connects to the next point.
Model auditing – I argue – will be larger than modeling itself!
In my opinion, as those machine learning models grow in popularity, spurring the next trillion-dollar industry, there will be a core issue with them.
As those models, mostly leverage on deep learning, this means that you know the input and output, but in the middle, it’s a black box. In short, there is no simple way to audit the model, to know how it got from input to output.
In that scenario, being able to devise smart ways to audit machine learning models, will be extremely valuable.
Indeed, if you ask me, where I’d start an AI company, I’d tell you right on that I’d start by building a toolbox to audit machine learning models.
Why? Because, any company using AI services, will need to be able to audit these models, and there is no right way to do it now. It’s still a Wild West!
Commoditizable content is the starting point
If you want to start leveraging these machine learning models, you might want to start by looking at the part of the business which has the potential to be commoditized.
For instance, if you have an e-commerce store, with thousands or millions of product descriptions.
Language models can be pretty effective in generating them, at scale.
In fact, the advantage here is that the language model can dynamically change the product description, also based on seasonal searches of users. This is an application that I’ve seen been used very effectively.
However, one thing is to give the machine the ability to change a byte of text on a product page, another is to have the machine rewrite the whole page.
In short, you want to start from, a small section, which can be controlled, easily measured, iterated, corrected, and scale from there.
Scaling up models
One of the most difficult parts of using machine learning models at an enterprise level will be to scale those models reliably.
For instance, if you have the machine generate text for 100 product pages, is completely different, from a thousand, or ten thousand.
At each level of scale, the complexity the machine handles grows exponentially, and the chance, that a few product descriptions, get way off, becomes a possibility.
Another interesting application of deep learning models is in the realm of paid campaign optimizations, where the machine can work in two ways.
First, the machine takes unstructured data and makes it into structured data. Imagine if you’re spending millions on Facebook ad campaigns.
Those are mostly handled by a performance manager.
For how organized that person might be in handling these campaigns when the budget gets very large, it might become also very complex to understand campaigns that are performing well.
That’s because those campaigns might miss proper labeling. In short, the trivial, but time-consuming task to label these campaigns (like organizing them in clusters that make sense) becomes very hard.
Second, handling a large number of campaigns might also slow down the experimentation process.
Indeed, successful campaigns, over time, needs to be continuously tweaked, changed, and re-tested, to keep the ROI on these campaigns stable.
Deep learning, with neural nets, might be very good at that! Labeling, adjusting, testing, and iterating, these campaigns at scale.
How do you price AI services?
So far, I’ve been working and seen these primary pricing models:
Here the AI company, builds a custom model, through a pilot (either free or charged at a flat fee).
Once the model is ready, it goes through a transitional phase, which starts to run as a subscription/retainer (based on the volume handled by the model).
From there, once ready for scale, the price carries a base subscription, which beyond a certain volume, will spike up.
The subscription has to have a threshold, after which, the volume the model is able to handle might be unlimited (this happens in a very advanced phase, which based on the service, might take at least 18-24 months to develop).
These comprise services of AI companies, which might provide standard libraries of machine learning models, and the company that picks one will pay based on the consumption.
These comprise AI models which are standard but can be customized to a certain extent.
In that case, a subscription model, combined with a pay-as-you-go model might do.
A cut of the saving
In the case of a company that uses AI models to improve campaign performance, the company might introduce the customer to these models, with a free or flat fee pilot, and once this proves successful, only charge a % of the savings.
This model might be more effective to reduce the friction and acquisition cost of the enterprise customer.
With ♥️ Gennaro, FourWeekMBA