Large language models (LLMs) have taken over the business world, and now every company is attempting to use Generative AI. Although tools like ChatGPT are clearly powerful, it is just not clear how businesses can reliably use this technology to drive value.
For many businesses I’ve interacted with, “using AI” means constructing a chatbot, co-pilot, AI agent, or AI assistant. Nevertheless, because the initial excitement about these solutions wanes, organizations are realizing the important thing challenges of constructing systems around LLMs.
A central challenge is that LLMs are inherently unpredictable (much more so than traditional machine learning systems). Subsequently, it’s challenging to get them to unravel a selected problem predictably.
As an illustration, one solution to the hallucination problem is to have “judge” LLMs review system responses for accuracy and appropriateness. Nevertheless, increasing the variety of LLMs increases the system’s cost, complexity, and uncertainty.
This is just not to say that Generative AI (and friends) should not price pursuing. AI has made countless corporations very wealthy, and I don’t think that may stop anytime soon.
The important thing point is that value is generated through solving problems, not using AI (in itself). AI’s promise is realized when businesses discover the right problems to unravel, e.g., Netflix’s personalized recommendations, UPS’s delivery route optimization, Walmart’s inventory management, and lots of others.
While “solving the correct problem” is simple to say, it is just not easy to do. To assist with that, here I share 3 AI use cases for something every business cares about — sales. My hope is to get your imagination going and reveal easy methods to implement them with concrete examples.
The three use cases are:
- Feature Engineering — Extracting features from text
- Structuring Unstructured Data — Making text analytics-ready
- Lead Scoring — Identifying your best opportunities
Featuring engineering consists of creating variables that could be used to coach machine learning models or perform some evaluation. For instance, given a set of LinkedIn profiles, extracting things like the present job title, years of experience, and industry, after which representing them numerically.
Traditionally, this is completed in two ways. 1) you manually create features, or 2) you purchase features from a third party (e.g., credit scores from FICO, company revenue from D&B). Nevertheless, LLMs have created a 3rd way to do that.
Example: Extracting Features from Resumes
Suppose you might be qualifying leads for a SaaS offering. The software helps protect mid-market corporations against cybersecurity threats. The goal customers are IT leaders who resolve which vendors suit their corporations.
You may have a stack of 100,000 skilled profiles and resumes gathered from various sources based on the tags “IT,” “Cybersecurity,” “leader,” “VP,” and a number of other others. The issue, nonetheless, is that the leads are low quality, often including non-IT leaders, entry-level IT professionals, and others who don’t fit the client profile.
To be certain that sales efforts are focused on the correct customers, the goal is to filter down the leads only to incorporate IT leaders. Listed here are a couple of ways to unravel this problem.
- Idea 1: Review all of the 100,000 leads manually. Problem: Impractical for a single person or small sales team
- Idea 2: Write rule-based logic to filter resumes. Problem: Resumes are available a wide range of formats, so logic performs poorly.
- Idea 3: Pay an information vendor for this information. Problem: This significantly increases the associated fee of customer acquisition (~$0.10 per lead)
Given the problems with the ideas above, let’s consider how we could solve this problem with a big language model. An easy strategy is to craft a prompt that instructs an LLM to extract the specified information from a resume. An example is given below.
Analyze the next text extracted from a resume and determine whether the
person works within the IT industry. Return a `0` if the person doesn't work in
theIT industry, and a `1` in the event that they do. Then, provide a transient explanation for
your conclusion.Resume Text:
{resume text}
This solution is an ideal mix of the three ideas above. It (1) reviews each lead in search of specific information like an individual, (2) is automated by a pc program, and (3) you pay less money (~$0.001 per lead).
**Bonus**: For those all for implementing something like this, I share an example Python script here that extracts Years of Experience from a LinkedIn Profile using the OpenAI API.
Data from emails, support tickets, customer reviews, social media profiles, and call transcriptions are all examples of unstructured data. This simply means it is just not organized in rows and columns like an Excel spreadsheet or .csv file.
The issue with unstructured data is that it is just not analytics-ready, making it difficult to realize insights. This contrasts with structured data (i.e., numbers organized in rows and columns). Translating unstructured data right into a structured format is one other area by which recent advances in natural language processing (NLP) and deep learning may help.
Example: Translating Resumes into (Meaningful) Numbers
Consider the identical business case from the previous example. Suppose we successfully picked out 10,000 IT leaders from the 100,000 leads. While your sales guy could start picking up the phone and crafting emails, you first need to see if you happen to can distill the list to prioritize leads just like past customers.
One approach to do that is to define additional features that provide more granularity to the perfect customer profile (e.g., industry, compliance requirements, tech stack, geographical location), which might be extracted similarly to Use Case 1. Nevertheless, identifying such indicators is likely to be difficult, and developing additional automated processes comes at a value.
An alternate approach is to make use of so-called text embeddings. A text embedding is solely a numerical representation of a piece of text that’s semantically meaningful. Consider this like translating a resume right into a set of numbers.
The worth of text embeddings is that they translate unstructured text right into a structured table of numbers, which is way more amenable to traditional analytical and computational approaches. For instance, on this context, one can use text embeddings to mathematically evaluate which leads are most just like past customers and that are most different.
The ultimate use case is lead scoring, which consists of evaluating the standard of a lead based on key predictors (e.g., job title, company revenue, customer behavior, etc.). While that is nothing latest, recent advances in AI have enabled a greater ability to parse unstructured data that could be fed into lead-scoring models.
Example: Grading Leads Based on Quality
To conclude our ongoing business case, let’s discuss how we are able to use text embeddings to prioritize potential customers. Suppose we’ve an inventory of 1,000 past leads, 500 of whom bought and 500 of whom didn’t. For every lead, we’ve a profile that features key information comparable to job title, work experience, current company, industry, and key skills.
These leads could be used to coach a predictive model that estimates the probability that a customer will buy the product based on their profile. While there are various nuances to developing a model like this, the essential idea is that we are able to use the predictions from this model to define grades for every lead (e.g., A, B, C, D), which could be used to categorize and prioritize the ten,000 latest ones.
**Bonus**: For the more technical readers in search of to implement these approaches, I walk through all three use cases applied to real-world sales data from my business on this video. Moreover, the instance code is freely available on GitHub.
AI holds tremendous potential for businesses. Nevertheless, realizing that potential requires identifying the right problems to unravel with it.
With the ubiquity of tools like ChatGPT, solution ideas can easily be limited to the AI assistant paradigm. To assist expand the space of possibilities, I shared 3 practical AI use cases that use alternative approaches.
More on AI for business 👇
