How to Pick the Best LLM for Your Sales Activities
Since the introduction of OpenAI’s ChatGPT a little more than a year ago, large language models have captured the imagination of sales professionals, who are eager to see how generative artificial intelligence could simplify their workflows and relieve them of some of the mundane tasks that take up the time they could use for actively selling.
In just a short amount of time, genAI and LLM options have grown extensively. The major LLMs—ChatGPT, Google’s Bard, Microsoft’s Bing Chat, Meta’s Llama, Salesforce’s Einstein GPT, Anthropic’s Claude 2—are all serious competitors in the space. Or some really ambitious companies might even try to build their own LLMs, combining their own technology with some foundational components from these other companies. With that many choices, how do you know which one is right for you or within which applications to deploy them?
The best LLM for sales? It depends, experts agree.
First, sales executives need to determine how the LLM will be used for sales, says Frank Schneider, vice president and AI evangelist at Verint. The LLM might be used to generate creative content for email blasts, for product descriptions, or some other sales-related need.
Different LLMs might work better with a particular company’s data and systems, so experiment to see which integrates best, Schneider recommends. “A lot of it is going to depend on where I work. If I’m at a Microsoft shop and use lots of Microsoft products, you’re probably going to lean toward [Microsoft’s Copilot].”
Similarly, if Google or Salesforce products make up the bulk of the technology stack, that company’s LLM is likely to be the preferred choice, Schneider says. “It comes down to cost, accuracy, performance, speed, and data security on some level.”
For companies that don’t lean to one particular technology vendor, Schneider recommends that sales leaders examine Hugging Face, an open-source platform where the machine learning community collaborates on models, datasets, and applications.
The Hugging Face platform is helpful because LLMs are evolving quickly. In early December, for example, Google launched its latest large language model, Gemini. Google executives see the new LLM as a huge leap ahead, eventually impacting practically all of its products.
Other LLMs are evolving as well, Schneider points out. “Right now, there is no best of breed. The best of breed is in the eyes of the beholder.”
One LLM might perform better for some tasks, while another will excel at others, so the right LLM for one organization’s sales strategies can be very different from the right one for a different organization with a different sales strategy, he adds.
The LLMs are becoming a platform, Schneider says. “The apps are spinning on the edges, but you’re going to see a lot of collaboration on this technology. Instead of having a bunch of point solutions, you’re going to start to see a suite of solutions centered on an LLM as a foundation.”
DON’T BE HASTY
With the rush to add LLMs, sales executives might move too quickly for fear of being left behind, Schneider cautions.
Determine the potential sales uses for the technology, then prioritize them, Schneider recommends. Then plan training and policies for use of the technology.
Before choosing any LLM, you need to be clear on your sales strategy, says Kevin Raybon, founder of Phacyt Advisors, a consultancy to help go-to-market teams with their tech investments. “Otherwise, you’re putting the technology before the sales. That’s the problem that we have fallen victim to for years: We see a vendor who is coming to us with features, functions, and new technologies. We jump on board, sometimes without even knowing what it is and how the capabilities are going to help us reach our goals. When I talk to people about the technology, I’m not going to talk about it from the LLM level. We’re going to start at the strategy level. We’re going to think about the challenges you have today and your go-to-market strategy.”
After discerning those answers, it’s time to determine which LLM, if any, will help reach those goals, according to Raybon.
Different companies have different organizational structures, Raybon adds. Some organizations make sales responsible for heightening awareness about products and services. That responsibility falls to marketing at other organizations.
Similarly, other organizations might make their decisions based on which LLMs they think would be best at capturing, compressing, and reporting meaningful information about their market, competitors, etc., to give salespeople a competitive advantage.
Experts agree that before selecting any LLM or multiple LLMs for different types of tasks, sales needs to consult with IT. LLM technology is still in its infancy, and any choice can be foundational for the company if the choice is correct.
“Large language models on their own are useless without a robust technology strategy for proper governance, implementation, and use cases,” agrees Peter van der Putten, director of Pegasystems’ AI lab. “ChatGPT, BingChat, Bard, etc., are all consumer products. Enterprise AI will use the same models that sit underneath—GPT-3.5 or GPT-4 for ChatGPT and Bing, PaLM for Bard, etc.—but the enterprise AI experience will be very different. It won’t be a free-form chatbot, but instead, generative AI will be built into very specific business processes and customer journeys and interactions.”
For that reason, van der Putten says that whether going with an emerging LLM or a more established one, companies should pay particular attention to the technology’s governance and compliance attributes.
Use cases also cannot be ignored, he suggests.
For example, a salesperson could use generative AI to automatically summarize client conversations from diverse points of view, van der Putten adds. The technology will extract sentiments and topics and formulate clear, tangible calls to action, such as suggesting top executives as email targets. Sales managers will get autogenerated status reports around opportunities, accounts, or entire portfolios, based on sellers’ activity, as well as inbound emails or marketing activity and response. There will be no need to type in prompts or interpret model responses.
This will all be fully automated, and generative AI will pull in relevant knowledge and data to formulate plans and courses of action, van der Putten says.
In this scenario, the underlying foundation model, whether it is GPT, PaLM, Meta’s Llama 2, or something else, will become a commodity, according to van der Putten. “You actually want to abstract away from which service is being used and be able to switch services in the back end seamlessly based on quality, cost, and governance, without affecting any of the front-end experiences.”
He adds that a salesperson will typically use an LLM to determine the best wording for vital emails to prospects and customers; an upper-level sales executive, meanwhile, is more likely to use the technology to summarize trends across different sales portfolios. So salespeople and executives might prefer different LLMs.
With that in mind and with the evolution of various LLM offerings, there is no single best choice, according to Raybon, who likens the current state of LLMs to that of computing in the early 1980s, when different solutions took turns as No.1 as they evolved at different rates.
COMPARING THE LLMs
Most experts agree that building an LLM from scratch is a risky and unnecessary process. Zac Sprackett, chief technology officer of SugarCRM, for example, calls it “a big mistake.” “The barrier to entry is low. There are partners out there who can help companies get the most value out of an LLM so that they can focus on their business.”
So if it’s better to rely on a third-party system, which one is better? There is no easy answer, according to the experts.
“ChatGPT captured a lot of people’s attention because it was the first to wrap a great user interface around the large language model, and [OpenAI was] bold enough to put it out in front of people, warts and all, and let it capture people’s hearts and minds,” Sprackett says.
ChatGPT has advanced from where it was a year ago, with better tools, better ability to pull the most recent information from the web, and the ability to grab information via APIs, according to Sprackett.
ChatGPT also offers multimodal capabilities, as it can now evaluate and generate images as well as text, Sprackett adds. However, Google’s latest advance could add similar capabilities.
Keeping in mind that the different models are adding new capabilities at different rates, Akkio, an AI company that delivers generative analytics and machine learning to businesses, recently benchmarked ChatGPT and Anthropic’s Claude 2. Among its findings were the following:
• GPT-4 scored in the 93rd percentile in reading comprehension while Claude 2 reached the 86th percentile.
• Claude 2 outperformed GPT-4 in analytical writing, scoring in the 96th percentile compared to 89th for GPT-4.
• Claude 2 achieved 88 percent accuracy in math vs 83 percent for GPT-4.
• Claude 2 scored 71 percent on the Codex HumanEval Python exam compared to 67 percent for GPT-4, meaning that Claude 2 can generate code more accurately.
• GPT-4 edged out Claude 2 in common sense reasoning, with 83 percent accuracy vs. 82 percent.
• GPT-4 showed a better understanding of nuance and ambiguity in language.
Sprackett and others maintain that LLMs coming from CRM providers will provide the added benefit of being able to pull information—including customer interactions, lead generation sales success, customer support details, etc.—directly from those systems without requiring a lot of additional integration and coding work.
“ChatGPT, Google, and others don’t have that kind of customer-level detail, and they’re not going to be able to produce it,” Sprackett says. “For a business that has a CRM system, having access to all of that data and using it in conjunction with large language models is critical for generating responses that aren’t generic, that are unique and tailored to your specific business and the way that you support your customers and the products and services that you deliver.”
There is one major caveat when it comes to choosing the right LLM: After your sales organization determines how it will use an LLM, conducts its due diligence, consults with IT and with other experts inside and outside of the company, and then chooses and starts using one LLM, another might advance ahead in terms of usability and benefits. If that happens, switching from one to another is relatively easy for peripheral uses of the technology, according to Sprackett.
But if a company has LLMs integrated with APIs, it becomes a more complex transition.
Sprackett says that SugarCRM enables companies to switch from one LLM to another and to use different LLMs for different purposes.
Sprackett added that the advancement of LLMs in sales and related industries is also producing a cottage industry of prompt engineers—people who are well trained and experienced in structuring queries to elicit the best responses from LLMs.
“Not everybody should have to be a prompt engineer to get all of the value from a large language model,” Sprackett says. “Our responsibility as a vendor is to gather all of the relevant information that will produce the best possible response and assemble it behind the curtain so that the end user doesn’t have to.”
An example would be using the LLM to gather and summarize all relevant customer information for a quarterly consultation with that customer.
“It’s our job as a vendor to pull together all of those different data sources and ensure that they are available for the creation of that crisp answer,” Sprackett says.
And at the end of the day, there is one last caveat that will, and rightly should, impact any and all LLM decisions: Some new LLMs are likely to emerge, and there is likely to be some consolidation in the industry as well, van der Putten cautions. “There will be a lot more companies in production with large-scale generative AI. In a year or 18 months from now, I’m sure that the back-end ecosystem will be in flux. I don’t think that will settle anytime soon.”
Phillip Britt is a freelance writer based in the Chicago area. He can be reached at email@example.com.