How Will the Arrival of DeepSeek Impact CX Operations and Workflow Automation?

As the Chief Scientist here at Verint, I get a lot of questions about AI. But in the past couple weeks, I’ve been getting a nonstop flood of pings, emails, texts and calls.

It hasn’t been this crazy since the launch of ChatGPT.

As you may have guessed, all these folks want to know about DeepSeek-R1, the latest large language model released from Chinese AI company DeepSeek. They’re curious about the open-source model’s touted low development cost, its advanced conversational abilities, and how it will affect the way other companies build generative AI applications.

First off, a few things about DeepSeek-R1 are important to understand.

How DeepSeek compares to other AI models

As much as people are talking about the cost and compute use of DeepSeek compared to US companies, the quoted cost of $6 million comes from the paper on v3, the predecessor of R1. As DeepSeek fully admits in the paper, the quoted cost to train the model does not include anything but the official training time for v3.

It doesn’t include all of the experiments and development work that went into building their custom training pipeline, generating data examples from previous models to train v3 and R1, optimizing intra node communications protocols, and any practice and ablation runs before the final training.

So, it is inaccurate to say it took “only” $6 million to create R1 when in fact that number is not even the complete amount needed for them to develop v3, and the R1 paper does not specify how much additional time and resources the training required to advance from v3 to R1.

But the cost issue aside, it’s nevertheless an impressive accomplishment to match or exceed OpenAI’s current flagship model—and especially doing so with less powerful hardware.

So, to answer one of the questions I’ve been constantly getting as of late—how does Verint plan to use DeepSeek?

To answer that, it’s important to explain how we pick the right foundation model for enterprise use cases. This is more complex than using the current benchmark leadership—in this case, DeepSeek. While public benchmarks can give us a starting point for a set of models to begin evaluating, many times we find the current top-ranking models fall short in other areas that are important to our products and the outcomes they deliver to our customers.

That’s because at Verint, we choose the right model for each specific application workflow we are automating. To deliver quality customer service and better workforce optimization, you need the right AI model for the job.

When selecting AI models, Verint considers a variety of criteria:

• Accuracy at the task with application-specific data and stability performing the task across a diverse set of input.
• Speed in task is within product constraints for both Time to First Token and Time to Last Token. Not all applications stream responses, and some use chain pipelines that block waiting for complete responses.
• The cost for projected usage is important. We want to make sure we’re not incurring unneeded expense for our customers.
• Regional availability or ability to self-host. We deploy products worldwide with on-premises, hybrid and cloud customers.
• If the model is third-party hosted, do they provide all of the security guarantees we require, such as PCI 4.0, SOC2, HITRUST, GDPR, etc.? And how long can we expect the model to remain available? Model migrations for existing products can be time consuming and expensive to ensure there is no regression in quality or behavior.
• Multilingual performance—our products support more than 50 languages globally.
• License terms and conditions.
• Is the model safe? How likely is the model to produce undesirable output, and how resilient is it to prompt injection and coercion?
• Is the model responsive to prompt engineering? For instance, if we want to make customer-specific alterations to the output, how much work will that require?

At Verint, we realize that the development of new and better LLMs is integral to the evolution of better CX automation and automated workflows. This evolution of AI helps our bots get better over time as new, better performing, and lower-cost models become available.

While we’re definitely interested in seeing how DeepSeek evolves and how we might be able to put it to work in Verint bots, here are a few things that currently keep us from making that a reality right now.

Security and compliance considerations

First—and this is important—the DeepSeek API does not meet our strict security requirements and is not currently available in regions outside of China.

In terms of speed, DeepSeek is much slower than other LLMs. And when it comes to cost, a user would need to self-host the model, given that it’s not available outside of China. This results in costs that would be as much as 100 times the cost of other available third-party LLMs on the market.

As mentioned above, Verint solutions work across 50 different languages, but DeepSeek is predominantly trained in Chinese and English. This is far less than Verint’s current linguistic capabilities, as well as those of other available third-party LLMs.

Will Verint use DeepSeek for customer engagement AI?

So, from this brief review of just a few of our enterprise use criteria, DeepSeek-R1 does not appear to offer us anything substantially better than what we currently use, even if the model may excel in public benchmarks.

We may yet find some low volume, internal use cases for it, such as automating prompt refinement or model-as-a-judge uses. However, for powering our bots in high-volume, low-latency contact center use cases, it doesn’t appear to be a good fit.

To learn more about AI research from Ian Beaver and the rest of his team, see our Verint Da Vinci research page.