Generative AI (Gen AI) swept through the telecoms sector with the force of a wildfire in 2023, and US operator AT&T was among the first to grasp its potential for a company-wide transformation.

The operator launched its Ask AT&T tool in June using an early version of OpenAI’s ChatGPT. More than 68,000 employees are now using Ask AT&T for tasks such as writing code, assisting customers, summarising meetings, patching security vulnerabilities and analysing the vast amount of data that crosses its network daily.

“We’re not anywhere close to hitting the barrier of what’s possible here,” Andy Markus (pictured), chief data officer for AT&T, told Mobile World Live (MWL). “Things are moving fast, but in the telco space I think we’re moving faster than most.”

Markus noted AT&T has long leaned into AI and ML to enable automation across its network and for use cases such as planning, administration and configuration, “but generative AI just allows us to do more in a way that a human would, and not in the way a machine would”.

Drafting the GenAI blueprint
With the realisation that Gen AI large language models (LLMs) have the potential to create incorrect results, or hallucinations, based on false information that is gathered, AT&T crafted its AI guidance principles.

One of the other primary concerns over the use of generative AI is the lack of clarity about which source code or open language is being used as information is gathered, which could lead to LLMs accessing an author’s or company’s intellectual property without their permission.

AT&T created its own set of core principles for the use of GenAI, which Markus said mirrored similar guidelines issued by the US federal government and other entities.

“There are foundational aspects that you have to have,” he stated. “We’ve centered our policy around those. If you follow those, you’re in good shape.”

The operator’s AI policy and operating guidelines were designed to educate employees on AT&T’s overall strategy while also establishing guardrails for how it’s developed, deployed and executed.

Markus said one of the core elements of the guidelines was having a human in the loop at every step of the GenAI process. While it’s a great facilitator for tasks such as coding or operating the network, the human element keeps hallucinations in check.

“We think it’s [GenAI] a huge facilitator, a great productivity enhancer and something that allows us to do things faster and quicker, but there’s always a person in the process,” he said.

As for IP concerns, AT&T made the decision to keep everything in-house for its Ask AT&T tool by training the LLMs on its own vast amount of internal data.

“We collect the telemetry. We evaluate it to make sure that sensitive and personal information isn’t included into the query or the output and we check for plagiarism,” Markus explained.

With its GenAI tool, AT&T has the ability to create actions based off its knowledge corpus from documents, policies and procedures.

“If we’re asking an HR question that is spun up through the API that we’ve connected to the HR interface that our employees use, we ensure that the generative function is only polling from our policies and procedures,” Markus said. “In other words, it’s not going to answer based on HR information that might diverge from our training. We ensure that it’s going to answer from our information.”

If the information is code-based, Markus stated AT&T checks to make sure it’s not copy written or has copyleft licencing obligations. All of those checks and balances happen through AT&T’s own APIs.

To further shield the company, AT&T requires vendors to use its APIs that are embedded in the Ask AT&T tool.

“We were very concerned that AT&T’s intellectual property would be leaked out in the public domain,” Markus noted. “If not, our 5G plans could end up in the public domain because these models are neural networks that learn with the information they see. Our data would be everywhere.”

Markus said AT&T tells vendor partners to use its APIs because “we’re further along than you are, in many cases, and we’re really concerned about you using a process that’s not as controlled as ours”.

The operator worked with Microsoft to make Ask AT&T secure and safe for its employees and corporate data.

Because AT&T has a heavy focus on interoperability, it has collaborated with Nvidia, Microsoft’s Azure OpenAI and used open-source models such as Meta’s LLaMA 2 and Falcon for Ask AT&T. AT&T employees also went to the San Francisco Bay area to meet with Google DeepMind, Microsoft and Nvidia employees as well as Marc Andreessen.

“I want to make sure we’re not in an echo chamber so that we’re not hearing what we want to hear because we’re close to it,” Markus said. “We’re moving fast and we want to talk to people who are moving fast with us.”

Evolution of Ask AT&T
While AT&T started out with the earlier version of ChatGPT, Ask AT&T has evolved with the various iterations since OpenAI launched in November of last year. But with the addition of LLaMA 2, Falcon and additional language models, the operator is now able to take a varied approach.

“Instead of just applying one foundational model to the problem, we’ll apply several and see if we get a better result,” Markus said. “We might use a Falcon model with a LLaMA model.”

Markus stated another benefit of the open-source models is they’re a lot cheaper than OpenAI. AT&T now has the capability to spin up a foundational model for specific use cases such as coding.

“Sometimes a good answer doesn’t need to be the most verbose answer,” he said.

Building use cases
On top of Ask AT&T, Markus stated the operator is developing an “ask data” tool to create additional functionalities. It provides subject matter experts with the ability to interrogate their data “with human language being the new computer language,” according to Markus.

He cited examples such as a structured query language that runs against a table to determine the number of potential subscriber additions for the company’s fibre-based broadband service based on market segments and then project those results over the next four quarters.

“Tell me the reasons why that forecast might be wrong or right? It’s taking the coding component, the ability to logic and reason, and applying that to data,” he said. “It really can supercharge that subject matter expert in the business. They know their business, but they may not be great coders, and they might not have a data scientist by their side.”

For the network, Ask AT&T is used to interrogate documents and data to create new information. That type of classification of a problem can be applied to the network for proactive issue identification and management.

 “It’s helping our network engineering team understand how to act quicker, and in some cases, before the issue actually occurs,” Markus noted. “We do some of that today using traditional AI and traditional methods, but we’re seeing the possibility that this is going to really help us at a different scale.”

The tool also gives AT&T a more detailed look at root cause analysis (RCA), which has been traditionally a very manual, human focused process.

“Generative AI can learn from all of our tickets and all of our previous RCAs,” Markus stated.

As open-source GenAI models have matured, AT&T is now starting to build its own custom models for its network use cases. Markus stated the initial cost of ChatGPT meant it wasn’t always cost-effective to create its own customer models.

“It had to be the killer use case for us to invest that type of funding to build our own custom models,” he said. “The ability to create our own custom models has become a very real possibility, so we’re testing that out at AT&T.”

Markus said the summarisation of image components by GenAI is also of interest for use cases such as customers or field technicians taking pictures of an issue or problem that could then be resolved with an automated process.

“Understanding the images and quickly taking action on them is a really big set of use cases that GenAI can help us with,” he stated. “That’s an area we’re heavily looking into.”