Mastering general counsel's dual generative AI role

General counsel are playing a two-tiered role as companies explore how to deploy generative AI in their business and support functions, Kevin Clem, chief growth officer at HBR Consulting, says. 

Since generative AI exploded on the scene last year as an add-on to existing AI-assisted technology functions, CEOs have been asking their C-suite executives to explore use cases for the technology.

As part of those marching orders, general counsel and chief legal officers are looking for ways to deploy the technology in their legal-team processes.

But they’re also taking a higher-level view as they look from a risk-management perspective at what types of uses, by any business or support function in the organization, are likely to be okay, and which aren’t.

Additionally, legal chiefs are examining what policies and guardrails are needed for the use cases that eventually get deployed.

“There’s an expectation that the GC or the legal team will weigh in on acceptable use and set appropriate guardrails,” Clem told Legal Dive. “They may not get down to the level of actually reviewing use-case by use-case examples, because there are going to be hundreds if not thousands of use cases in an organization. So, the expectation isn’t that every potential use case is explored by the legal team on a substantive level, but that the team will help define the parameters around which generative AI can and should be used from a policy standpoint.”

Support functions

An example of a non-legal use case that companies are looking at, Clem said, is in the HR function.

Rather than have someone on the HR team answer common employee questions, that function would be handed over to a generative AI tool that’s been trained on and draws from the company’s employee policy manual.

One company working with HBR is thinking about training its generative AI tool on its 70-page employee policy manual. Should it follow through and do that, he said, the tool would keep its answers within the parameters of that manual. 

“You’re not looking at the totality of the data that’s out in the large language model,” Clem said. “It’s pointing to a source that’s relatively contained.”

Another example is how a company’s communications team might use the tool. Rather than have a public relations person write a first draft of a press release, the tool would be deployed for that purpose and the PR person would come in later to refine it. 

Risk watch

These kinds of use cases are what Clem considers examples of relatively safe experiments of the new technology, which trains on and draws from natural language databases to create content that’s expressed in the same way a person would in a document. By creating these initial document drafts, it can free up staff for more value-added work. 

Risk is high on the minds of in-house counsel and company leadership after a well-publicized mishap in early June by an attorney who used a generative AI tool to conduct research for a court filing that cited half a dozen made-up cases. 

Analysts say the technology will get better as it's further developed. But for now, the best practice is to find relatively safe use cases to experiment with, which means looking for applications in which the tool trains on, and draws from, trusted and self-contained data sources.

“You want to make sure it’s answering correctly,” Clem said. “So, run some experiments within the walls you’re going to control before opening up more broadly.”

Legal function

For the legal department, it can make sense to experiment with using the technology first in non-legal functions, like training attorneys on the use of a new tool or workflow. 

If you bring in a new document management or e-discovery system, for example, you can use generative AI to write the first draft of the training manual. That frees up staff time while keeping the experiment within a relatively safe space. 

“Someone isn’t sitting down and writing the material from whole cloth or trying to adapt from a prior training document,” Clem said. “So, that’s an example where you’re rolling out new technology, leveraging generative AI to help with a new training curriculum. It’s not going to be perfect, but if you can get a good head start, why not leverage generative AI to help draft the script?”

You can even mix and match AI assisted tools to free up more staff time, he said.  

A project HBR is working on is combining a generative AI-drafted training script with an AI-assisted avatar that will conduct the actual training. Should this be deployed as planned, involvement by a staff person will be minimal, because the script will be largely written automatically and the presentation, once it’s refined, will be automated. What’s more, for a global operation, the training can be conducted in whatever language is needed. 

“It’s stitching together generative AI with more traditional avatar-based AI training technology that can read a script in whatever language you want it deployed and you’re now deploying a training around a new technology that is entirely based on AI,” he said.

That kind of training can apply to any function within an organization, including legal, which could use it for training on new tools or procedures. 

On the legal side, it’s best to stick with trusted, long-validated sources, like Westlaw and LexisNexis, for legal writing. Both Thomson Reuters and LexisNexis are developing tools incorporating the new technology such as Lexis+ AI.

“Your drawing from reputable sources,” Clem said. “These have been used for legal research for decades. If it’s just drawing on the data you know is contained, you’re at less of a risk that the technology is going to create hallucinations or invent content.” 

Bottom line, as CEOs ask their executive team to look at possible generative AI use cases, in-house legal teams can expect to be called on to formulate company-wide risk-management-based policy on the use of the new technology, while at the same time, the team will look for possible use cases within its own operation.

Niki Armstrong, the chief legal and compliance officer at Pure Storage, says generative AI has the potential to help legal teams “fundamentally change the way they think and operate.”

As a result, her legal department is testing out use cases, including in the area of regulatory compliance support.

The publicly traded data storage technology company is seeing if generative AI can help it monitor changes to the many different laws and regulations that govern its activities.

For example, Armstrong said, her compliance team has started testing whether the latest artificial intelligence can provide useful information about updates to privacy and employment laws in different countries.

The team is also examining whether generative AI can assist with tracking pertinent new regulations for tech companies issued by the U.S. Securities and Exchange Commission.

This type of information helps Armstrong’s compliance unit determine whether it needs to update internal policies or create new training materials for employees.

“We are using generative AI to help us better understand how regulatory changes impact us so that we can both form and develop content that is useful and applicable to us, while maintaining ethically sound and legally compliant practices — a strategic priority for our team,” Armstrong told Legal Dive.

Additionally, she is hopeful generative AI can assist Pure Storage’s legal team with automating repetitive tasks and addressing potential compliance issues.

Outside counsel 

Armstrong said Pure Storage could take the information produced by generative AI about legal and regulatory changes and share it with outside counsel to review for accuracy.

This approach could produce cost savings for the legal department because it is not asking its outside law firms to start their research from scratch.

“We're almost cutting out a little bit of the middleman where we're doing some more initial research on our own and then going to outside counsel to say, 'This is what we have found. Can you review and confirm?’” Armstrong said.

Some law firms have already begun using generative AI, reports indicate, and Armstrong said Pure Storage’s firms may be among them.

A member of Pure Storage’s legal ops team is helping lead an effort to determine whether and how the company’s outside law firms are using the latest AI.

Armstrong said she expects law firms to closely review any work product produced by AI to ensure there are no mishaps like the ChatGPT fake cases scandal out of New York. The two lawyers involved were fined by a federal judge after one of them used OpenAI’s chatbot for legal research without independently verifying its work.

“Certainly, we've seen some of the debacles that have happened with attorneys who have relied on ChatGPT,” Armstrong said. “I can't have inaccurate, unreliable or flat-out false information coming from our trusted partners in these law firms.”

Knowledge management

Meanwhile, Pure Storage’s legal team also hopes to use generative AI in the knowledge management context.

Armstrong said that once information about legal and regulatory changes produced by an AI tool is verified either externally or internally, Pure Storage aims to develop knowledge base articles it can reference moving forward.

For example, an internal knowledge base article could provide information about key employment law changes in countries in which Pure Storage has many employees.

“If generative AI could actually help us draft those knowledge management articles, I think that could be huge,” Armstrong said.

Pure Storage’s in-house legal department is also considering potential document automation use cases for generative AI.

But Armstrong said that before moving ahead in that manner, she wants to make sure the deployment of AI will speed up work rather than create more of it.

If the latest AI proves effective with document automation, Armstrong said, this could free up her legal team to spend more time on higher-level strategic work.

The recent advances in artificial intelligence have prompted interest across Airbnb in how employees can use technology to enhance their work processes, according to Iris Chen, a vice president and deputy general counsel at the company.

She said Airbnb’s legal department has responded by creating a small working group that is testing out AI products for potential use. The unit is serving as a focus group of sorts for the in-house legal team.

“We're trying to see if this actually has an impact before we go off and say, ‘Everybody's got to adopt this,’” Chen said. “I think that's been a really smart way that we've thought about, ‘Okay, how are we going to start integrating some of this really cool technology into our daily practice?’”

Chen made her remarks during a recent webinar focused on digital transformation that was hosted by Ironclad, an AI-powered contract lifecycle management provider.

Chen said one area where the Airbnb legal department is experimenting with AI is a tool that can help the litigation team write initial briefs. She said she was excited about the potential positive impact.

“That's going to be amazing if that [tool] actually can reduce the turnaround time that it takes for the litigation team to produce memos or briefs,” Chen said. 

Along those lines, she highlighted that when testing out new technology, it is helpful to start in an area where there is something that can be measured.

Top-down support

Chen said the company’s AI exploration has been a top-down initiative, and she shared that the legal team’s working group has benefited from the involvement of the chief legal officer.

Even though the CLO is retiring, Chen said, he has participated in focus groups regarding the use of AI in legal and asked important questions about the broader implications of the emerging technology.

“He has been super excited about the possibilities and encouraged us to experiment,” she said.

Airbnb recently announced the company’s outgoing legal chief, Rich Baer, will be replaced by former White House Chief of Staff Ron Klain. He will start at Airbnb on Jan. 1 and will report to Airbnb CEO & co-founder Brian Chesky.

Advisory work

Meanwhile, Chen said the AI experimentation has also served the legal team well in its work advising other parts of the business about the emerging tools.

Hands-on experience with the technology allows legal professionals to understand AI’s benefits and limitations, as well as the risks AI can introduce, she said. 

This knowledge helps the in-house legal team provide thoughtful advice to colleagues. 

“I think having legal be adopters and helping the business evaluate these tools is actually a really good thing,” Chen said.

Legal software company Luminance hosted reporters in its London office this week to show its AI tool called Autopilot negotiating a nondisclosure agreement with a customer without involving lawyers in the process except for sign-off. 

When a laptop belonging to Luminance General Counsel Harry Borovick received an email with an NDA attached as a Microsoft Word document, Autopilot opened, scanned and started making changes to the document without input from any person, according to accounts of the demonstration by CNBC and BBC. 

“A six-year term is unacceptable, so it's changed to three years,” says the BBC report, describing the back-and-forth between the two laptops.  

The software flags risky clauses and automatically makes changes based on what Luminance says is 150 million legal documents that the software has been trained on.

“[One] risky clause imposes an unlimited liability, meaning there's no ceiling on how much Luminance might have to pay if the terms of the NDA were breached,” the BBC report says. So, the software replaced the clause with a proposed liability cap of £1 million and removed a hold harmless clause that “the AI knew ... wasn’t okay.” 

The laptop belonging to the counterparty, Connagh McCormick, the general counsel of research company ProSapient, opened the revised contract and inserted a liquidated damages provision in response to the removal of the hold harmless provision. “That effectively turns the £1 million maximum liability into an agreed compensation to pay if the agreement is broken,” the BBC report says.  

The laptop belonging to Luminance’s GC replaced the liquidated damages provision with language stating the company is only liable for direct losses incurred.

“Version four of the contract is acceptable to both parties,” the BBC report says, and the AI belonging to Luminance’s GC “accepts all the changes and sends it to DocuSign” for signing.

“At that point you get to decide whether we actually want the human to sign,” Jaeger Glucina, a managing director at Luminance, told reporters at the demo. “That would be literally the only part that the human would have to do. We have a mutually agreed contract that's been entirely negotiated by AI.”

The company says its tool is trained entirely on legal documents, not on content from the public internet like ChatGPT. That could reduce exposure to incorrect, private or protected information that has some AI critics concerned.     

To teach the AI what contract terms are and aren’t standard for the user, what terms are negotiable and what terms are unacceptable, companies create what Luminance calls a knowledge bank based on their signed documents.

“The AI [is] not only legally trained … but also understands your business,” said Glucina in CNBC’s report. 

Lawyers have the final say. “It ends up in the laps of both lawyers signing,” said McCormick. “At that point, I'm going to read the contract and the other lawyer's going to read the contract. If there's anything I disagree with, I've got the opportunity to flag it. I'm not committed to anything the AI has done.”

Autopilot isn’t out yet; it’s slated for beta testing before the end of the year with plans for a 2024 rollout.

If it catches on, said McCormick, “a lawyer's time is going to be spent doing something more interesting, more valuable” than negotiating NDAs. 

General counsel tasked with developing generative AI policies for their organizations should start by working to identify “must-avoid outcomes,” according to recent guidance from Gartner.

These potential outcomes can be determined through risk tolerance discussions with the organization’s senior management about the possible applications of generative AI models within the business.

Additionally, legal chiefs should discuss with other leaders which potential end results feature acceptable risks given the anticipated benefits of emerging artificial intelligence such as improved efficiency. 

“Guidance on using generative AI requires core components to minimize risks while also providing opportunities for employees to experiment with and use applications as they evolve,” said Laura Cohn, senior principal, research at Gartner, in a prepared statement.

Use cases and safeguards

Gartner recommends legal leaders organize a list of possible generative AI use cases by perceived risk. This includes determining the likelihood and severity of the risks.

Common risks associated with generative AI usage include cybersecurity, data privacy and IP issues.

General counsel can help determine what types of controls to implement for various use cases based on their risk levels. 

For lower-risk use cases, such as translating regulations written in a foreign language, GCs could require human review.

Meanwhile for higher-risk use cases, such as producing content for customer consumption, approval from an AI committee or manager could be required. 

“General counsel should not be overly restrictive when crafting policy,” Cohn said. “Leaders can consider defining low risk, acceptable use cases directly into policy, as well as employee obligations and restrictions on certain uses, to provide more clarity and reduce the risk of misuse.”

Additionally, outright AI prohibition could be considered for the highest-risk cases, Gartner suggests.

Decision rights

GCs also have an important role to play in determining who has the authority to make decisions regarding various generative AI use cases.

These determinations should be made in tandem with executive leadership. 

Once such decisions are made, in-house legal teams must work with other business units to review their AI duties and risk ownership. 

This approach can include highlighting which department or departments that employees should turn to when seeking the go-ahead for AI use cases requiring approval.

“Document the enterprise unit that governs the use of AI so that employees know to whom they should reach out with questions,” Cohn said.

Disclosures

Gartner recommends organizations have a policy of disclosing the use and monitoring of generative AI technologies to both internal and external stakeholders.

General counsel are advised to help companies decide what to disclose and with whom.

These disclosures could include labeling text that is machine generated and placing watermarks in AI-generated images to the extent possible.

“Consumers want to know if companies are using generative AI applications to craft corporate messages, whether the information appears on a public website, social channel, or app,” Cohn said.

Overall, she said generative AI policies and guardrails “will better prepare enterprises for possible future legal requirements.”

Leading legal chiefs say they recognize AI has the potential to rapidly transform their industries, so they are supportive of employees testing out the latest technology for work purposes.

But in order to ensure the exploration of emerging artificial intelligence occurs in a manner that minimizes risk, in-house legal departments are playing a key role in developing AI acceptable use policies.

These corporate AI policies should start by emphasizing basic dos and don’ts, three legal chiefs said during an AI-focused webinar hosted by contract lifecycle management provider Ironclad. 

Dana Wagner, chief legal officer at customer engagement platform Twilio, said it’s important to highlight which AI-powered platforms the company has relationships with and are safe for employees to use.

It also helps to make clear there are differences between putting publicly available data into a public AI model that will use it for training purposes versus inserting proprietary data into proprietary models that are not training on the data.

“You want the experimentation; you just want it to happen in a responsible way,” Wagner said.

Alan Tse, global chief legal officer at commercial real estate company JLL, agreed that emphasizing the need to be careful about inputting confidential information into various AI tools is essential.

He also said it is important to remind employees that just like if they were using Google to find content such as a photo, they need to make sure they have the legal authorization to use images or other information produced by AI-powered technology.

This guidance is necessary because generative AI tools such as ChatGPT can quickly produce text, images and video in response to user prompts.

Along those lines, AI policies should require that outputs from AI-powered platforms are reviewed by humans for accuracy, said Eleanor Lacey, general counsel at work management platform Asana. 

This approach can help prevent disasters that could harm a business, such as a lawyer citing fake cases produced by a generative AI tool. 

Collaboration

During the webinar moderated by Ironclad Chief Community Officer Mary O’Carroll, the legal chiefs also shared that collaboration with other business units is vital when crafting corporate AI policies.

Lacey said Asana’s legal team has worked with the company’s security and engineering units on AI-related policies.

For example, she said bouncing questions off her engineering colleagues about AI has enabled her to learn more about the emerging technology and informed policy drafting.

“I go to the engineers and I say, ‘This is the thing that I want to understand more,’” Lacey said. “And they're like, ‘No, you think you want to understand that, it's actually this other thing that you don't even know exists. And now we're going to educate you on it.’ 

“I think having that back and forth is incredibly helpful as a lawyer because I'm not also a coder,” Lacey continued. “It's helping me learn and better identify where I think the risks are.”

Wagner said legal departments also benefit from asking executives and leaders of other business units how the company is planning to use AI.

This level of specificity can assist with crafting AI policies and principles that are more concrete and useful, he said.

Policy iteration

The legal chiefs said companies should be prepared to update their policies as AI technology develops further and internal use cases change.

“We’re not trying to get it perfect,” Tse said. “We know that we're going to have to move a little bit as we learn.”

The legal leaders said this approach is a much better option than companies instituting AI, or generative AI, bans.

Tse said employees would still end up experimenting with AI, and they would do so without guardrails in place.

“They are still going to use it, they just won’t tell us how they are using it and we can’t help,” he said.

In that vein, a recent Asana survey of knowledge workers found 48% want more guidance from their employers on how to use AI.

Creating first drafts of contracts and flagging risks as you go back-and-forth with your counterparty are two of the most common jobs in-house counsel are putting generative AI tools to use for, Pedram Abrari, CTO of contract lifecycle management company Pramata, said in a webcast.  

Other uses are reviewing, summarizing and pulling data insights out of contracts, Teju Deshpande, principal of Legal Business Services at Deloitte, said in the webcast, hosted by Pramata. 

The emerging AI tools can be especially useful to in-house counsel if they’re pressed for time to weigh in on a rush of contracts at the end of a period because they can cut the amount of time spent on each contract from hours to something much shorter. 

“You have emails, Slack messages, and so on,” said Abrari. “You go through and pick one of them to process. Then you have to find out what account it’s for, and if you have all the agreements for that account. Sometimes it’s a challenge to even find those agreements, and hopefully you have the best copy of each agreement.”

Then you have to review the master service agreement and any amendments and orders to get a good picture of the status of the account before you can add another amendment or order to it. And anything you add must be consistent with your company’s business objectives and your team’s contracting playbook. 

“You have to worry any type of agreement you put together is in compliance with that,” Abrari said. “It requires spending hours on each request under extreme time pressure. And since we’re all human, accuracy may take a hit.”

Smart assistant

Thinking that a generative AI tool will take all of that off your plate isn’t accurate, Abrari said. The more accurate way to think about it is like having a smart legal assistant or paralegal reviewing the agreements, amendments and orders, sorting through earlier and later versions of them to find the final ones and writing the first draft that you as in-house counsel can then refine.

“It can help you get through these contract requests more rapidly and with more accuracy in less time,” Abrari said.   

Generative AI can also help you align your contracting playbook with your negotiation strategy and your corporate objectives, so that if your counterparty wants a term or price that doesn’t match your playbook, it will flag it so you can seek to change it in negotiations. Or, if you can’t get the other side to agree to do that, it will let you know who needs to sign off on the increased risk you’re embedding in the new document.

“For example, for your payment guidelines, if a payment is made in more than 45 days, that’s high risk and requires finance desk approval,” Abrari said. “For price increases, if none are allowed, that’s a high risk and requires vice president of finance approval, and so on.” 

Clean data

The way to prevent a generative AI tool from producing incorrect responses to prompts or otherwise increasing your risk is to ensure it’s only drawing from documents that are current and accurate, following good governance instructions and that your prompts are well-written. 

“Prompting is a new skill set,” said Deshpande. “So, building good prompts is a skill that needs to be developed.”

Companies offering generative AI tools typically include pre-programmed prompts, but it’s the nature of the tools that users can create their own prompts. However, that doesn’t mean the tools will respond to new prompts equally well if they’re not crafted with an understanding of how the tools work. 

“I can’t underscore how important prompt engineering is,” said Deshpande. “It’s the highest and best use of what you’re trying to achieve.”

Even if you do everything right, there remains a risk the tool will output an incorrect answer or hallucinate or, depending on the modeling the tool is based on, provide an answer that includes copyrighted work.

“It’s still early days,” said Deshpande. 

But there’s a risk to not using the tools as well, the legal tech specialists said.

“If your policy is not to use it, people will start using it on the side,” said Abrari. “The cat’s out of the bag. Everybody knows that if you have to use genAI to get ahead, and my company’s telling me not to use it, I’m going to use it anyway on my personal computer. That’s where it starts getting risky … Having a genAI model that is vetted for security, available to your organization, is probably lower risk than not doing that.”