In-house and AI roundtable: 'It feels like we got on a moving train'

The Gazette roundtable on in-house lawyers and AI brought together counsel from a broad range of sectors, from university publishing to real estate and heavy engineering, to discuss how they are using AI to support and leverage their legal teams. AI deployment also means establishing guardrails for the use of generative AI (GenAI) models, as well as managing risks around output validation and controlling data retention and model training. There are also bigger issues around sustainability, legal privilege, and the cost and value of legal services. 

The promise of AI for general counsel is that it speeds up routine tasks and processes, freeing up GCs and corporate legal teams for strategic work, and interacting with their internal stakeholders and external law firms. Popular applications include enterprise licences for generalist models, such as ChatGPT and Microsoft Copilot; and legal-specific tools such as Thomson Reuters CoCounsel, and point solutions for tasks like notetaking and compliance/verification. And this year has seen chatbots evolve into increasingly capable AI agents. 

However, when legal departments introduce AI into their processes, they need their internal clients to accept new ways of interacting with legal, and their AI systems to integrate with existing IT systems and applications. And because they are dealing with sensitive company data, it is critical to establish appropriate guidelines, governance and controls for using AI responsibly. AI in corporate legal also raises familiar considerations concerning the accuracy/validity of output, managing sustainability and whether it is or should be replacing junior roles. 

'The management consultants have enjoyed using Ask Elisabeth – they feel they can ask questions 24/7 and get an answer which they can then validate with my team'

Elisabeth Sullivan, Efficio Consulting

Rebecca St Requier, vice president Prime Services Legal and co-chair at University of Law London Alumni Network, highlights the rapid evolution and adoption of GenAI. ‘There’s been a huge change in the landscape in the last three years. For me, [AI has] changed from being a clunky tool that I was unimpressed by to a useful tool that I’m embracing. I see it as something that improves the quality of life of people working in my space by relieving us of the more boring tasks.’ She notes that ‘within the derivatives space it is known that in terms of collateral management, AI can detect where the issues are for a collateral manager – for example, first thing in the morning. So they can go straight into the key points, rather than having to, laboriously, work their way through’. St Requier adds that ‘even with issues such as dispute resolution over collateral, and eligible collateral, AI can actually manage a raft of those points before requiring human intervention’. 

Elisabeth Sullivan is group general counsel and company secretary at Efficio Consulting, a procurement and supply chain consultancy which employs about 1,500 people. A couple of years ago she launched an AI chatbot, Ask Elisabeth, which ‘had all our policies, our playbook, any FAQ I was keeping in our legal SharePoint. People could put in a document and effectively check it against company policies and obligations’. She observes that ‘the management consultants have enjoyed using Ask Elisabeth – they feel they can ask questions 24/7 and get an answer which they can then validate with my team’. 

This year, following an RFP (request for proposal) involving 10 legal technology providers, Efficio invested in an AI tool for its legal team. ‘We are also trying to understand how Copilot can help with administrative tasks. We also have a Copilot add-on, which means you can “redline” in Word using Copilot. 

‘I have been trying to think about how I can push my business to self-serve, so they can do their own mark-ups. We also have an AI agent in ChatGPT doing horizon scanning with legislation. We are in 15 countries, [so it can identify] what new legislation topics we need to be aware of,’ explains Sullivan, adding that the prompt is updated regularly.

KONE’s AI journey is similar. Legal director Brenda Albert says: ‘We have had Copilot for a while, across the business, and some of us have the Enterprise licence, so we’re able to create our own agents. Within the legal team, we tested a lot of products over the last three or four years. Copilot’s usability is better because it’s already integrated into Microsoft 365.’

Albert is currently building an AI agent using Copilot that she has named Connie; she is working with a third party to build out the agentic processes. The idea is to upload playbooks so that Connie can respond to questions and review documents in the same way as Ask Elisabeth. 

Ultimately, the plan is for ‘Connie to be able to… review documents, apply our playbooks, and eventually, I want her to integrate into other areas too', Albert says. ‘Copilot has a lot of add-ons. It can [integrate] into Salesforce and… into Docusign. So it’s just mining all of that, so that you end up with an agentic AI that is like having a senior paralegal and an EA mixed together.

‘That is my dream – that we should be utilising AI to get rid of the repetitive tasks we don’t need to be doing and free ourselves and our teams [to be] more strategic, and to find more time to [meet] our stakeholders face to face.’

Managing agents

'We also ask our external counsel which AI tools they use. Transparency is important, especially if you want to get a sense-check on something'

Michael Harwood-Smith, Marshall Wace

While corporate legal teams are not regulated to the same extent as professional services, they need to establish effective AI governance and controls, and understand how their external law firms use GenAI. 

Michael Harwood-Smith is a lawyer at Marshall Wace, a leading provider of alternative investment solutions which has used quantitive and statistical models for over 20 years. ‘We look at processes and workflows that sit in different teams to identify which parts of them can be enhanced or sped up by AI. We have rolled out agentic workflows, and we focus on training and governance.’

While Marshall Wace handles most of its legal work in-house, it instructs outside counsel for specific issues. ‘We also ask our external counsel which AI tools they use,’ he says. ‘We think transparency is important, especially as a client if you use AI to produce a draft and you want external counsel to review the output, or you have a document-heavy task that you want external counsel to look at.’

General counsel Catie Sheret explains that ‘while Cambridge University Press & Assessment is embracing AI, it is quite risk-averse because alongside its publishing activities, it is responsible for delivering high-stakes exams such as those for [examination board] Cambridge OCR’. It is particularly exploring the use of AI in three areas: auto-marking, customer services and content generation. The 50-strong legal team operates a sophisticated AI governance set-up, which combines policy rules, controls and ethical principles. 

Sheret sees potential in both legal and compliance, ‘replicating the risk assessment process across sanctions, data privacy and even AI governance itself’, Sheret continues. ‘We’ve got a centre of excellence and quite an involved declaration of interest process, and last year we got the AI centre of excellence to have us build a chatbot to help people through the process. This year, we built [another chatbot] to do all the reporting on who in their teams hadn’t done the tasks they were supposed to do for the year, pulling data from various different systems.’ 

Experienced financial services and insurance general counsel Jo Nayler has used AI to check whether people across the business have done all their training by running an automated report. And several attendees use AI tools, including Copilot, to produce reports and meeting minutes, although there is some concern around ensuring that confidential information is deleted permanently. 

Data and guardrails

This led to a discussion about the risk of AI models retaining residual knowledge. Harwood-Smith at Marshall Wace notes that ChatGPT Enterprise supports compliance requirements for regulated industries as customer prompts and company data are not used to train models. He and others 'always request zero data retention, which stops the model provider being able to reuse or even retain the data.'

The dilemma is that the more the model knows about an organisation, the more useful it is. St Requier recounts a similar experience while organising an AI moot at the Royal Courts of Justice, where a ChatGPT bot was trained to defend a father’s custody rights against human lawyers representing the mother. ‘Aruna Verma, campus director Moorgate of University of Law, was very keen to ensure fairness by making sure the model did not recollect any residual knowledge of the data that it was fed during testing and preparation, as well as verifying outputs to check for hallucinations,’ St Requier recalls. 

‘But we did have to feed into the bot all of the rules of the court, how to address [people], and the codes of conduct, how to refer to the judge. However, this [residual knowledge] could be a desirability factor for in-house legal because as the bot starts to understand the needs of the business, you have to give it fewer instructions.’ 

Tamara Ekanayake from Thomson Reuters is interested in whether regulatory compliance is factored into AI implementation in in-house legal departments. While all attendees focused on providing appropriate training and guidance while maintaining strict controls over AI models’ access to data, there are concerns about shadow AI, whereby people use unlicensed or public AI models for work purposes. And as AI enables people to build their own solutions, it is opening up new skillsets and capabilities which bring new risks. The response is to introduce effective controls and guardrails around the use of AI. 

Albert draws an analogy with KONE’s engineering equipment, and the need to monitor the data relating to new lifts and escalators. This is even more important for GenAI because adoption was relatively sudden. ‘It feels like we got on a moving train, and we’re building the guardrails as we go… we adopted this learn-as-we-go [approach] and now we have Copilot with guardrails – there’s a work version and an external version – we effectively have our own search engine.’ 

The company’s AI use and the training it provides continue to evolve: ‘We learn fast, we fail fast, and we fix fast. There are pros and cons, but we are definitely learning fast and adapting fast.’

AI’s environmental and human impact

While AI efficiencies free up general counsel and corporate legal teams for more strategic work and to interact with the business, AI adoption also has negative implications, notably the impact on the environment and the risk to junior roles.

Sheret is involved in discussions about AI and sustainability: ‘I’m on the Law Society’s Climate Change Committee, and we’re thinking about some guidance, for in-house, on sustainability and AI. We’ve got a digital sustainability team at Cambridge and are doing some quite groundbreaking work.’

She continues: ‘If we, as in-house counsel, encourage our organisations to ask the right questions, in the same way we do about other environmental issues, we can start to push [big tech] to be more responsible. And generally, we should all be more thoughtful about our use of AI. So not using it to outsource what we should be thinking about ourselves, not using it for things that should be automated but don’t need AI – just a better process. So we’re not using it because it’s a lazy, easy option. I think token-based pricing will also drive more thoughtful use.’ 

In addition to using secondments and trainees to create a more joined-up approach to AI, there are additional opportunities for in-house counsel to collaborate with their external law firms, for example regarding AI procurement as major vendors move to token pricing and the legal AI market becomes more crowded.

This is not just about pricing or passing on savings from AI efficiency. Increasingly, knowledge and guidance about AI products are considered part of external firms’ value-add. Firms know this, of course. Nayler references Sidley Austin becoming the first firm to introduce a mandatory training seat in its AI lab. 

While cost-effectiveness is an important benefit of AI, this is likely to be affected by the introduction of token-based pricing and the rising cost of tokens. This ‘has been a big topic for my company, a procurement consultancy, both for ourselves and for our clients, and how we can try to get that under control’, says Sullivan. 

St Requier brings the issue closer to home. ‘When we’re looking at a purchase order for outsourcing to private practice, in-house legal counsel will want to know how the external law firm is using AI, to make sure that we’re getting value for money from that service provider, in the same way as an in-house [legal] function ultimately serves the business,’ she says.  

Finally, what progress are in-house counsel hoping will be made in the next year? The two main themes that emerge are value for money and improved clarity concerning the implications of AI adoption for legal, society and the environment. 

Nayler and Sullivan focus on cost-effectiveness. ‘I would like to come back in a year and feel like AI efficiencies have been passed on in terms of cost, in a material way,’ says Sullivan. Ekanayake at Thomson Reuters is looking for AI to add value for all stakeholders, not just senior leaders. 

St Requier would like to see ‘increased trust and confidence replacing current uncertainty’. Sheret too is looking for certainty in respect of legal privilege and the status of AI outputs. ‘Can we have absolute certainty on what kind of technical architectural structure is required for material to stay privileged in a tool? At the moment, we are all guessing,’ she says. This is a sensible aspiration, given concerns about shadow IT. If case details are uploaded to a public AI model like ChatGPT or Claude, they lose legal privilege. And failure to verify AI outputs has implications for case outcomes and professional indemnity insurance. 

Harwood-Smith brings the conversation back to corporate legal, hoping for 'more transparency from law firms about how they are using AI, whether it's through training and use of AI within the legal teams, or whether it's through embedded legal technology teams that work alongside lawyers.’ All these aspirations are indications that the legal AI market is maturing. Since AI is now an established element of the legal tech stack, firms and in-house legal departments need to ensure that the right guardrails and governance are present. 

Joanna Goodman is a freelance journalist

This roundtable was kindly sponsored by