GenAI finds its voice

A Harvard Business Review article titled ‘How people are really using GenAI in 2025’ by Marc Zao-Sanders identifies a shift from technical use cases to personal and professional applications. The evolution of agentic AI towards more emotive use cases is leading people to interact with large language models as digital colleagues rather than tech tools. This underpins a tentative shift to voice-first as a more intuitive GenAI interface. Just last week, OpenAI CEO Sam Altman responded to a question on X about energy use, stating that saying please and thank you to ChatGPT costs the company ‘tens of millions of dollars well spent’ in electricity bills. And ChatGPT’s Voice Mode is free to access – try it out, but remember not to share anything confidential!

Agentic AI is driving innovation

This shift from AI agents to agentic AI is reflected in the evolution of law firm GenAI tools. These range from internal ChatGPT-style chatbots with standard prompts for expediting specific tasks and processes, to agentic applications that manage entire workflows, and select the relevant application for each task. This has led to the rapid take-up of agentic AI platforms such as Harvey and Legora, which is expanding across Europe and the US. Incumbent legal tech vendors responded by introducing agentic AI capability into their product portfolio. This means that, to stay in the vanguard of GenAI, firms need to go further and boost innovation, through partnerships, acquisitions and incentives.  

Growth and consolidation are indicators of a maturing market, but create complexity for procurement. Although there is an embarrassment of choice when it comes to GenAI products for legal, 2025 has seen significant consolidation. This has led to dissatisfaction among early adopters of products developed by companies that were subsequently acquired. There have been issues around tech support, and discontinued products/capabilities.

Michael Kennedy, head of innovation and legal technology – research and development at Addleshaw Goddard, who is tasked with piloting and assessing the latest GenAI products, favours point solutions over catch-all offerings. ‘The offerings are better, and everyone is trying to do everything,’ he says. ‘I would rather they did different things well than try to cover all the bases. There’s a lot of overlap, and we will see more consolidation as people figure out what they need.’ Kennedy adds that, as large language models are getting better, suppliers are getting more customers, who are explaining what they want and how they work, which is in turn improving the products. ‘If they get feedback from, say, 50 customers,’ he notes, ‘they can make their product more specific to how lawyers actually work.’

Voice vibes

As agentic AI acts more like a digital colleague than a tech tool, the way people interact with it is also changing, and there is a tentative shift towards voice-first interaction.

While Addleshaw Goddard’s internal large language model AGPT does not have a voice interface yet, Copilot is widely used for recording and transcribing meetings, and generating meeting summaries and follow-up actions. ‘If you join a meeting late, you can ask Copilot what’s been discussed so far and it produces useful bullet points,’ says Kennedy. ‘It means you can focus on the meeting rather than on taking notes. I do both – I write down just the important points, and then I upload my notes and the Copilot notes into AGPT and ask it to create a merged outline of the meeting.’

When it comes to AGPT, ‘we are having conversations about accessibility for people who are visually impaired or who struggle with typing’, says Kennedy. He uses voice notes to gather his thoughts, ‘for example, to build a business case for technology I’m looking at buying. I use it for ideation when I’m not at my desk and to structure my thinking around a plan or report’.  

Peter Duffy, CEO of Titans legal tech and AI consultancy, believes that voice is ‘the new frontier’ for GenAI. Rather than reducing human interaction as people get used to working with AI agents and agentic systems, he sees voice as a more natural and rapid interface. It is quicker and less laborious than typing a query, and has the potential to transform prompt engineering.

Duffy used AI to develop a voice-vibes web app. This allows him to record voice notes, transcribe them and enrich them using ChatGPT by adding a structured reasoning prompt, as well as images and screenshots.

Another use case is delegating tasks – for example, recording a voice note instructing ChatGPT deep research to do market research, in the same way that you would brief a junior employee, without having to type in a detailed series of prompts. ‘All of a sudden it’s like a superpower for experts,’ he says, adding that ‘a lengthy voice note is quicker and provides so much more context than typing a series of prompts’.

While Duffy recognises that it is early days for GenAI voice interactions in a legal context (as distinct from voice commands), there is at least one trailblazer product. LexisNexis’ personalised voice AI assistant Protégé (currently only available in the US) includes voice prompts and advanced reasoning capabilities.

Duffy’s rationale for voice-first interactions with GenAI is the ability to intuitively share ‘tacit domain knowledge that isn’t codified anywhere. Yet getting that information out of your head and into the context window within AI can have a dramatic impact on the quality of the output’.

AI for knowledge amplification

Kingsley Napley is addressing precisely that tacit knowledge dilemma. It is partnering with legal tech start-up Let’s Think to develop The Knowledge Exchange, an AI knowledge amplification tool which uses behavioural science to facilitate tacit knowledge sharing.

For Sarah Harris, director of innovation and knowledge, the rationale was to replicate the supervisory model, whereby junior lawyers learn by working with experienced senior colleagues. That was broken by the shift to remote and hybrid working. Culturally, it replicates the value that the profession places on lawyers’ experience, in that their status and remuneration are measured by post-qualified experience (PQE), or the number of years a lawyer has been practising.

'We are effectively backing the humans. This is about embracing AI to supercharge our people, rather than codifying processes'

Sarah Harris, Kingsley Napley

Harris was looking for a way to harness the tacit knowledge lawyers develop over the years and make it accessible to everyone in the firm. She met Wendy Jephson, CEO and co-founder of Let’s Think, who is a qualified lawyer and business psychologist. Together, they developed The Knowledge Exchange, which involved training an AI app (built on AWS Bedrock) in behavioural science. It would ask senior lawyers the right questions to elicit tacit knowledge. Then it would analyse their responses in order to capture the complex thought processes that underpin expert knowledge and decision-making, classify the output, and store it in a centralised database, accessed by a conversational (voice-first) user interface. As Jephson says, ‘until GenAI, there wasn’t a way to automate capturing tacit knowledge, organise it and put it in the hands of lawyers’.

The Knowledge Exchange is being piloted in the litigation department ‘because we wanted to apply it to nuanced decision-making’, says Harris. ‘We picked six heavyweight litigators in the criminal team and went through the [knowledge capture] process manually – with Wendy and her team interviewing them. The next step was building the technology to automate that process.’

Jephson explains that the interviews are designed to articulate and encourage reflective and critical thinking, and capture tacit knowledge which then gets analysed in the repository. The difference between accessing a precedent, and accessing knowledge is context – not just what was decided previously, but the thought process that led to a particular approach. ‘We are effectively backing the humans,’ observes Harris. ‘This is about embracing AI to supercharge our people, rather than codifying processes.’

Voice-first for Knowledge Exchange – and Windows 11

Again the modality is voice-first ‘because it’s quick and easy, and you get a different kind of thinking when you think aloud’, Harris says. The Knowledge Exchange provides junior lawyers with a psychologically safe environment to ask straightforward questions about how similar matters were handled previously, for example, before approaching a partner.

On the value of voice, Harris adds that this approach ties in with Windows 11, where the dictate function is front and centre for Outlook and Word, again encouraging voice interaction with Microsoft and Copilot. Knowledge Exchange, and the other use cases for voice-first interaction with AI agents, are another way of moving AI up the value chain. Such developments make AI feel like part of the team rather than part of the tech stack.