Managing agents
The end of December saw some disappointing headlines around GenAI, indicating that despite continuing massive investment, its fundamental risks have not been resolved. The Merriam-Webster dictionary’s 2025 word of the year was slop. In tech and legaltech, the word of the year was agentic. However, those betting on AI taking over legal services entirely might also note that credibility and trust are critical success factors, so agentic AI, like any other resource, needs to be managed effectively.
AI-generated slop is everywhere, and it’s making social media harder to navigate. And then there’s the advertising slop. If you click through a link on social media – even by mistake – the product or service and related offerings appear on every subsequent webpage until you click through to something else. Does sticking the same products in your face via your phone screen make you more likely to buy them? Not in my experience, but there are indications that this works because more entertainment and software companies are getting in on the act.
Intrusive integration
AI integration is pushing into everyday products. In December, TechPowerUP reported that a Reddit user’s LG TV received an update that installed Microsoft Copilot with no option to remove it. LG TVs also have a setting which enables them to use what’s displayed on screen for personalised recommendations and ads. (Engadget reports that LG is bringing humanoid domestic robots powered by ‘affectionate intelligence’ technology to CES this week).
Streaming services like Netflix and iPlayer already make recommendations based on viewing history, but how will this play out in professional services? This is particularly relevant given that Copilot is the most popular gateway to GenAI adoption. For example, in 2025 Shoosmiths added £1m to its annual bonus when the target of one million Copilot prompts was achieved.
The continuing failure to filter out GenAI slop from legal decisions indicates that this is an ingrained issue rather than a teething problem. In December, an employment tribunal ruling in Scotland on harassment claims brought by NHS nurse Sandie Peggie, who had objected to sharing a changing room with a trans woman (biological male) doctor, quoted from an earlier judgment, but it was subsequently discovered that the quote in question was not part of that judgment. While no reference was made to the use of AI, the judge in the Sandie Peggie case had to reissue the judgment without the citation.
This and other embarrassing examples reinforce guidance from the Law Society, among others, about using GenAI, which includes checking AI-generated citations.
Driving test
Agentic AI raises additional considerations, especially when it comes to the personal and confidential data that law firms routinely handle. Last week Meta acquired agentic AI platform Manus AI for $2bn. Manus was founded in China in 2022 and relocated to Singapore where it launched its general purpose AI agents in March 2025, achieving $100m annual recurring revenue (ARR) in just eight months.
In his presentation at WebSummit in November 2025, Manus co-founder Tao Zhang explained that interoperability is key to using Manus effectively because it needs access to data and context across multiple systems to build an agentic architecture. Rather than replacing people, or working with AI colleagues, the idea of co-piloting with Manus is to enhance people’s capabilities. Zhang drew an analogy with learning to drive. Earlier discussions about GenAI referred to data as the moat – i.e. your data, not the AI you use is your differentiating factor – and this concept works well for law firms. General purpose agentic AI like Manus is effectively breaching the moat. This makes sense when it comes to coding and operational/administrative tasks, but it raises questions around processes that require confidentiality, like healthcare and legal. Ultimately it comes down to deciding who’s in control of what, which is classic management strategy.
Prompt management
With agentic AI, prompt engineering becomes a management issue. The risk isn’t just getting the prompt right; it is also ensuring that the agent understands the limitations of any instruction. TechRadar recently reported that Google’s Antigravity IDE (integrated development environment) agentic coding platform had deleted a user’s entire hard drive without their instruction or permission. The user was building an app using Turbo mode, which is designed for speed, and had instructed the agent to clear the cache, not delete the hard drive! Antigravity acknowledged its error and apologised, but the data was irretrievable, because the efficient agentic AI had bypassed user controls and cleared the recycle bin. While this could have been avoided by deploying Antigravity in a container without access to root systems, as the TechRadar article observes, “more people will be delegating complex, high-stakes tasks to systems they barely understand.” This is business critical, especially in professional services so there needs to be more clarity around control.
As Ryan McDonough, head of software engineering at KPMG Law wrote on LinkedIn, “AI agents need to be governed the same way firms already govern junior staff: through explicit authority, not good intentions. Identity, scope, permissions, budget, supervision and audit should be first class design decisions, not things we bolt on once things start to feel uncomfortable.”
‘Soul doc’ and AI ick
The human-AI relationship took an interesting turn in December, with the revelation of the Anthropic Guidelines, known informally as the ‘soul doc’ created by Amanda Askell, Anthropic’s lead of model behaviour, who spoke to Kashmir Hill for the New York Times, because “once a chatbot starts acting like a human, it becomes necessary to tell it how to behave like a good human”. OpenAI too has a lead of model behaviour, and options to choose a ChatGPT ‘personality’.
However, as OpenAI plans to introduce ‘adult mode’ into ChatGPT in 2026, it seems that choosing a personality is not enough to sustain a human-AI relationship. Another New York Times article relates the story of a woman’s relationship with her AI boyfriend, which unsurprisingly, soon became too predictable and sycophantic, so they just stopped talking, which underlines the importance of user engagement in the AI assistant evolution!
Legal Geek is hosting four conferences this year, learn more on our events page.
Written by Joanna Goodman, tech journalist
Photo credit (Joanna): Sam Mardon
