When AI first began to make its way into our work at Freshfields, I thought the hard part would be the technology itself. Surely the challenge was simple, I thought at first. ‘We find the right tool, plug it in, and it will give us 90–95% of what we need.’
It didn’t work out that way. Because AI underperforms.
On live projects, the tools did get us maybe 60–70% of the way there. That sounds decent, but in legal services the remaining 30–40% isn’t optional. And quality can’t be optional.
Check it?
What about checking the AI’s work ourselves? Well, if you end up verifying every line of an AI output, you haven’t really saved yourself any time at all, have you? Not to mention giving yourself a huge billing headache.
This was all so frustrating. Even more so when we realised the tools had other weaknesses:
- They always tried to give some answer, even if there was nothing relevant. That just wasted time.
- They struggled with scale. A 190-page annual report or a 300-question checklist was simply too much to handle.
- And the tools were pretty good at summarising, but very poor at giving precise yes/no factual answers.
So the question became: how do we make AI useful despite these limits?
The answer, we found, wasn’t more technology. It was more system.
Get with the system
We started applying lessons we’d learned from our work in due diligence. First, work with a sample. Get a few documents right, then scale. Second, refine the prompts relentlessly until the AI output was good enough to serve as a roadmap. Third, accept that validation was essential — but find ways to focus it where it mattered, rather than checking everything blindly.
This helped. But it still wasn’t enough unless people across the team could work the same way. And that’s where leadership came in.
Worldwide leadership
Earlier this year in Berlin we brought project leads from Manchester, Germany, and Hong Kong together — the first time we’d all aligned on a single approach. Until then, each hub had worked in its own way. Getting everyone in the same room (literally and metaphorically) was the breakthrough: it meant that from then on, knowledge could spread across the whole organisation, not just within one office.
The goal wasn’t to showcase the AI tool itself — it was to teach everyone how to use it properly:
- How to design and refine prompts.
- How to test with samples before scaling.
- How to validate in a way that balanced speed with quality.
At first, many of the team were hesitant. Within a few days, though, they were running the conversations themselves — sometimes better than I could. That was the point. The knowledge was spreading and I wasn’t a bottleneck.
The tool is not the point
Looking back, the lesson for me is simple but easy to miss: AI on its own won’t transform legal work. The tool is not the point.
The real craft lies in leadership — in building systems that people can trust, in training colleagues so the methods spread naturally, and in knowing where human judgment must remain at the centre.
The technology will keep evolving. Today it gets us 60–70% of the way there. Tomorrow it might be 80% or even 90%. But the firms that thrive won’t be those with the ‘best’ tool. They’ll be the ones with leaders who can connect people, process, and AI technology at whatever level the AI technology reaches.
**********
WANT TO KNOW MORE?
If you need more insights into AI and its impact on management, please message me here, or reach out directly at Freshfields.