Skip to content
At this point, most organisations have probably had the conversation.
Someone in your team has tried ChatGPT. Someone else has raised concerns about governance and data. Someone has read a headline saying AI will transform productivity. Someone else thinks it’s mostly hype.
And somewhere in the middle sits a slightly more practical question:
What does actually introducing one of these tools into a business look like?
Not in theory, just day-to-day.
Where do people use it? What’s useful? What turns out not to matter? How do you introduce something like this without creating more friction, risk or admin than you remove?
About a year ago, we found ourselves asking those same questions.
In Adam’s previous blog on AI adoption at 8fold, he shared how we approached introducing AI safely and what governance principles helped us create space for experimentation.
This article is more about the practical side.
What happened once we actually put a general-purpose LLM into people’s hands and started using it in the real world?
Could a large language model actually improve the way we work, and if so, what would that look like in practice?
We weren’t looking for an AI compliance specialist
We didn’t want an LLM to do compliance work. We already have people for that, and they’re very good at it.
What we wanted was something much less ambitious. Think less artificial compliance consultant and more new-and-improved Clippy.
(If Clippy passed you by, it was Microsoft’s famously enthusiastic writing assistant that used to appear in Office documents and offer help, whether you wanted it or not.)
That was much closer to our intended use.
We just wanted to reduce friction in some of our internal processes and help people communicate more effectively. That meant helping staff tailor reports to different audiences, improve written communications, reduce admin overhead and create more space for the work humans are actually good at.
The challenge wasn’t the technology. It was making it usable.
Like lots of organisations, we experimented with tools like ChatGPT and Claude. They worked well enough, but we kept hitting two practical issues.
The first: information management.
Even when we removed confidential information and disabled training settings where possible, using standalone tools created another information asset to think about and control. Every upload created another process. Every check added time. Eventually we realised we were spending enough effort managing the process that we were losing some of the benefit.
The second issue surprised us more:
People simply weren’t reaching for the tools.
Not because they didn’t like them. They just weren’t where work happened. Switching tabs, copying and pasting text, moving between systems. None of that sounds difficult, but small amounts of friction add up quickly. How many extra steps are worth taking to make an email slightly clearer?
That became the real question for us
Crossing the Rubicon
Then Google made the decision for us. Gemini appeared inside our Google Workspace environment (if I’m honest, our first reaction wasn’t excitement. We were slightly irritated).
Like any change in tooling, we wanted to understand what had changed, what controls existed and what that meant for our governance processes. We added monitoring around developer updates and treated it as something to review properly.
But it also made us curious.
If AI is already appearing inside the systems people use every day, what would responsible adoption actually look like?
So we set up an entirely separate Google Workspace environment and tested Gemini independently. What surprised us wasn’t that Gemini was dramatically more capable than the alternatives. It was that for our intended use, it worked well because it was already embedded into existing workflows.
There was less switching between systems, fewer information management concerns and much less effort required to make it part of day-to-day work.
What using Gemini actually looks like at 8fold
A year in, our use of Gemini has stayed fairly practical and low risk.
We use it to;
- help rewrite communications and adjust tone
- shorten emails and improve clarity
- help locate information and documents more efficiently
- transcribe and summarise meetings
We also use it for early-stage research, although with an important caveat. We ask for references and links and verify information ourselves rather than treating outputs as authoritative.
One of the more useful roles it’s found is acting as a final review layer. Sometimes it’s less about generating something new and more about being a third pair of eyes:
- Does this align with guidance?
- Is something unclear?
- Have we missed anything obvious?
That’s often where we’ve seen the most value. At the same time, we’ve been careful not to overstate what these tools can do. We still review transcripts, we still check sources, we still make decisions.
What hasn’t worked quite so well?
Some of the limitations have been useful to discover, too.
We’ve tested more specialist use cases and found that current LLMs still struggle when tasks depend heavily on interpretation, judgement and context.
For example, document gap analysis sounds like the sort of thing AI should be good at. In practice, we’ve found the level of detail and interpretation required means it doesn’t perform at the standard we’d expect.
Interestingly, we’ve also become reasonably good at spotting communications that have been heavily LLM-generated. Not because they’re wrong, but because they feel slightly disconnected from the audience or situation. The words are there, but the judgment often isn’t.
Thinking about implementing an LLM yourself ?
One of the outcomes of doing this work internally is that it’s refined the questions we’d ask if we were approaching implementation again.
Rather than starting with the technology, we think it’s best to begin by asking:
- What problem are you actually trying to solve?
- Does the tool fit naturally into existing workflows?
- Will people realistically use it?
- Will it save time, improve quality, or both?
- What unintended consequences could it create?
- What still needs human oversight?
Those questions ended up being more useful to us than comparing model benchmarks.
What hasn’t changed in a post-Gemini world
The most valuable thing Gemini has done for us hasn’t been replacing expertise.
It’s reduced friction around communication, drafting, finding information and admin. That’s created more capacity for the work that still depends on people:
- Interpreting requirements
- Making difficult judgment calls
- Understanding context
- Helping organisations implement change in a way that works in the real world
That’s ultimately why we’re not especially worried about AI replacing what we do.
Clients don’t work with us just because we can generate information quickly; they work with us because they trust us to apply judgment and help them make good decisions.
Ready to introduce an LLM in to your organisation?
If you’re exploring how to introduce LLMs into your organisation safely and proportionately, we’d be happy to share what we’ve learned.
