The good and bad of clients using AI tools and more notes from 19th August 2026
Posted by Paul Silver
On the 19th August, fifteen freelancers met in the beer garden of the Battle of Trafalgar in Brighton to talk clients, tech, cats, and other random things. This is a little of what we talked about:
- Conversational interfaces vs traditional interfaces
- New work!
- Clients exploring ideas via Claude AI
- Clients exposing confidential information via n8n without realising what they’re doing
- Getting non-techie people at a client to understand risk
- CTO work
- Penetration testing and trying to get people to read the reports
- Working with people who have only developed with LLMs
- Trying to bring version control and various standard disciplines into an existing project
- Difficulties trying to get Claude to commit to git on changes
- Benchmarking local models vs Claude
- Dealing with lack of sleep
- Cat babysitting
- AYGO console
- Live ginger cat capture
- Stag/hen do and wedding stories
- Car insurance
- The BBQ emergency alert
- Red Bull variants
Clients using AI tools
There were various stories of clients using AI/LLM tools to try things out, which also expose potential problems across the current tech industry.
On the one hand we had some folk using Claude to build a rough versions of tools they know they want in their business and wanted to make alpha versions to see if they were actually useful, and then talked to a full on programmer about turning them into ‘real’ tools – i.e. something reliable and properly built that they will be able to use for years. This feels like a very good use of an advanced tool, where people inside a business can make early versions without the cost of a developer as they will likely ditch some ideas once they’re explored.
On the other hand, a client was exposing their private data by using a tool on a cloud based platform, which is a bad risk. This shows that a powerful tool can cause problems if used by people who don’t have a grounding on what is required when making a publicly hosted service. They didn’t understand how they were making private information easily browsable by the public as they’d never had to work through such issues before. Fortunately, it was caught before it became a problem.
Part of the problem here is the old unknown-unknowns issue – if someone is using a tool they feel is quite magical to show some of their private information to the public, they are not going to know the simple ways a nefarious person could see if more information is available than the company planned. They are using a tool that seems to know what it is doing, but do not know how they need to tell it to work. They could presume the tool will make their service secure by default, but usually, wouldn’t even think about the security side of things.
It is the view of many developers in the Farm that there will be freelance work available now and in the future cleaning up problems caused by people using AI tools to make services when they don’t have the grounding needed in making public tools. If you are not an AI/LLM fan, it is important to remember that people are making mistakes via naivety, not stupidity or malice. If you’re using a tool that can build a whole website to talk to your data just by asking it, that will feel magical when you have not known how to even approach that in the past. Finding out later that the process was not quite as magical as you thought is going to hurt for some people, and they don’t need that feeling rubbed in.
Ginger cat entertainment
We were joined in the beer garden by a very nervous, but also very curious ginger cat, who thoroughly investigated the various bushes, trees and tables until their owner came to find them. Apparently, they shouldn’t have been out at all. Good luck escaping again, puss!
Thanks to Chris for the photo