The idea of having people without formal software development skills being able to use vibe coding to create simple applications is both compelling and a potential nightmare for busy IT departments.
The idea of citizen developers is nothing new; Microsoft’s Excel and Word macros have helped people automate repetitive tasks using a relatively simple programming language.
Now, with the advent of large language models (LLMs) that do a decent job of coding, the barrier to entry to software development has effectively been lowered. Anyone who is prepared to describe what they want to achieve as a series of prompts submitted to a coding chatbot can get program code generated automatically.
But how useful is that code, and would the easier entry into programming be compelling enough to get more people in the organisation to build their own programs? This is one of the questions Computer Weekly asked Starling Bank’s deputy CIO, Frédéric Laurent, during a podcast discussion about the bank’s use of artificial intelligence (AI).
Laurent says vibe coding increases the number of people who can build applications, but this does not necessarily mean anyone can write software. “I think you need to have the appetite to build programs,” he says. “You need to have an attention to detail if you want to build them correctly, such as verifying that the outputs are what you expect.”
These outputs may not necessarily be just automatically generated code. “It could literally just be an agent that deploys a website for you,” says Laurent. “You’ll still need to go and check that it’s correct and, in my opinion, I don’t think everybody out there has the patience and attention to detail needed.”
Without an attention to detail, he says vibe coding is relegated to something fun for people to play with, rather than a scalable approach for citizen software development in business.
Laurent does quite a bit of coding himself outside work, using Claude Code. “In a coding context, Claude Code embarks on the architectural solution it is going to build based on the functional problem I give it,” he says. “I then ask one follow-up question because – based on my engineering experience – something feels off. So, I’m going to ask one question and it responds saying, ‘That was really stupid of me’ and then changes the plan entirely.”
As an engineer, he says this behaviour does make him think twice about how much confidence to put in such tools. Nevertheless Laurent believes the future is very much AI-assisted coding or having applications entirely generated by AI.
“I do think there are still skeptics within the engineering community and I can totally see why that is the case,” he says. “I think until those [the erroneous behaviour of the AI models] are ironed out, I can see why some engineers will be skeptics.”
Understanding new concepts
In his day job, Laurent says he uses AI to help understand a new subject matter quickly. “I think AI is an incredible tool with regards to education,” he says. “When I come across new concepts, my first port of call is always to go and ask an AI model to explain it to me as if I was a five-year-old. I also ask it to give me a list of good references, because we all know that AI models can hallucinate.”
While he is not yet using OpenClaw as an AI personal assistant, Laurent does use NotebookLM from Google to help him go through any articles or archived research papers he finds. “It builds an ‘outsourced memory’ which I can then query directly, and certainly when getting a grasp on a topic quickly,” says Laurent.
When asked if AI has changed how he works, he says: “I think anyone who is saying that AI hasn’t changed the way they work maybe has missed the boat.”
While Laurent prefers doing the research himself, he says he will sometimes submit a problem to the AI just to see whether it comes up with an alternative angle or another perspective. “Then you ask it almost to contradict itself immediately after it gives the answer,” says Laurent.
Devil’s advocate
He says the LLM chatbots are trained to answer a question and provide some output, but due to the risk of AI hallucination, there will always be a question mark over how much credence is put on a particular response. “I quite like playing the game of saying, ‘you’ve told me something, now play the opposite role’ or ‘if you needed to disprove this [response], what would you do’, and it’s quite interesting to see how the AI behaves,” says Laurent.
In his experience, the quality of today’s models means they seem to think their original response is fine. However, he says: “There are a few cases where the AI actually comes back and says, ‘actually, I think I haven’t considered this properly’, and then presents additional facts I should consider.”
This is Laurent’s technique for testing the robustness of the AI’s responses to his prompts. “There was a period of time where I would systematically ask the AI to ask me 10 clarifying questions before it tried to answer my query. Funnily enough, a lot of AI providers realised that this piece of advice had been doing the rounds, and so they’ve incorporated it into the AI services themselves, so I seldom use it these days.”
As AI technology improves and the AI industry identifies how best to use it, Laurent says Starling Bank has tried to make sure those technological advancements benefit both the bank and also its customers.
“We started with spending, which was natural language processing based on LLMs,” he says. “Then the AI companies added multi-modal capabilities, so we created scam intelligence, which can take a picture and decide whether there are some red flags relating to fraud.”
The bank considered reasoning capabilities when it created the Starling Assistant. “Now the industry seems to have coalesced around that concept of skills, which is a way of telling an AI model how to perform a task through a succession of steps that you’ve essentially encoded,” he says.
Starling Bank recently added “smart tools”, a suite of AI tools that act as skills for the Starling Assistant. The tools are being built based on customer feedback, and are designed to help the bank’s customers better manage their money.
S 004