AI helps us code faster, but it doesn’t get the final say at Ideal Postcodes.
Artificial intelligence is changing how teams work across nearly every industry. At Ideal Postcodes, we’ve been exploring how AI can support our day-to-day work, and it’s already making a difference.
We’ve found that it plays an important role in making our work faster and a little more enjoyable. It’s also helped us deliver better outcomes for our customers.
We focus on using AI where it adds real value, such as improving the pace of development, supporting learning, and helping us explore new approaches to problem-solving. We also use it as a tool that supports our work and helps us use our time more effectively. It can speed up our repetitive tasks, assist with research, and generate ideas, but it does not replace our judgement or experience.
As a remote business, we also place a high value on work–life balance. Encouraging the team to use AI tools to work more efficiently has created more time for deep work and flexibility during the day. It’s one of the ways we support a healthier, more productive working environment.
At the same time, we recognise the limitations of AI. We use it selectively and responsibly, always ensuring we understand and verify the outputs before putting them to use.
For our developers, AI has become part of the daily toolkit. It helps with tasks that are time-consuming but relatively straightforward, such as writing boilerplate code, generating tests, or fixing small errors. That leaves more space to focus on complex challenges and higher-level design.
Chris our CTO uses AI to reduce the mental load that comes with repetitive development tasks. In many cases, the structure of the solution is already clear, it just needs to be written out. AI can take on this type of work quickly, so the team is still fresh and energetic by lunchtime.
Our software engineers often run multiple AI sessions in parallel, each handling a different task. While one model writes a feature or fixes a bug, another might generate tests or update documentation. Running several models in tandem allows us to compare outputs, add context or data, and explore different solution paths. This has significantly increased our throughput and made collaboration more dynamic.
Large language models have also become useful for learning and research. The team use it to understand new topics, review research papers, or test their knowledge through back-and-forth dialogue. It’s not just limited to desk time, we use it while travelling, taking a walk, or even doing personal chores. These tools have sparked interesting discussions within the team about how each of us applies AI in our own life.
While AI is a useful tool across many areas of our work, we set boundaries around where and how it’s used. We do not allow it to make major product or business decisions, but it does support nearly every other part of our process in some way.
Of course, AI isn’t always reliable or adaptable. One of the reasons we continue to use it is to better understand where its strengths and limitations lie. Over time, this helps us build an intuition about how it works, what to ask, how to interpret its responses, and when to double-check the results.
We rely on AI for well-understood, low-risk problems. Outside those cases, its role is more exploratory. Still, it’s hard to imagine an area where it won’t have some form of utility.
We focus on outcomes and on whether a technology is genuinely useful for our customers. In address validation, accuracy depends on access to reliable data. You can use AI to help validate an address, but you can’t simply provide it with one in a closed environment and expect it to know the answer. Without access to ground truth, an AI model cannot confirm what is correct, and this distinction is essential when accuracy matters.
Our use of AI goes back many years, well before transformer models became popular. Earlier on, we trained models to classify and separate address strings into fields. While these worked in certain situations, they also came with clear limitations.
We are now exploring how transformer models could help us reach what our CTO describes as “human-level matching, but at the speed and throughput of a machine.” Because these models can interpret the meaning and context of text, much like a person does, they open new possibilities for more intelligent and reliable address validation.
Our early results are promising, and we’re excited about the potential. We’ll be sharing more soon!
We believe that consumer and business expectations are changing rapidly. It is not yet clear exactly how they will evolve, but it is certain that they will. Our priority is to ensure that as these expectations shift, our products evolve with them.
We are already seeing signs of this in how people interact with technology. For example, we now create documentation designed not only for developers but also for language models. This makes it easier for our customers and AI systems to understand and integrate our tools, and the feedback so far has been very positive.
We have been loving using AI as it is helping us adapt to these shifts in practical ways. It allows us to experiment, improve our processes, and respond more quickly to what customers need.