Cedric Chin over at Commoncog recently wrote a great article on how to make sense of what’s going on with AI, when so much of what’s said on the subject is hype and bluster. His advice was to pay attention only to “detailed field reports” of people using AI for real-world purposes in real-world businesses. From those reports, we should ask:
- What new outcomes are suggested by this usage?
- What are some actions I may take in response to it?
- What are the relative values of these possible outcomes to me?
- What are the causal relationships here?
As someone who has been working heavily with AI, and doing so in a way that’s potentially different to others given my background in software development, I’m going to write some field reports of things that I’ve been working on, with the aim of providing people with fodder to answer those four questions and ways to make sense of what’s going on.
In the first of these reports, I explore how I developed an AI-powered store finder.
The situation
Honest Umami manufactures a range of MSG-based seasonings for the consumer market, which are sold primarily in independent food and drink stores in the UK. One of the perennial problems we need to solve is finding new potential stockists and getting the brand listed with them.
As with any brand, targeting is an important part of that job. The more we target the right stores, the more cost-effective it is for us to gain listings and the faster we grow; the more we target the wrong stores, the more time and money we waste visiting and sampling inappropriate stores, and at some point we potentially even start to lose money on customer acquisition. That’s a particular problem for Honest Umami, because we sell MSG, a product that can be divisive, and so not every retailer is actually a potential stockist. It’s easier for us to get targeting wrong than it is for other brands.
The previous status quo
There are existing tools for prospecting for new stores, such as Bowimi. They typically allow you to search Google Maps for stores and then add those stores to lists. If they have custom functionality it’s generally around route planning, assigning leads to salespeople, monitoring which leads salespeople are visiting, etc.; more field sales management than store evaluation.
What they don’t offer is a way to predict, before visiting a store or doing detailed research into one, whether it’s going to be a good fit for your brand and your products.
For Honest Umami, that’s a dealbreaker, because targeting is the whole issue. Having a list of every store in the UK is far less helpful than having a list of the few hundred that are perfect for us.
What I did
I developed a tool called Findie. It has lots of information on thousands of stores, and if you tell it where you’re currently stocked then it can recommend to you new stores that are the best possible fit for your brand. It started as an internal tool at Honest Umami and has become a standalone product.
How AI was involved
It straightforwardly wouldn’t have been possible to develop Findie without AI. There are two aspects to this: the first is that some critical functionality within the Findie app is only possible with AI technologies; and the second is that I couldn’t have developed Findie as quickly as I could without the help of Claude Code.
In terms of Findie’s reliance on AI, the following features were not possible – or at least not practical for a solo developer to implement straightforwardly – until AI came along:
-
The comparison of stores to one another in a way that takes into account semantic similarity, not just literal word-for-word similarity. So Findie knows that stores that talk about “bbq” are similar to ones that talk about “barbecue”, but also that they’re quite similar to ones that talk about “hot sauce” or “brisket” or “smoking”. (It does this using an embeddings model.)
-
The evaluation of photos and website screenshots as well as of text. So Findie is able to look at photos of a store and assess its vibe, its interior design, etc.; things a human would do instinctively, but that we couldn’t automate easily before AI vision models came along.
-
The generating of high-quality text summaries of stores. So Findie can generate several paragraphs of accurate, highly readable text about a store, and can summarise the content on a store’s website in a way that actual conveys meaningful information.
-
The natural-language conversational search. So you can search for things like “delis in Manchester similar to my current stockists” rather than faffing with complicated search filters.
The other aspect, the speed of development, is also important. I have a background in web development, have spent hundreds of hours on Findie, and could definitely have developed something like it without the help of AI. But I know that I wouldn’t. In particular, I wouldn’t have spent that time on it if I hadn’t been able to validate extremely quickly that it would work. With Claude Code I was able to make a proof of concept in a morning that convinced me it was worth spending more time on.
This removal of any friction to experimenting and creating a rough prototype is what made me go through with developing it. The cost of getting started is basically zero. But the speed of development is also important because it allowed me to get Findie to a state of polish that made it feasible to release it as a standalone app. Findie in the pre-AI world might have languished as a janky internal tool that was just about usable by me but that couldn’t be used by anyone else. Claude Code’s speed meant I could polish it up, add more features, support multiple brands, and do all of the other things needed to make it a general-purpose app.
General takeaways
From my experiments with Findie, I’ve drawn several general conclusions that have changed the way I look at AI, internal tools, software products, and more. They are:
-
Internal tools can be custom-made and can have a level of polish that’s as good as a proper app. We should look at our businesses and identify places where things can be automated, places where there is friction that software can remove. It used to be uneconomical to create one-off tools to fix problems like this; that has changed.
-
Creating many of these tools still (for now) requires actual software development skills – I don’t think I could’ve created Findie without my web dev background. But many don’t, and even where they do need that professional help, they’re still cheaper and easier to create than they were before. My sense is that where these tools require the persistence of data in databases, integration with external services, or complex authentication and permissions, they still need the help of a developer. But where they’re standalone, straightforward tools (think something that converts data from one format to another, or a margin calculator, or something for generating a PDF, etc. etc.) they can easily be created by non-technical users.
-
In the olden days, developing software required a lot of up-front investment, but then had zero marginal costs – so it cost a lot to serve your first customer, but then your second (or your two millionth) represented almost pure profit. This led to “software as a service” businesses trying to serve as big an audience as possible, to maximise the return on their initial investment in the software. But now, with AI, software businesses can be niche and target individual segments while still being profitable, and can therefore potentially serve their markets better than if they were forced to have a lowest-common-denominator offer. Bowimi suffers from needing to work for car parts manufacturers and soft drinks brands and marketing agencies and anyone that does field sales. But by ignoring most companies and focusing on food and drink brands, Findie can potentially better serve the needs of those customers. We should expect more specialist, industry-specific or otherwise-niche tools like Findie to emerge.
From there, I remind you of Cedric’s four questions:
- What new outcomes are suggested by this usage?
- What are some actions I may take in response to it?
- What are the relative values of these possible outcomes to me?
- What are the causal relationships here?