An AI deserves a good workshop too
If you want to make the most of a craftsman's talent and skills, you need to create a good workshop with high-quality tools.
An AI assistant is no different. At Jafix we are building Jafix Assist, an AI assistant for repair businesses that lets freelancers and small businesses automate their business processes.
Now I can hear you thinking: a SaaS solution? Why wouldn't a business just do this itself? Have employees install Claude Code and automate their own work. There are indeed innovative companies doing exactly that, but it's far from feasible for every business, because it takes a lot of technical expertise.
Think of setting up a knowledge base, integrations with other systems, security, and what about personal data and GDPR? And of course you don't want your assistant to only work while your laptop is open.
From chatbot to domain-specific AI assistant
At Jafix it started with Handige Harry ("Handy Harry"): a chatbot that used AI to translate a question like "my vacuum cleaner makes a strange noise" into the right search filters, so you ended up at a repair guide or a repairer. Intent detection and query understanding, mostly. We went through several iterations after that, but Jafix Assist only came about with the rise of AI agents.
No more question-and-answer game, but AIs that can actually do things, carry out actions and keep going until the task is done.
That is where Jafix Assist fits in: it does work and can be used to automate business processes. Think of CRM tasks such as processing intakes and customers, drawing up quotes, and building and maintaining the website. Specialised, then, in solving problems in our domain, which in turn is possible thanks to our years of knowledge of the sector.
Agents do need a so-called harness, though. Think of the harness you use to guide horses, or the climbing harness you put on for safety: something that reins in forces, keeps things safe and gives direction. If you use Claude Code or ChatGPT, you are working inside such a harness.
Write a quote for yesterday’s e-bike
Link goal e-bike request Read procedure write a quote Fetch request 2 photos Look up prices 3 lines Draft quote draft The quote is ready: € 82.50 incl. VAT. Shall I send it?
To complete the metaphor: the workshop is the harness, where the craftsman (the LLM) has the right tools, can put on safety glasses and works within the rules of the workshop. Set up the workshop well and he can be put to the best possible use.
The work also shifts here, from a chatbot that is (annoying) for the customer to an assistant that does work inside the business and doesn't communicate with customers directly.
Most work doesn't need frontier models
Jafix Assist doesn't run on the biggest and most expensive frontier models, but on smaller open (open source/open weights) models, hosted entirely in the EU. Every model is replaceable, which lets us experiment with new and better models and gives us sovereignty: we don't depend on a single vendor.
Challenges from practice: does a "this one" in the chat refer to the quote or to the website we're working on? Is an email a real request or dangerous spam? Which tools should be used?
For questions like these we use a separate, super-fast and small decision model. It doesn't write text, it classifies options. When in doubt, we ask which context is meant, really just as a person would.
You see this idea more and more as a way to make your agents work more efficiently and safely next to the thinking model, and it fits the rest of the workshop: the craftsman shouldn't have to sort the mail as well.
How do you know it's going well?
Evals! Evals! Evals! According to Anthropic, evaluating is the hardest part of agents: you have to look at the outcome, not at what the model says it did.
We do that in several ways, but my favourite is the mystery shopper: a stronger model plays an owner using the app for the first time, following various personas and scenarios, and afterwards judges what was broken, wrong or annoying.
Users can also share conversations with us for review. With these results we keep improving the harness and making it more robust (self-recovery). But within a single organisation, too, the assistant learns by itself without any code changes, because it learns from the corrections and instructions the user gives. And when something goes wrong, that is passed on immediately as well.
Put the new opening hours on the site
Update page opening hours Build website time-out This didn’t work. I’ve passed it on to Jafix.
Report with trace sent to Jafix
Chatified applications and UX/UI
An AI chat interface inside a "chatified" app comes with quite a few challenges. That's why you see OpenAI and Anthropic shipping a steady stream of updates to improve the user experience. Even there, I sometimes lose my way, and above all my conversations.
In Jafix Assist everything happens in one chat. No separate conversations. That makes it simpler, but it also brings challenges. Conversations and (background) processes can easily run side by side. On top of that, the application's own screens also shape the context. What is the user looking at? What do they mean by "this one" in the chat? Which workflows and procedures are currently running?
On mobile it gets even harder, because you can't (and don't want to) do everything in the chat. So sometimes the assistant prepares something in the workspace for you to review, but on a small screen the workspace can't be shown at the same time.
And the web interface is just one of the channels that can be used. Users can also communicate via Slack, WhatsApp and email, for example.
Write a quote for the e-bike
Draft quote draft It’s ready in your workspace.
Sent! I’ll let you know when the customer replies.
| Replace spokes (2×) | € 45,00 |
| True the wheel | € 25,00 |
| Call-out fee | € 12,50 |
| Total incl. VAT | € 82,50 |
Human in the loop
The assistant could do a great deal autonomously, but trust has to be built. So for every communication to the outside world there is a moment where you approve the message, the new page on the website or the quote.
Compare it to hiring a new employee: on their first day you probably wouldn't let them make important decisions without approval either.
Next steps
It's early days, things are moving fast and we're experimenting with where the opportunities lie for our customers. Try it yourself and let us know how it goes!
About the author
- Albert Skibinski is a product engineer and co-founder of Jafix.
- I write about web development, long bike rides and food!
