Over the past two years we've shipped roughly fourteen AI features into production apps. The good ones have one thing in common: nobody talks about them. They just save time.
When a tax advisor photographs receipts in TaxCastle, he isn't thinking about the vision model that turns the receipt into DATEV-compliant entries in 1.8 seconds. He's thinking about leaving on time tonight. That's the bar.
Internally we call it "tool AI". It sits beneath the surface. It doesn't announce itself. It has no avatar with a name. And it works exactly when it's needed — not one action before.
Three principles we measure features against
We have a short checklist every new AI feature has to pass. It lives in a Confluence doc, but honestly it's also taped to the wall by the coffee machine.
- Does it save clicks or typing? If the answer is "no", the feature is cut. Magic without time savings is a toy.
- Does it work without the model? Every workflow needs a fallback — manual, rule-based, something. AI is an accelerator, never a requirement.
- Would the user miss it if it were gone? We ask this after three weeks. If nobody complains, it wasn't a missing feature.
The classic: photographing receipts
TaxCastle is the app where we've worked on this principle longest. Tax firms get hundreds of receipts every month — entertainment, fuel, hotel bills. Classically, all retyped.
Our model reads date, amount, VAT rate, vendor and purpose in under three seconds straight from the photo. But the interesting part isn't the model, that's standard. The interesting part is what doesn't happen:
- No chat dialog opens.
- No "AI is thinking…" indicator appears.
- The user isn't asked to confirm the result.
- No avatar pops up saying "Hi! I'm Tax-Bot".
Instead: take a photo → receipt flies into the right client folder. Sorted wrong? Drag-and-drop into another. That's all. And exactly because that's all, tax firms love the feature.
What happens when you do it the other way around
On another project — client unnamed — we built a chatbot into a field-service app. With avatar, personality, the full show. Adoption rate after three months: 4%. The same function without avatar, embedded into the existing search bar: 71%.
The lesson isn't that chatbots are bad. The lesson is that form follows function, always. If the user already has a search bar they're used to, AI belongs behind it. If they're focused on the receipt, AI belongs on the photo. Not in a separate dialog.
Where we're going in 2026
The next wave we're working on is predictive tool AI. Meaning: before the user asks, the answer is already there. Example from RoadReview — from the mistakes the instructor tagged during the lesson, the suggested focus points for the next lesson should be ready before they step out of the car.
It's not magic. It's just the question: what does the user do next, and how can I prepare it? Answers before they become questions. That's the blueprint. How we build this kind of tool AI into existing apps and workflows is on our AI development page.
