Applied AI: what we learned using it in real products and client work
No hype: how we use AI in our own SaaS, what we built for clients (including real-time avatar agents) and the six lessons we wish we had known earlier.
We have been using language models since they stopped being a toy: in our development workflow, inside our products and in client projects. This post is not futurology. It is about what works, what does not, and what we wish we had known before starting.
Where we use it ourselves
In our own SaaS. Our products carry AI where it shortens real work: document review and data extraction (a photographed expense becomes an accounting entry with its fields filled in), plain-language explanations of technical results, and recommendations that prioritize what to fix first. None of these features are marketed as "magic". They are marketed as what they are: minutes saved every day.
In client projects. The most striking case: avatar agents that talk in real time. An assistant with a face and a voice that serves a client's users, trained with a custom system prompt (personality, limits, brand tone) and connected via RAG to the company's knowledge, so it answers with the business's data rather than generic internet knowledge. We wrote a whole article about it, because the architecture deserves one.
In how we work. Code assistants have been part of our daily workflow from the start. They do not replace judgment: they amplify it. Review stays human.
The six lessons
1. Start from a measurable process, not from "we want AI"
Projects that go well start with a sentence like "reviewing each invoice takes us 4 minutes and 300 come in every month". Projects that go badly start with "we want a chatbot". If you cannot measure the before, you will never prove the after.
2. RAG beats retraining almost every time
For a model to answer with your company's knowledge you do not need to train it on your data. You need retrieval: index your documents and hand the model the relevant fragments with each question. It is cheaper, updates instantly (upload a new document and it is live) and you can audit which sources each answer used.
3. The system prompt is a product, not a sentence
The difference between an assistant that represents your brand and one that invents refund policies lies in a well-worked system prompt: what it knows, what it must never do, how it speaks, when it hands over to a human. Ours goes through versions, tests and regressions like any other piece of software.
4. Keep a human in the loop
For anything touching money, customers or legal decisions, AI proposes and a person disposes. A draft reply that a human agent approves in two seconds captures almost all the value at a fraction of the risk.
5. Budget cost and latency from day one
A large model on every request can ruin a feature's margins. There is almost always a better tiered architecture: small models to classify and route, the large model only when needed. And in real time (voice, avatars), latency rules over everything else.
6. Evaluate before launch, not after the incident
A set of test questions with expected answers, run on every prompt or model change, catches regressions before your users do. It is boring, and it is what separates a reliable assistant from an anecdote on social media.
Where to start in your company
Pick a process that hurts (repetitive, measurable, frequent), build the smallest version that could work, keep a human in the loop, and measure for two weeks. If the numbers hold, expand. If not, you have lost little.
If you would rather walk that path with company, that is exactly what we do in consulting: identify where AI brings real value to your business and build the integration, from automating one process to an avatar agent talking to your customers.