Process automation
The repetitive work somebody does by hand today: copying data between two systems, routing requests, producing the same document every week. We start with one process, the one costing the most hours.
Business process automation and AI agents
Almost everyone has tried AI on something, and almost everyone stopped at the same place: the demo worked, Monday morning did not. The gap is not the model, it is the graft. An automation truly enters a company when it knows your data, handles the cases that do not fit, and writes down what it did and why. And if it has to know your data, better that it runs on your own machines.
What we build
Automations and agents that take repetitive work off the people doing it by hand.
AI agents and internal assistants
Assistants that answer on your documents and your data, not on the internet at large. Semantic search over a vector database, so every answer cites where it came from. Where documents are confidential the model can run on a machine of yours: the data never leaves your network, and no supplier contract has to stand in for that guarantee.
Chatbots and reply channels
Automatic answers to the questions you get a hundred times, with a handover to a person when a question leaves the map. Where that line sits is decided together, and it can be moved.
AI inside existing systems
AI inside the system you already use, not in a separate browser tab. Reading documents, extracting fields, consistency checks, classification: the model works where you work.
Who it makes sense for
Who it makes sense for
- Companies where a skilled person spends hours copying data between two programs.
- Anyone answering the same questions from scratch every day.
- Anyone who tried a general-purpose AI tool and stopped because it did not know the company's data.
- Anyone who has to show how a decision was reached, and so cannot accept a closed box.
Frequently asked questions
Frequently asked questions
How do you fit AI into a process that already exists?
You pick one step, measure what it costs today, and automate that while leaving the rest alone. If it holds for a few weeks on real data, you move to the next one. Rewriting the whole process around AI is the fastest way to never ship it.
Does our data end up training a model?
No. We use the OpenAI and Anthropic APIs in modes that do not use submitted data for training. Where data is sensitive, though, a contractual guarantee is not the point, and the answer is different: an open model running on a machine of yours, or processing on the device of whoever uses it. The data never crosses the network, so there is nothing left to guarantee. Which of the two routes to take is decided before we start and written down.
What happens when the AI gets it wrong?
It does get things wrong, so the system is designed knowing that. Where a mistake is cheap, the automation proceeds alone. Where it is not, it proposes and a person confirms. Every step is recorded, so a mistake can be traced rather than discovered months later.
How long before we see a result?
For a single, well-bounded automation, a few weeks to something running on real data. The time goes into understanding the process and handling the cases that do not fit, not into wiring up the model: that part is quick.
Do we have to replace the system we use?
Almost never. If it exposes an API, or even just a readable database, the automation grafts onto it. Replacing the system is a different project, and if it is genuinely needed we will say so plainly instead of dressing it up as automation.
Shall we talk about it?
Tell us about the process that does not work. We reply within 24 hours, with a first 30-minute call and no commitment.
Let's talk about your projectThe other solutions