AI
AI AUTOMATION
- 5+ projects
- from 1 week
- on your own infrastructure
We take the manual routine off your team: processing requests, reading documents, answering the same questions, putting reports together. We break the process into steps and automate the ones where AI behaves predictably.
WHO IS IT FOR?
Three situations where automation pays for itself fastest
- .01
THE ROUTINE IS EATING THE TEAM
Every day people do the same things: move data between systems, retype the same answers, pull reports together, sort through incoming documents. The work is not hard — there is simply a lot of it. The costlier part is elsewhere: those hours are recorded nowhere. The reports show salaries and workload, but not that a third of the working day goes on actions that do not need a person at all.
- .02
YOU TRIED AI AND IT DID NOT STICK
You bought a subscription, gave the team access, showed a couple of examples. A month later two people use it and everyone else is back to the old way — because the tool knows nothing about your data, your rules or your context. Something quietly swaps places: instead of an automated process you get one more window to copy text into by hand. There is more work now, not less.
- .03
THE DATA CANNOT LEAVE THE BUILDING
The process needs automating, but it holds contracts, clients’ personal data, finances or internal documents. Security — or plain common sense — says none of it is going into somebody else’s cloud. The problem is almost never the model. It is where that model runs, who sees the requests, and what stays in the logs once the answer has been given.
WHAT'S INCLUDED
We do not build below this line
- - Process discovery and where AI fits.01
- - A model chosen for the job and the budget.02
- - Connected to your data and services.03
- - Tested on your real data.04
- - An interface for the team.05
- - Guardrails and control over answers.06
- - Logs and usage statistics.07
- - Launch and instructions for the team.08
- THIS IS THE FLOOR, NOT A PACKAGE
Everything listed ships in every project — even the simplest and cheapest one. There are no stripped-down versions.
- AI IS NOT NEEDED EVERYWHERE
Where accuracy to the last cent and repeatability matter, ordinary code is more reliable and cheaper. During discovery we say plainly which steps stay deterministic and which go to the model.
- THE DATA CAN STAY WITH YOU
The model is deployed on our infrastructure or on yours — requests and documents never reach an outside provider. The infrastructure is run by someone with twenty years behind them.
WHAT'S INSIDE
THE DATA STAYS WITH YOU
The model can be deployed on a dedicated server — yours or ours: requests, documents and conversations never reach an outside provider. The infrastructure is run by someone with twenty years behind them, and it is the same infrastructure the rest of our projects live on. If nothing sensitive is involved we work through a provider, which is cheaper and quicker. Which of the two it is gets decided before the start, not discovered after launch.
AI IS NOT NEEDED EVERYWHERE
Where accuracy has to hold without exception — calculations, money, legally binding decisions — ordinary code is more reliable and cheaper. A model works well where something has to be read, understood, compared or drafted. During discovery we say plainly which steps stay deterministic and which go to the model. Selling you AI where a script would do is a bad deal for us too: projects like that do not work, and we are the ones who end up explaining why.
EVERY REQUEST COSTS MONEY
Building it and running it are two separate bills. The second depends on how many users you have and how large their requests are, and it does not go away once the project is handed over. We work out the cost per request before development and set it against the price you charge, across several growth scenarios. And we build in what brings it down: caching repeats, splitting simple and hard requests between different models, limits and alerts when spending spikes.
THE MODEL CHANGES, THE PRODUCT STAYS
Talking to the model lives in a layer of its own: changing provider, or moving to a local model, does not rewrite the product. This is not future-proofing for its own sake but insurance — over the past two years prices, access terms and the quality leaders have all changed. Plus rules and limits on answers, a journal of every request, and statistics: what people ask, where the model gets it wrong, and what is worth caching.
CATEGORIES
Three levels of scope. Yours is the one you recognise yourself in
PROCESS DISCOVERY
WHEN IS THIS YOUR CASE?
It is clear the team loses time to routine, but not where AI will actually help and where it would only add work. We break the processes down step by step, count the manual operations, and mark what can be automated now, what belongs to ordinary code without AI, and what is better left alone. The output is a document: a map of the process, the points where AI fits, and an estimate of time and budget for each. The result is yours whether or not we carry on. If we do, the cost of discovery goes towards the project.
FIRST PROCESS
WHEN IS THIS YOUR CASE?
The entry point is chosen, and it has to work on real data rather than in a demo. We build the automation of one process end to end: the model selected, connections to your systems and documents, an interface the team uses without us, limits on what the model may and may not answer, logs and statistics. We test it on your data before launch and tune it on the results. One process taken to a working state is what shows whether the rest are worth automating.
CLOSED LOOP
WHEN IS THIS YOUR CASE?
What gets automated is not one process but the way the company works: several scenarios, different departments, access rights of their own for every role. Or the data is such that it never leaves the perimeter under any circumstances. The model is deployed on a dedicated server — yours or ours — connected to internal systems and databases, running without a single call to an outside provider. A role model, a journal of every request and answer, monitoring and redundancy, documentation and handover to your team. The logic is designed around your processes rather than bent to fit what somebody else’s service can do.
FAQ
Answers to the most popular questions
From $1.500 for a single automated process to $5.000 and up for a system with a local model and several scenarios. The figure depends on how many steps the process has, how many systems it has to connect to, and whether the model runs on our infrastructure or on yours. That is exactly why the first step is discovery at $1.000–2.000, after which you get a map of the process with an estimate for every entry point rather than a guessed bracket.
Process discovery, 1–1.5 weeks. A first process taken to a working state, 1–4 weeks. A system with a local model and several scenarios, from 2 weeks. Most of the time goes not into development but into testing on your real data: until you can see how the model behaves on actual documents and questions, it is too early to put it to work.
AI works well where something has to be read, understood, compared or drafted: sorting incoming documents, preparing first drafts, searching an internal knowledge base, reconciling data from several sources. It works badly where accuracy has to hold without exception: calculations, accounting, legally binding decisions. We leave those steps to ordinary code, and we say so during discovery rather than after the invoice.
It depends which option you choose. Working through an outside provider means the requests go to them — cheaper and faster to launch, and fine when nothing sensitive is involved. With a local deployment the model runs on a dedicated server, yours or ours, and the data never leaves the perimeter. This is settled before we start rather than halfway through: both the architecture and the budget follow from it.
For most jobs an existing model tuned to your context is enough: your data, your rules, your standards for what an answer should look like. Training one from scratch is expensive and almost never pays for itself. Open models deployed locally cover the cases where confidentiality matters, or where the cost has to stay predictable at a high volume of requests. Which one suits you becomes clear after discovery — it is a question of the job and the budget, not of fashion.
It will get things wrong — the question is where, and what follows. So we separate the steps: where a mistake is expensive, the model prepares an option and a person makes the call. Where it is cheap, the model acts on its own. Add to that limits on what the model is allowed to answer and a journal of every request: if something goes wrong, you can see exactly what was asked and what came back. We do not build systems that take irreversible actions unchecked.
Usually not — what changes is what they spend the day on. Steps get automated, not job titles: sorting the inbox, drafting, moving data, searching documents. The person stops being a relay and starts working with the result. If the goal is to cut headcount, say so plainly — the calculation is different, and we will tell you honestly whether it is realistic.
Access to the people who run the process and to what they work with: sample documents, threads, spreadsheets, procedures. No specification is required — it is what discovery produces. The closer the samples are to reality, the more accurate the estimate: on polished examples a model shows a result that will not repeat in real work.
You do. The source code, the configuration, the documentation and everything built along the way are handed over with the project — this is not a subscription to a service of ours and not a tie to us. If you decide to change contractor or build your own team, the system leaves with you. If the model is external, you pay the provider directly — we do not resell access.
The system can stay on our infrastructure: monitoring, model updates, and reworking the scenarios as the process itself changes. With AI projects that is the rule rather than the exception — quality only shows over a long run, and the first months usually go into tuning against the real logs.