Artificial intelligence in a company only makes sense if you can measure what it changed. That's why we don't sell "AI transformation" — we automate one process, measure it before and after, and only discuss full deployment once the pilot clears a success threshold that you set yourself.
AI Automation
Measured before it's deployed
AI process automation that starts by measuring a single operational process, tests it in a fixed-price pilot against a threshold you set, and moves to deployment only once that threshold is met.
Orders arrive by email — each in a different format. Complaints have to be read, classified, and re-entered into the system. Supplier invoices are PDFs that someone extracts data from by hand. Your team spends hours a day on this, and volume keeps growing faster than headcount.
You've tried rules and macros. They worked — until the first unusual case, and then exceptions started outnumbering rules, and someone had to babysit the automation that was supposed to run itself. Managers ask for more headcount, but operational roles are getting harder to fill, because nobody wants a job that consists of retyping data.
At the same time, you keep hearing from every direction that "AI will handle it" — from people who have never seen your process. And your doubts are reasonable: models make things up, company data isn't something you send out into the world, and hype has a way of fading.
We automate processes that couldn't be captured in a plain algorithm — where rules kept losing to exceptions. Not the whole company at once: one process, measured before and after.
The order is non-negotiable: first measurement (how much this process costs today, in hours and errors), then a pilot on a sample of your data, and only once an agreed threshold is cleared — full deployment. You can stop at any of these gates. Details below.
We start with measuring the process, not the technology. If the numbers don't justify deployment, we say so plainly and stop at the advisory stage — a deployment that doesn't deliver would come back to us as a complaint, so this honesty is in our own interest, not just good manners.
We do the programming and the ERP/CRM integrations ourselves, not through a subcontractor — automation that can't talk to your systems is a gadget, not a tool. And experience with operational processes means we talk about your document workflow first, and about models only after that.
And here's something unpopular in this industry: we don't promise that AI will "do everything." We automate work that is repetitive and verifiable; decisions with financial or legal consequences stay with a person — by design, not by oversight.
What you get
- Shorter handling time, counted in hours and in moneyNot an "improvement" — a number. Case handling time before and after automation, on your own data, measured in a way you approve yourself.
- Growing volume without proportional headcount growthA seasonal peak stops meaning a three-month hiring push. The automation scales with volume; people stay where they're irreplaceable.
- A team freed from work nobody wants to doRetyping data from PDFs doesn't develop anyone. Less of that work means less turnover — and less recruiting for roles that are already hard to fill.
- A number you can bring to the boardInstead of asking for more headcount — a before measurement, an after measurement, and the difference in money. The conversation about automation becomes a conversation about arithmetic, not about faith in technology.
We'll be honest with you, because the credibility of everything above depends on it: AI automation is a new line in our offering, and we don't yet have completed reference deployments to show you. Rather than pretend otherwise, we're offering an arrangement where both sides gain something: for our first clients, we offer a pilot on preferential terms in exchange for the right to publish the results — after you approve the content and scope. If you'd rather wait for someone else's references, we understand; but by then, the terms of the reference pilot will no longer be available.
Scope and pricing
Thirteen services — from the process-selection workshop to maintaining a live automation. Each has its own page with details:
Implementation path: workshop to select a process for automation · pre-deployment process measurement · AI automation pilot · production deployment · monitoring and tuning of deployed automation
Applications: document data classification and extraction · company-knowledge assistant / chatbot · customer service chatbot · voicebot · content generation system · automated content quality control · AI integration with ERP and CRM · custom machine learning model
The path is always the same — and you can stop at any stage:
- Process-selection workshop — two hours, fixed price. A review of your operational processes and the choice of one: the one that hurts the most and can be measured. You leave with a list of processes and a savings estimate — regardless of whether we go on to work together.
- Measuring the current state. Handling time, volume, error rate. This becomes the baseline for everything that follows.
- Setting the success threshold — together, before the pilot. You decide, on your own data and your own edge cases, what the automation has to achieve for a deployment conversation to make sense.
- Pilot on a data sample — fixed, capped price, known before it starts. An anonymized sample, an NDA, and a data-processing agreement before anything is handed over.
- Evaluation against the threshold. Threshold met — we discuss deployment, priced on the basis of the pilot. Threshold not met — we stop, and you know exactly what that knowledge cost: the price of the pilot, nothing more.
- Deployment with a human in the loop, plus monitoring. Inference cost billed separately and transparently — you know what the automation costs to run, not just to deploy.
Why does this structure pay off? Because it reverses the usual risk distribution with new technology. You're not buying a promise of deployment — you're buying an answer to the question "will this actually work for us," for a fixed, capped pilot price. The most expensive scenario in AI is a deployment that doesn't work; in this path, that scenario is structurally impossible, because deployment never starts without a threshold being met.
Our guarantees
You set the success threshold — before the pilot starts. On your own data and your own edge cases. We don't grade our own work.
No threshold met = no deployment, and spending stops. Your risk is the price of the pilot, nothing more. No "just one more sprint, this time it'll work."
Data under control. The pilot runs on an anonymized sample, under an NDA and a data-processing agreement. Processing takes place on infrastructure within the EU, or locally on your premises if the nature of the data requires it. You decide what leaves your network.
Human in the loop. Every decision with financial or legal consequences goes through a person. The automation prepares it; the person approves it.
Decision log for audit. Every response the automation gives is recorded and validated against its source — you can see where each piece of information came from. "AI makes things up" stops being a systemic risk, because fabrication becomes detectable and measurable.
Code, prompts, and documentation in your repository. Plus an abstraction layer that lets you switch model providers. No dependency on us or on a single vendor.
Availability
We run a limited number of pilots in parallel, because each one requires work on the client's data and weekly contact with someone on your side. We'll tell you the nearest start date for a team at the first conversation — no artificial pressure; we'd simply rather you knew the state of the queue before deciding, not after.
Propose a workshop date
Write or call and propose a date for the two-hour workshop. You'll come away from it with a list of processes suited to automation and a savings estimate — regardless of whether we end up working together. We'll be upfront about one condition now: we need an identified process, access to a data sample, and one decision-maker on your side. Without that, we won't take on the project, because it would end up as a demo nobody uses.
Something worth weighing before you shelve this decision
The alternative to automation isn't zero cost. The alternative is more headcount for work nobody wants to do, and more recruiting for roles that are already hard to fill. That cost is already in your budget today — just scattered across line items nobody adds up. Measuring the process adds it up for the first time; what you do with that number is your decision.
In short
One process, not the whole company · measured before and after · a pilot at a fixed, capped price · you set the success threshold, before it starts · data anonymized, processed in the EU or on your premises · human in the loop and a decision log · code and prompts in your repository · no threshold met means no deployment.
Frequently asked questions
AI makes things up. How can you base company processes on that?
Models can fabricate — which is exactly why architecture, not trust, is the answer. Responses are validated against their source, every decision has a log showing where the information came from, and matters with financial or legal consequences are approved by a person. In the pilot we measure the error rate on your data — and it's that number, compared against the human error rate in the same process, that settles the conversation.
I'm not handing over our data. How do you deal with that?
In stages, and with contracts. The pilot works on an anonymized sample, handed over after an NDA and a data-processing agreement. Processing takes place on infrastructure within the EU, and for especially sensitive data — locally, on your premises. What actually leaves your network is something you decide, and it's written down, not assumed.
Do you already have deployments? Who's using this?
We'll say this plainly here too: this is a new line in our offering, and we don't yet have completed reference deployments. That's why we offer our first clients a pilot on preferential terms in exchange for the right to publish the results — after your approval. The design of the success threshold means you don't have to take our word for it: the automation proves itself on your data, or you don't move to deployment.
How much does the pilot cost?
A fixed, capped amount, known before it starts — plus the workshop and the measurement beforehand, each with its own price. That's the complete list of costs up to the deployment decision; there's no line item for "additional work that turned out to be necessary." Inference cost in a live system is billed separately and transparently.
What if the pilot doesn't meet the threshold?
We stop — and that's a result, not a failure hidden in a slide deck. You get a report: what the automation achieved, where it got things wrong, and whether — and what — could change that. Sometimes the conclusion is "too early, let's revisit in a year," sometimes it's "this process isn't suited, but the one next to it is." Costs stop at the price of the pilot.
Will this replace my people?
It replaces tasks, not people: reading, classifying, retyping. Decisions and exceptions stay with your team — which stops being a data-entry machine and starts doing work that actually needs a person. In practice, automation more often lets you handle growing volume without new hires than it reduces the headcount you already have.
At what company size does this make sense?
The threshold isn't company size, it's process volume: if some task repeats often enough to take someone hours a week, it's a candidate for measurement. The workshop and the measurement are separate, low-cost stages precisely so a company can check whether it pays off before spending anything on technology. Sometimes the measurement concludes "this doesn't pay off yet" — and that, too, is a valuable answer.
Who will maintain this, and what happens if you or the model provider change?
Code, prompts, configuration, and documentation are in your repository from the start, and training your team is a condition of acceptance. An abstraction layer separates the system from any specific model provider — switching vendors means swapping a module, not rebuilding from scratch. You can have us handle monitoring and tuning, or take it over yourself — you have everything you need either way.