All services

AI, data and automation

Automation where the data exists and the process can carry it.

03
AI, data and automation

Machine learning models, data pipelines and automations that remove repetitive work and return information at the right time. We measure before we ship.

We start from what in your process is worth automating. We use the data you have, measure the results and put into production models that someone owns and monitors. A real example: large volumes of invoices in different formats, read without human intervention.

What it covers

  • 01Document extraction and classification
  • 02Machine learning applied to processes
  • 03Data pipelines and reporting
  • 04Agents and automations under supervision

When automation makes sense

Two conditions, both necessary: the data already exists in a form a machine can read, and the process can absorb an occasional wrong result without serious harm. If the first is missing, you build the data first. If the second is missing, the automation proposes and a person decides.

That is why we start by measuring: how many documents a month, how many formats, what a mistake costs today, how long a person takes. Without this baseline you cannot say whether the system works, and a model that looks good in a demo can cost more than the manual process it replaces.

Starting from scratch or automating a process that already exists

A process that already exists has one precious thing: history. Invoices already classified, cases already handled, emails already sorted. From there you build the evaluation set, and from there you know when the system is ready. That is how we approached the automatic reading of invoices in every format for a large infrastructure operator.

A new process, like a sales pipeline that finds the right companies, has to be built with the human inside from day one: the system gathers signals, computes a score and drafts a message; the final decision and every send stay with a person. Autonomy widens only when the measurement justifies it.

How we work inside a system in production

The automation enters alongside the process, at first in shadow: it produces its result, a person produces theirs, and we compare. Only when the difference is within the agreed threshold does the system take the wheel, and even then with a switch to go back and doubtful cases routed to a person.

Models are treated as dependencies that change. We log prompts, versions and evaluations, so a vendor update does not silently change behaviour. Where it makes sense, models stay interchangeable and data does not leave your perimeter.

What we actually deliver

The pipeline in your repository, with tests and the evaluation set. A dashboard with the metrics you chose at the start: accuracy, cases passed to a person, time per document, cost per run. Integrations with the systems you already use. And a document stating what the system can do, what it cannot, and how to switch it off.

What we also build for ourselves

We use the same tools on our own products: agents that generate and review content, an engine that turns teaching material into tests, sales scoring pipelines. It is the most honest way we have to know what holds in production and what only works in a demo, before proposing it to a client.

Questions

Before writing to us

When does a language model make sense, and when not?+

It makes sense where the input is text or heterogeneous documents and an error is recoverable. Where a clear, verifiable rule exists, we use the rule. We choose starting from the cost of error.

How is accuracy measured before production?+

On a real sample of your data, against a success criterion agreed beforehand. Only with numbers on your case do we decide what to fully automate and what goes through a person.

What data should a company never give to ChatGPT?+

No data you could not publish: personal data about customers and staff, health data, passwords and access keys, contracts and offers with real figures, unpublished financials, confidential code and projects. With a personal account that data leaves the company's perimeter without anyone knowing. A business plan with a data processing agreement lowers the risk, and written rules on what gets uploaded and where are still needed. In the systems we build, the data perimeter and the model provider are decided with you before any code is written.

What is shadow AI and how does a company handle it?+

It is the use of ChatGPT, Gemini or other assistants from personal accounts, for work, without the company knowing. A ban on its own usually moves the use to personal phones. What works better is providing an approved company tool, sorting data into three groups (public, internal, confidential or personal) and writing on one page what can be uploaded, with which tool, and who reviews the output before it goes out.

Describe the system. We'll tell you what we see.

Ten minutes, one question at a time. At the end, our honest read of the case. It's the filter, before the call.