Not every job needs the top rung. How much of the work stays inside the system is a decision: the higher you go, the less manual effort — and the more oversight it takes.
How much should the system do?
Whatever system we build, one question stays the same: how much of the work stays inside the system? There are four steps. With each one the manual effort falls and the need for oversight rises.
01 — Integration
Systems talk to each other
Between two systems the rule is clear: which data, when, in what format. The connection is built once and runs every day.
Exporting, fixing and re-importing goes away.
02 — Automation
Work runs on its own
Nobody has to start the step that repeats: a trigger, a schedule, a condition.
Who hears about it when something goes wrong is part of the rule — automation that fails silently is worse than none.
03 — AI integration
The system reads and understands
For the places where no rule can be written: free text, invoices, emails, contracts, voice recordings.
The model pulls the fields out, the software checks them, a person approves the result. The model estimates; the checking happens in the software.
04 — AI automation
The system decides and acts
Understanding and doing meet: sorting, routing, drafting a reply, opening a record.
Where it carries on alone and where it waits for your approval is settled up front; every decision is logged.
Most jobs are solved on the first two steps. If a rule can be written, no AI is needed — and we will tell you so.
Rule or model?
This distinction drives the cost and the risk of a project, so we settle it first.
If a job can be solved with a rule, solve it with a rule. Rules are cheap and fast, the result is the same every time, and why it came out that way is traceable. AI guarantees none of that.
A model wins where no rule can be written: when the same information arrives in a different shape every time. Fifty suppliers with fifty invoice layouts; requests written in free text; documents with no standard.
In practice the two work together: the model reads and proposes, the software checks, and whatever falls outside the rule goes to a person. Leaving the checking to the software rather than the model is the most important design decision here.
Where your data is processed
Sensitive data are processed with open models on our own servers; the data stay within the EU. For everything else we use commercial services. What is processed where is put in writing in the discovery output — it is not a topic for later.
Running a flow that uses AI is also different: it needs a cost ceiling, a log of what it produced and a regular sample check. We set those up from the start; what it costs and how reliably the system reads are both in the monthly report.
Let us start with a discovery.
Tell us, we will listen. Discovery is short and paid; at the end you hold the process map, scope and budget range. If we conclude that an off-the-shelf tool would do the job, we will say so just as plainly.
DARFA · Berlin
We use your details only to answer this enquiry and never pass them to third parties.