FDE Műhely

Method

Discovery → Evals → Deployment

The loop doesn't stop at the first process. As soon as one workflow works, the next bottleneck becomes visible — usually immediately upstream or downstream of it.

Timeline

What the first quarter looks like

This is a typical schedule for a moderately complex back-office process. The dates move. The order doesn't.

  1. Week 1

    Observation

    We sit with the team. The first day is mostly quiet: they work, we watch. We ask questions when something unusual happens — the memory is fresh then, and we're talking about a real case rather than an abstract rule.

  2. Week 2

    Exception log

    We write down every deviation. That list becomes half the specification later. In parallel we pull the system logs: the ERP and CRM tell us what nobody says out loud.

  3. Week 3

    Map and matrix

    The operating map comes together, and we mark each step: what stays code, what becomes a model call, where a human stays in the loop. Alongside it, an ROI matrix of what to do first.

  4. Weeks 4–6

    Golden dataset and evals

    We build a measurement set from historical cases, run the system against it, and see where it stands and where the errors actually come from.

  5. Weeks 7–12

    Shadow mode

    The system runs but doesn't decide. Your team sees what it would propose and can flag when it's wrong. That feedback goes back into the evals.

  6. Week 13+

    Staged go-live

    High-confidence categories go automatic first, the rest to a human. The threshold moves on measurement, not impatience.

The decision

We decide step by step what goes to a machine

This is the real product of discovery: not "we're introducing AI", but a per-step answer for what should happen to each one.

  • Deterministic
  • Agent
  • Human
  • Exception
  1. 01

    Intake and normalisation

    Unpack attachments, drop duplicates, land on one format. No model involved — none is needed.

  2. 02

    Extraction

    From PDFs, screenshots, forwarded threads. Judgement belongs here, because no two senders are alike.

  3. 03

    Schema validation

    Required fields, totals, vendor master. If it fails, it doesn't move on.

  4. 04

    Matching and coding

    Tie to the purchase order and delivery note, propose a cost centre, flag uncertainty.

  5. 05

    Approval

    The finance clerk sees the proposal and the source on one screen. They press Enter, not the model.

  6. 06

    Write back to the ERP

    Through the existing API. We don't migrate systems, we build on them.

  7. 07

    Exception queue

    Whatever the system won't finish doesn't vanish — it queues up, with a reason.

Principles

What we don't compromise on

We build on what you already run

If someone spent two years rolling out an ERP, the last thing they need is another migration. We build on the APIs.

Everything is logged

Every step is reviewable: what came in, what the system decided, on what basis. There is no trust without it.

The human moves up, not out

Approval stays with a person. The system prepares, proposes and explains; the responsible colleague decides.

The exception is a deliverable too

What the system won't finish doesn't vanish quietly. It queues, with a reason, an owner and a deadline.

We build model-agnostic

Your own eval set decides which model runs. When a better one ships, switching is a measurement, not an article of faith.

We close in numbers

Hours saved, error rate reduced, risk avoided. If we can't state it in numbers, we don't claim it.

Frequently asked questions

What's the shortest engagement that makes sense?

A standalone discovery, which is two to four weeks. Any shorter and you can't map a process with its exceptions included — and the exceptions are the point.

Do you need to be on site?

For discovery, yes — typically two to three days a week. The build is mostly remote. If the team itself is distributed, we adapt.

What happens at the end?

Handover: documentation, runtime, eval set and training. The goal is that your team can operate and extend it. We'll stay on for support if you want, but we don't build dependency.