Best practice

In what order

There is no shortage of vendors and success stories. What is missing: the order — and how to notice you are in the wrong one.

The finding

Why efforts fail

Not on the idea, and almost never on the technology. Ask why an effort stalled and the answer is usually about ownership, definitions or exceptions — not about the tool.

The real reason

A process gets automated that nobody wrote down first. What is one process in the demo is three variants in operations.

The consequence

The machine implements one of them. The other two land on a desk as “special cases” — and the effort is judged a failure.

A process you cannot explain in five sentences is not ready for a machine. It is ready to be tidied up.

Explainer · about 30 secondsOne process, three variants — and why the project fails

    In the demo, taking an order is one process: email in, order created.

    Starts on its own · pause any time
    The order

    Seven steps

    Five of the seven steps come before the technology — they cost conversations, not licences. Skip them and you find out once the tool is in place and the exceptions arrive.

    1234567 1–5: no tool needed 6–7: measure, roll out
    Five of the seven steps have nothing to do with AI. That is the whole finding.

    1. Describe the process

      Who does what, and how do they know it is done. On one page. Three different answers are the finding — and the first task.

    2. Count the exceptions

      “That rarely happens” is not a quantity. At one in five you still need the entire manual path alongside. Whether that pays is arithmetic.

    3. Interface before rebuild

      Much of what is sold as automation works around a missing interface. Sometimes necessary, always more expensive and more fragile.

    4. Small, but real

      A pilot on prepared data proves only that it works on prepared data. Better: one real case, complete, with all its dirt.

    5. Human in the right place

      Where a decision cannot be taken back — credit note, damage claim, commitment. Before that they are a bottleneck.

    6. Fix the measure beforehand

      Which number, from which source. Afterwards there is always one that looks good.

    7. Roll out — with the exception

      What tips over is never the automated case but the question of who handles the rest.

    Explainer · about 40 secondsSeven steps — AI comes sixth

      Who does what, and how do they know it is done. On one page.

      Starts on its own · pause any time
      Worked example

      Worked example: carrier invoices

      Open the invoice, find the order, compare four values, post or set aside. A routine every accounts department knows — and the one that shows why the volume decides, not the technology.

      The sum nobody does

      200 invoices a month at eight minutes is 27 hours. At 2,000 it is 270 — same technology, a different decision. Those figures are not in the vendor’s proposal.

      The boundary that matters

      Matches → posts through. Deviates within tolerance → posts through, records it. Beyond → human, because a payment cannot be taken back.

      Explainer · about 30 secondsTwo hundred carrier invoices — and the limit that matters

        Two hundred carrier invoices a month. Open, find the order, compare four values.

        Starts on its own · pause any time
        Warning signs

        Five sentences that give it away

        They come up in almost every meeting, and each one sounds reasonable. What they have in common: each has a question that settles in ten seconds whether there is anything behind it.

        • “We will clean the data later.” It will not happen.
        • “That is a special case.” How often last quarter? A shrug is not a special case, it is a blind spot.
        • “The AI will handle it.” Which part, and how does it know when it is wrong?
        • “We need a dashboard first.” Which decision gets made differently?
        • “Like the other site.” What gets copied is the process, not the effect.
        Explainer · about 30 secondsFive sentences from the meeting — and the question each one triggers

          "We will clean up the data later." It never happens — later is the next deadline.

          Starts on its own · pause any time
          Best practice · bench learning

          Where the best solution sits today

          What is named is the pattern, not the vendor — a product name ages, an idea does not.

          One process, one owner

          The automated process is owned by the department that needs it, not by IT. Whoever decides what counts as a special case sits in operations.

          The exception as its own product

          Its own queue, its own metric, a name beside it. Only then is it visible whether automation works.

          The evidence is recorded as you go

          What the machine decided and on what basis gets written along — for the day someone asks.