Decision Intelligence

The first decision you automate should be reversible

Not the most valuable one. Not the most frequent. The one where a mistake costs a correction rather than a customer.

First decision to automate: bu ne anlama geliyor

First decision to automate: from data through prediction to a recorded decision

The instinct is to automate the highest-value decision first, because that is where the return is. It is the wrong instinct, and it is why so many programmes never get past the first deployment.

The first automated decision is not primarily a value exercise. It is a trust exercise. Its job is to establish, with evidence, that the organisation can let a system act and survive being wrong.

The three properties to look for

Reversible: a mistake produces a correction, not a loss. Frequent: enough instances to learn from within weeks, not quarters. Bounded: a small action space where the worst case is known in advance.

Value is the fourth criterion, not the first.

What this looks like in practice

Stock transfers between locations rather than purchase orders to suppliers. Markdown recommendations on slow lines rather than price changes on headline products. Rescheduling within a shift rather than committing capacity.

None of these will impress a board. All of them produce the operating record that makes the next conversation possible.

Why the high-value decision resists automation anyway

It is not caution that blocks it. It is that a high-value decision is usually made rarely, which means the system cannot accumulate evidence at any useful rate.

A capacity commitment made quarterly produces four observations a year. Nobody will let an unsupervised system make the fifth on the strength of the first four, and they are right not to. Frequency is not a convenience; it is what converts a deployment into knowledge.

Reversibility has a specific test

Not “can we undo it” in principle. Ask: what does the correction cost, and who has to be told?

A stock transfer reversed the next day costs a truck movement and nobody outside operations learns of it. A purchase order reversed costs a supplier relationship and a conversation. A price change reversed costs customer trust and cannot really be reversed at all — the customer saw the price.

The three are technically all reversible. Only the first is reversible in the sense that matters for a first deployment.

What the first deployment is actually producing

An override rate, by decision class, over a few hundred instances.

That number is the asset. It tells you where the system understands the operation and where it does not, and it converts the next scope conversation from an argument about principle into a reading of evidence.

An organisation that has automated one small decision for a quarter can say: this class of decision was overridden four percent of the time, and here are the four cases. An organisation that started with the flagship decision has a stalled pilot and an opinion.

The board conversation

Reversible-first is a harder sell internally than high-value-first, and it is worth stating why plainly rather than dressing it up.

The pitch is not that the first decision is valuable. It is that the first decision buys the right to make the second one, and that the alternative path — start big, stall, restart — costs a year.

Most boards accept this when it is put in terms of sequencing rather than ambition. What they do not accept is a pilot whose success criteria were never defined, which is a separate and more common failure.

When to widen

When the override rate has been stable and low for long enough that a reviewer stops reading every case. That is the practical signal, and it usually arrives before anyone has agreed on a threshold.

Widen along one axis at a time — more instances of the same decision, or the same volume of a slightly larger decision. Doing both at once means a change in the override rate has two possible causes and teaches you nothing.

Request a Demo