Who still decides when the system acts
By Sascha Cutura

Every automated process makes decisions. Some of them are obvious: approve this invoice, route this enquiry, flag this order. Others are buried in rules that someone wrote two years ago and nobody has read since. Either way, a decision is being made, and when it goes wrong, someone will be asked to explain it.
In most organisations we work with, nobody can say in advance who that person is. The question only gets asked after something has gone out to a customer, been paid to the wrong supplier, or been refused when it should have been approved. By then the conversation is about blame. It is much more useful to have it before the automation goes live, when it is still a conversation about design.
Automation moves decisions without announcing it
When a process is manual, decision rights are usually clear even if nobody wrote them down. The credit controller decides whether to chase. The operations manager decides whether to expedite. People know who to ask.
Automate that process and the decision does not disappear. It moves. Part of it goes into the rules, part of it goes to whoever configured the tool, and part of it falls into a gap between the two. The credit controller now receives a report of what the system already did. The operations manager finds out about an expedited order when the courier invoice arrives.
None of this is a failure of the technology. The tool is doing exactly what it was set up to do. The problem is that nobody decided, deliberately, which parts of the judgement should move into the system and which should stay with a person.
Three questions for every automated decision
Before a process is automated, ask the same three questions about each decision point in it.
What happens if this decision is wrong? A misrouted internal ticket costs a few minutes. A wrongly declined customer application costs a relationship, and possibly a regulatory complaint. The cost of a wrong decision tells you how much human attention it deserves.
How often does this decision need judgement? Some decisions are genuinely rule-based. If the purchase order matches the invoice and the goods receipt, pay it. Others look rule-based until you watch people make them and notice how often they pick up the phone first.
Who will answer for it? One named role, with the authority to change the rule. If the answer is "the system" or "IT", the decision has no owner.
The answers usually sort decisions into three groups. Low-cost, high-volume, genuinely rule-based decisions can be fully automated, with a named owner who reviews the pattern each month. Medium-risk decisions can be prepared by the system and confirmed by a person. High-cost or high-judgement decisions stay with a person, and the system supports them with better information.
Where people stay in the loop
"Human in the loop" has become a phrase people use to reassure each other. In practice it often means one person clicking approve on two hundred items a day without reading them. That is a signature, and it gives a false sense of control.
A useful loop has three properties. The person reviewing has enough context to disagree. They have enough time to look properly. And their decisions feed back into the rules, so the system gets better at preparing the next case. If any of those is missing, the review is decoration.
This is also where volume matters. If a process produces more exceptions than a person can review properly, the answer is to fix the rule or the upstream data, rather than to ask the reviewer to go faster.
The government's AI adoption survey for 2026 found that 81% of UK businesses reported applying some level of human input or oversight to AI-generated outputs. Oversight only has meaning if the person involved can still change the outcome.
Write it down before you switch it on
The most practical thing a leadership team can do is produce a one-page decision register for each automated process. For every decision point: what the system decides, what a person decides, who owns the rule, what the threshold is for escalation, and how often the owner reviews the outcomes.
It sounds administrative. It is the document that will be asked for first when an auditor, a customer or a board member wants to know why something happened. It also forces the uncomfortable conversations early. It is quite common for the first register meeting to reveal that two departments each assumed the other was checking the same exception queue.
You do not need a full governance framework to begin. Pick one process that is already automated or about to be. Walk through it with the people who run it today and mark every point where a decision is made. Then answer the three questions for each one. You will probably find a decision nobody owns. Fixing that one is a good afternoon's work.



