AI Has to Earn Its Decision Rights
Big companies say they want AI involved in decisions, then stall, because nobody can answer one question: who is responsible when it’s wrong?
Plenty of executives have said something like “we should be bold and let AI take part in decisions.” Then nothing happens. What usually gets in the way is a question nobody can answer: if the AI gets it wrong, who is responsible?
One workable approach is to grant AI authority the way we grant it to new employees, letting it earn more step by step. Every decision goes into a log and can be reversed. Autonomy comes in levels, and each move up needs someone’s signature. Whether a type of decision moves up or down depends on the human veto rate, a number available every day and a more useful one than accuracy.
Responsibility has nowhere to land
Letting AI make a decision means someone has to answer for what happens when it’s wrong. The business owner doesn’t want to, since they didn’t build the system. IT doesn’t want to, since business judgment isn’t IT’s job. Legal is responsible for compliance risk, so the safest answer is naturally not to let it decide.
All three concerns are reasonable, but the result is that responsibility can’t land on a specific person, so decision rights can’t be handed over. The AI ends up labeled “for reference only,” and references can be ignored.
Yet we’ve never given people authority all at once either. A new employee’s work gets checked for the first three months before anyone lets go. Finance has approval limits: a manager signs off below 50,000 yuan, a VP above 500,000, and those lines move up with seniority and down after mistakes. Authority comes in levels, moving up needs a reason, and it can be taken back at any time. AI can be given authority the same way. In a system, that takes three things: a log, levels, and people who sign.
The log
Every decision the AI makes gets an entry: what was decided, which data and rules it was based on, what happened afterwards, and how to reverse it if needed.
The first three make a decision auditable. The last one is what makes people willing to use it. Few people will let a system make decisions that can’t be undone; if a decision can be reversed with one click, trying it costs far less.
The log has one more benefit. “Is the AI reliable?” is a hard question to answer. With a log, it becomes “which of last month’s three hundred decisions went wrong?”, and that’s one a meeting can actually work through.
Levels, and why the veto rate
Autonomy has four levels. The AI only suggests, and a person decides. The AI fills in the answer, and a person confirms each one. The AI’s answer takes effect by default, and a person can veto it within a set time. The AI acts directly, and people spot-check afterwards. Each type of decision gets its own level and moves up or down on its own.
Moving up or down needs a basis. Most people’s first thought is accuracy, but accuracy needs a right answer to compare against, and most business decisions don’t have one. Whether a visit was well scheduled may not be clear for a quarter, or ever. Waiting for accuracy makes promotion very hard.
The human veto rate is more practical: the share of the AI’s decisions that people change. It’s available every day, without waiting for outcomes. And it reflects the thing that matters most when granting authority: whether the person responsible trusts the AI. A veto rate that keeps falling is grounds for moving up a level. If it suddenly jumps, the level drops automatically and a review is triggered.
The metric has a loophole too: people get lazy and approve everything without looking, so approved decisions also need regular spot checks.
The people who sign
Each level has to map to a specific person. At the confirm-each-one level, responsibility lies with whoever confirms. At the default-takes-effect level, it lies with whoever holds the veto, and with whoever approved moving that type of decision up to this level.
The key point is that moving up a level is itself a decision someone signs. The system doesn’t promote itself quietly. A responsible person reviews the log and the veto rate and signs off: this type of decision can now be left to the system.
That gives “who’s responsible when AI is wrong?” an answer, the same one we give for a subordinate’s mistakes: the person who granted the authority.
A by-product
Once this is running, and every veto has to come with a reason, it builds up data the company never had before: why people overrule the AI.
“This doctor retires next month. The system doesn’t know that.” “This hospital just got a new director. Now isn’t a good time to go.” “The number is right, but this isn’t the moment to change it.”
These reasons used to live only in the heads of front-line staff and managers. Nobody wrote them down, and they left when the people did. Now every veto leaves one behind. Data like this can only grow out of a company’s own business. It can’t be bought.
Of course, all three of these cost something themselves. Make them too heavy, and the effort spent on approvals and paperwork can exceed what the AI saves.
