---
title: Before your first agent goes live, decide who may overrule it.
description: Who may overrule the AI, and who answers when it gets it wrong? A one-page decision-rights map for one workflow, and the classic study behind it.
---

[Blog](https://managementkits.com/blog)

# [Before your first agent goes live, decide who may overrule it.](https://managementkits.com/blog/who-may-overrule-the-ai)

 Written by [Olaf Bach](https://managementkits.com/blog/author/olaf-bach) | 28 Oct 2026

**When an AI tool enters a workflow, the organization around it changes whether anyone designs the change or not. Classic management research documented this long ago: Stephen Barley watched it happen in the 1980s, when two hospitals installed the same CT scanner and ended up with two different radiology departments. Your AI governance names the roles: use-case owner, data owner, user. This note is about where those roles meet the work: who may overrule the tool in your workflow, on what grounds, and who answers when it is wrong.**

Barley's study is one of the quiet classics of organization research. He observed two community hospitals from before each installed its first CT scanner through most of the year that followed. The scanners were the same. What formed around them was not.

- At one hospital, technologists who already knew the machine gained room to decide early on; as less experienced radiologists rotated in, the doctors withdrew to their offices and the technologists ran routine scans on their own judgment.
- At the other hospital, the radiologists kept the decisions and often ran the scans themselves, the technologists kept asking for direction, and an attempt to give them more independence slid back into the old pattern.

Barley published the findings in [Administrative Science Quarterly in 1986](https://doi.org/10.2307/2392767) and put his conclusion in the title: **technology is an occasion for structuring.** It starts the negotiation over who decides what. It does not decide how the negotiation ends.

**Forty years on, the same thing is happening on your floor, only faster.** An AI tool enters triage, drafting or review. Within weeks the unit has a new and unwritten order: who trusts the output, who re-checks everything, who has quietly stopped reading. If nobody decides, the order forms anyway, around the people with the most confidence rather than the most judgment. *A decision nobody owns is still being made, just by whoever is nearest.*

## Who may overrule the AI?

The person who is accountable for the outcome of that step of the work, on grounds that are written down before the tool goes live, and nobody by default. That is the short answer, and it is rarely what we find in place. Governance usually names a use-case owner and the users; it rarely says who may set the output aside at step three, and on what grounds. In most rollouts the vendor has defined what the tool does, IT has defined who has access, and the question of who may say no to the output, and when, has been left to the team to settle by habit.

It matters because the overrule is **decision rights at their smallest scale**, and decision rights are where organization design starts. When a tool proposes, drafts or triages, somebody has to be allowed to set it aside, and everyone has to know the grounds. Without that, your most experienced people face two bad options: sign off work they have not really read – or, as AI gets ever more powerful, even understood – or redo everything and treat the tool as a demotion. Both happen, often in the same team. And the grounds matter as much as the name. In one of the best-known controlled studies, “Navigating the Jagged Technological Frontier” by Fabrizio Dell'Acqua and colleagues, published in Organization Science, [758 consultants](https://doi.org/10.1287/orsc.2025.21838) did markedly better with AI on tasks inside its competence, and markedly worse than colleagues without it on a task just outside it, where the tool's answer sounded right and was not. Knowing where that edge runs in your own workflow is what an overrule is for.

The best field evidence from the current wave points the same way. A Stanford team studied [51 enterprise AI deployments that had actually worked](https://digitaleconomy.stanford.edu/publication/enterprise-ai-playbook), across 41 organizations, and asked what separated them from earlier failures in the same companies. **Deployments designed so that the system handles the routine and people take the exceptions reported more than twice the productivity gain of those where a person approves every single output.** And the most frequent resistance came from legal, HR, risk and compliance, ahead of the people actually using the tools. That is not a complaint about staff functions. It is what happens when the people whose job is to ask who is accountable are handed a tool without an answer. One caveat: the study looked only at deployments that worked, and the companies reported their own numbers, so take the size of the gap as a direction, not a proof.

At Management Kits we take a clear position here. The overrule rule is the one organizational decision in an AI rollout that you cannot hand to the vendor, to IT or to a programme office, because it is a statement about your people's judgment. Made explicitly, it turns the tool into an instrument. Left open, it turns the tool into a status question, and status questions get settled silently, which is the worst way to settle anything – not to speak of regulatory risks. AI governance that has an impact is not a board reviewing templates and use-case lists; it is how everyday decisions get made at the point where the tool's output meets someone's judgment. We made a related argument in [our post on what organizations really reward](https://managementkits.com/blog/ask-your-team-what-really-gets-you-promoted-here.-ask-them-separately): people read the system correctly, so the system had better say what you mean.

*Working on this in your own organization? [See how the MK Practice works →](https://managementkits.com/practice?utm_source=blog&utm_medium=cta&utm_campaign=p1-practice&utm_content=pn02-mid)*

## How do you map decision rights for one AI workflow?

The move is one page, the decision-rights map. It takes a week because the conversation at the end matters more than the page.

*We made the map into a canvas you can print on A3: [download the Decision-Rights Map](https://managementkits.com/decision-rights-map?utm_source=blog&utm_medium=cta&utm_campaign=p1-practice&utm_content=pn02-canvas).*

1. **Pick one workflow the tool has already entered.** Not the whole unit, one flow: incoming cases, first drafts, a reporting step. If the tool is still in a pilot, pick the pilot.
2. **Write down the steps where the tool acts.** Usually three to five: where it proposes, drafts, sorts or flags.
3. **For each step, write four names.** Who decides. Who may overrule, and on what grounds, in one line. Who reviews, and how often. Who answers when the output is wrong and nobody caught it. Names, not roles, for now.
4. **Leave the blanks blank.** A cell you cannot fill is the finding.
5. **Spend thirty minutes on the blanks with the two people whose names appear most.** They will fill half of them from what they already do, and tell you which of your entries is fiction.
6. **Log the first real overrule.** When someone sets the tool aside, write down who, on what grounds and what happened next. That entry is the first line of a quality norm, which a later note in this series takes up.

In our work with leadership teams, the blanks cluster in the same column every time: "decides" and "reviews" fill up quickly, and almost nobody has thought about the overrule, least of all the grounds. It is also the question the most experienced people ask first, usually in a meeting and usually too late. Asking it on paper, with names, before they ask it out loud is most of the value of the map.

If you have read [our earlier piece on delegating power rather than tasks](https://managementkits.com/blog/2020/6/16/from-delegating-tasks-to-delegating-power), you will recognize the logic: a right that is not explicit is not delegated, it is merely assumed. In the MK Practice, the follow-through sits in the working topic *Ensure clarity on decision-making*: a milestone of five actions, from clarifying the current process to a debrief after the first significant decision, that goes onto your work agenda and comes back for review.

A decision rule we find useful: **if the tool is live and the overrule column has a blank, that blank is where your next quality incident comes from.** Make the map this week, hold the conversation next week, look at it again after the first overrule, and read it once more a quarter in. If a quarter passes and no overrule has been logged, nobody is overruling, which means either a very good tool or a team that has stopped reading. You will want to know which. And if you wrote the one page about your own unit that [our opening note](https://managementkits.com/blog/your-organization-is-changing-faster-than-you-are-learning-to-lead-it) asked for, put the map next to it and see whether the blanks sit where that page said "nobody has decided yet".

## What the map will not do for you

Obviously, a map cannot fix a tool that is wrong; if the overrules pile up in one step, that goes back to whoever bought or built it. And some rights sit above your level: where a committee or a regulator sets the grounds, the map records that and escalates the blank rather than inventing an answer.

Back to the two radiology departments. Neither chose its new order. It happened to them around a machine, over months, through hundreds of small deferrals and refusals that nobody wrote down. Yours can be chosen, on one page, this week, with the two people who will have to live with it.

The next note is about the rest of the unit, and why your team is not resisting the tool so much as resisting being redesigned without a say.

You don't read a practice. You run it.

The MK Practice is a structured process you run on your own organization: a diagnostic that reads where your unit needs work first, working topics like the ones in this series, a work agenda that keeps the moves coming, and a practitioner who looks at your problem within 48 hours. Founding Member: USD 990 a year, 30-day money-back guarantee.

[See how the MK Practice works →](https://managementkits.com/practice?utm_source=blog&utm_medium=cta&utm_campaign=p1-practice&utm_content=pn02-decision-rights)

*Want to talk it through first? [Book 20 minutes with a Management Kits team member](https://calendly.com/managementkits-demo/talk-first?utm_source=blog&utm_medium=cta&utm_campaign=p1-practice&utm_content=pn02-call). Running it with a whole leadership team? [See the MK Practice for organizations](https://managementkits.com/enterprise-solutions?utm_source=blog&utm_medium=cta&utm_campaign=p1-practice&utm_content=pn02-org).*

[View full post](https://managementkits.com/blog/who-may-overrule-the-ai)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Olaf Bach"
  },
  "dateModified" : "2026-10-28T23:14:30.313Z",
  "datePublished" : "2026-10-28T23:14:30Z",
  "headline" : "Before your first agent goes live, decide who may overrule it.",
  "image" : {
    "@type" : "ImageObject",
    "height" : 1400,
    "url" : "https://26208959.fs1.hubspotusercontent-eu1.net/hubfs/26208959/Management-Kits_Decision-Rights-Map.png",
    "width" : 2400
  },
  "mainEntityOfPage" : "https://managementkits.com/blog/who-may-overrule-the-ai",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60,
      "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
      "width" : 60
    },
    "name" : "Blog"
  }
}
```