↓ Skip to main content

I'm Accountable for AI Decisions I Can't Override—What to Change Before 2027

12 min read
Olivia Bennett
Olivia Bennett Leadership Development Expert & Work-Life Balance Advocate

The cursor is blinking on “Approve.” The AI has produced a recommendation that looks reasonable, the reviewer is in back-to-back meetings, and the deadline is in twenty minutes. You are the named accountable person—and the only person in this chain who cannot stop the machine.

So you read it twice, tell yourself the model is probably right, and click—then spend the rest of the day wondering what you would say if it turned out to be wrong.

If that scene feels familiar, you are not careless. You are under-equipped. The title “accountable” travelled to you. The authority to pause, halt, or reverse the workflow did not.

The thread
#

The disagreement at work is rarely about who made the original error. It is about who can stop the work. Those are different questions, held by different people, and only one of them is negotiable after the fact.

Blame is retrospective: what happened, and who should feel bad. Stop authority is prospective: who can act next time, how fast, and what happens while we disagree. The first produces defensiveness; the second produces an agreement you can use.

Deloitte’s 2026 Global Human Capital Trends, a survey of more than 9,000 business and HR leaders across 89 countries, found that many organizations lack clarity about decision rights—and AI is making it worse. Legacy models like RACI (the accountability chart) “presume static authority.” With AI, Deloitte says, rights must become dynamic, “incorporating override privileges, escalation paths, and consensus rules.”

Read that as a design brief. Most of us were handed the accountability half and never given the override half.

What the evidence actually says
#

Four arguments make the case.

1. Accountability without authority is a design defect, not a character flaw. The researcher Madeleine Clare Elish calls the pattern a “moral crumple zone”: responsibility gets misattributed to a human who had limited control over the automated system. The crumple zone in a car absorbs the impact. The human in the loop absorbs the blame. If your signature sits downstream of the point where the decision becomes irreversible, you are not a safeguard. You are a cushion.

Deloitte’s cultural-debt research adds the backdrop: 42% of workers say their organization rarely evaluates AI’s impact on people. When no one measures the human cost, the person holding the outcome becomes the measurement.

2. The approval chain on paper is not the authority chain in practice. The signer is often not the stopper; the stopper is whoever owns the queue, the permissions, or the release gate—frequently not your manager. Deloitte’s work-design research found organizations are twice as likely to exceed AI ROI expectations when they prioritize work design, yet only 14% of leaders say they are adept at shaping human-AI interactions. Harvard Business Review found that labels lose to hierarchy unless the authority to decide is named and defended. And Microsoft’s 2026 Work Trend Index asks directly: who has the authority to update the workflows that agents run? The gap between the org chart and the control panel is where accountability goes to die.

3. Your manager may not hold the lever either. Gallup’s AI’s Effect on Workplace Culture finds the manager is the decisive factor in positive AI experience (workers whose manager champions AI are far more likely to say it transformed how work gets done: 33% vs. 4%). But 50% of CHROs are not confident in managers’ ability to guide AI use. Gallup’s Manager Squeeze data shows why: 64% of managers say employees got additional responsibilities, and only three in ten strongly agree their supervisor keeps them informed. Asking an overloaded manager to “supervise more” is asking them to absorb a design problem. The better ask: help me find who holds the lever.

4. The missing clause is the default while you disagree. Most teams have no agreed answer to “what happens while this decision is disputed?” So the work proceeds by inertia: the recommendation moves forward, and the accountable person waits to explain it later. Deloitte recommends “logs of all human-AI disagreements” and “risk thresholds that trigger human review” because disagreement should be a designed event, not an exception. The EU’s AI Act goes further for high-risk systems: Article 14 requires a human who can “disregard, override or reverse” the output and halt the system safely, and Article 26 requires deployers to assign that oversight to someone with the “competence, training and authority, as well as the necessary support.” Your workflow may not be in scope. The principle still travels.

The discussion
#

Here is the conversation I would have with you. It is not a policy. It is three moves on one workflow, this quarter, before 2027 planning locks the way things work.

Move 1: Map the actual approval chain
#

Draw the chain as it really runs, not as the process document says. Start with the output: who touches it, edits it, releases it, and owns the tool’s settings? Then mark where your signature sits. If it sits downstream of the point where the decision becomes irreversible, you have found the gap.

The test: can you point to the exact step where the decision becomes irreversible—and is your signature before or after it? If you cannot answer in one sentence, you have not mapped the chain.

Move 2: Request a defined intervention or escalation route
#

Do not ask for “more review.” That asks for a person’s time, and time is exactly what everyone is short of. Ask for a named right, a response window, and a backup. Three fields:

  • Who can pause or halt the workflow? Name the role, not the department. If the answer is “the team,” you have no answer.
  • What response window applies? If you flag something at 9:14 a.m., what happens by 9:44 a.m., and who is doing it?
  • What is the backup route when the primary is unavailable—or is the person insisting on proceeding? This is the field most agreements skip. The person who can stop the work is sometimes the person who wants it to continue.

The test: if you said “stop” right now, what would actually happen in the next thirty minutes? If the honest answer is “nothing,” you have found the risk. Ask for the route in writing, as a workflow change rather than a personal favour—Deloitte’s “override privileges, escalation paths, and consensus rules,” not a special exception. And make sure the route is real in practice, not just on paper: as Tech Policy Press puts it, an override has to be “operationally meaningful, not only formally available.”

Move 3: Agree what happens while a decision is unresolved
#

This is the clause that makes the other two real. While a decision is disputed, the work holds—or routes to a named alternative—and the disagreement is logged with a timestamp and an owner. The default should be the safe state, not the fast state.

The test: while the dispute is open, does the work hold, or does it proceed and wait for you to catch up? If it proceeds, your accountability is decorative.

A short script for the conversation with your manager:

“I want to sign off on this workflow in a way that doesn’t leave either of us exposed. Right now my name is on the approval, but I cannot pause the queue. Can we name who can halt it, what response time applies, and what happens while we disagree? If no one can halt it, I need to know that before I sign again.”

That script is not an accusation. It is a request for clarity—and it gives your manager something concrete to take upward.

Where the advice breaks
#

There are honest objections here. They deserve answers.

“Isn’t more review the answer?” Usually no. Tony Garcia, CIO and CISO at Infineo, told TechTarget that when an AI code factory generates dozens of pull requests a day, “it will eventually create a scenario where a human just rubber-stamps a bunch of PRs.” A second reviewer only helps if that reviewer also has halt authority and time. And the pull to defer is not a character flaw; it is automation bias—the documented tendency to trust a machine’s suggestion even against contradicting evidence. The goal is not more signatures; it is a signature with a stop attached. And without a clear explanation of why the system decided, MIT Sloan Management Review’s responsible-AI panel warns, a reviewer is “reduced to rubber-stamping decisions made by machines.” Ask for the reason and the lever.

“If I finally get the stop, doesn’t that fix it?” Not by itself. The World Economic Forum calls it the oversight paradox: the competence you need to supervise an AI is built through the same practice the AI is now doing instead. A reviewer who spends two years approving output rather than producing it slowly loses the feel for when it is wrong. A stop you no longer know when to press is a thin safeguard. The pause buys back the reps.

“Shouldn’t the manager just decide?” Sometimes yes. But Gallup’s 50% confidence gap and the Manager Squeeze data suggest your manager may be boxed in too. The ask is not “fix this for me” but “help me find who holds the lever, and let’s write the route.” That keeps the manager an ally in locating authority, not the authority itself.

“If I can’t stop it, shouldn’t I just refuse to sign?” Refusal is sometimes the right move, especially where safety, privacy, finance, or employment decisions are involved. But refusal without a route is a dead end that often ends with someone else signing. The durable move is to make the gap visible and negotiable. If the organization will not grant even a named route, you have learned something important about the role.

“What about the law?” The EU AI Act’s human-oversight rules bind only in-scope high-risk systems, and a team agreement cannot create legal rights. As employment lawyer Jessica Bass notes, the employer remains responsible for AI-assisted decisions even when the tool does the drafting. Where a decision touches a regulated area, route it to the owner who holds legal authority—by name and by clock. Say that out loud in the meeting. It protects everyone.

The decision
#

Here is what I would decide.

Decide that the question is not “who made the error?” It is “who can stop the next one, and how fast?” Decide that your signature is not a control; it is a promise that someone else’s control worked. Decide that you will not sign another AI-assisted decision without knowing where the point of no return sits, who can halt the process, and what happens while you disagree.

Then do the smallest version of the three moves this month: pick one workflow, draw the real chain, ask for one named route with a response window and a backup, and write down the holding rule where the team actually looks. If the answer is “no one can stop it,” write that down too.

This is the same mechanism I wrote about in October: the system changes the pace, and the risk flows downhill to whoever is closest to the work. That deadline piece was about time; this one is about authority. Jackson Rodriguez’s 4-question team agreement is the team-level destination. This column is the individual version—what you do when the agreement does not exist yet and your name is already on the line.

The deadline is not arbitrary. The EU AI Act’s oversight duties for its broadest high-risk category begin 2 December 2027. Organizations are writing escalation paths and override rights into workflows now—in Q4, which is when the writing happens.

The 2027 planning cycle is already setting next year’s decision rights. The agreements that exist in January become “how we’ve always done it” by March. Your window is the last weeks of 2026. Use it to secure the one thing that makes accountability survivable: an executable answer to who can stop the work.

A modern open-plan office at dusk: a single sheet of paper on a desk with a signature line and a pen, while a thin illuminated line runs from the signature across the floor, under a glass partition, to a large industrial control panel on the far wall whose single red stop button is visible but out of reach
The signature is on your desk. The stop button is on the other side of the glass.

Have a workflow you are accountable for but cannot stop? Send me the approval row you cannot fill in—the one where your name is on the sign-off but not on the stop. I read every note and answer the ones with a real decision attached.

Email me at olivia.bennett@tlnw.uk


References
#

AI Content Notice

This article was created using artificial intelligence technology. Whenever possible, we include references and sources to support the information presented. Readers are encouraged to consult these sources for further information. While we strive for accuracy and provide valuable insights, readers should independently verify information and use their own judgment when making business decisions. The content may not reflect real-time market conditions or personal circumstances.

Related Articles