AI Accountability for Leaders: Taking the Bullet for a Machine

Adam Peddicord

11 minutes ago

8 min read

Diagram of a moral crumple zone showing blame settling on the human closest to an automated system

I hadn't heard this one before. I was asking a couple of lawyers whether AI had actually helped them win anything.

What I got instead was a horror story making the rounds in their ranks.

A peer, in the middle of a trial, was asked by the presiding judge who made a particularly grievous error.

The peer knew it was the AI. Instead of telling the judge the tool was the root cause, they took it themselves, saying the mistake was theirs and they had not done a good enough job.

Hearsay like this is risky. I was not in the courtroom, and neither were the lawyers telling me.

What actually happened matters less than the question it left me sitting with.

Leaders face this every day. Where am I accountable? Where is someone, or something, else responsible? And when do I take the bullet for the team? For the record, I have taken real ones for a team. The metaphor is a great deal more appealing than the experience.

The stakes are climbing, and AI accountability for leaders is no longer a theoretical question. If you own retention, renewals, growth, or clinical quality, it is already sitting in your queue: when AI gets it wrong, who should take the fall? Let's see what the data says.

The Moral Crumple Zone

Here is the part that surprised me. We have been living in this world far longer than most people realize, and a lot of the protocol is already written.

In 2019, Madeleine Clare Elish published a paper in Engaging Science, Technology, and Society that coined the term “moral crumple zone.” It describes how responsibility gets misattributed to a human who had almost no control over what the automated system did. Your cruise control falters and you rear-end someone. A model update scrambles your health score weighting and torches QBR reporting. Technically, software errors.

Technology fails. The human checkpoint takes the hit.

You've been crumpled.

Right now the penalties look like a legal-industry problem. They will not stay that way. The AI Hallucination Cases Database logged 2,039 cases as of September 2026, up from roughly 200 in mid-2025. Every sanction landed on a human. They run from a $5,000 fine to about $109,700 in a single matter, and they now include bar suspensions. According to the database, no court has sanctioned a model. Not one.

Healthcare is feeling it too. Physician use of AI tools jumped from 38% to 66% between 2023 and 2024, and the FDA had cleared 1,300 AI-enabled devices as of January 2026. There is still no US malpractice verdict with AI at its center, and that is the problem. Courts are applying ordinary malpractice principles, so the physician is still the focal point of liability.

If you have wondered why consent forms, attestations, and sign-off gates keep multiplying, this is why.

If humans are absorbing the blame for the machine's errors, what is being done to level the field?

Not much. Here is what you should know:

There is no comprehensive federal AI statute. The main federal instrument, the NIST AI Risk Management Framework, is voluntary guidance. Useful, but it creates no duty and it protects nobody.
The federal posture is deregulatory and unresolved. Executive Order 14365 directed a national framework that would preempt conflicting state law, and that recommendation landed in March 2026. Preemption still takes an act of Congress, so Colorado's and California's laws stand. This is the Wild West. Do not expect help here.
Where accountability is clearly assigned, it is assigned to the human. California's SB 1120 bars health plans from letting AI deny, delay, or modify care on its own and requires a licensed clinician to make the final call. Maryland, Texas, Nebraska, Arizona, Alabama, and Georgia followed. ABA Formal Opinion 512 says the same thing to lawyers: know your tool's limits, and do not hand it your judgment.

Read that again. The professions that look protected are protected only in the sense that the law names a human who has to make the call. Neither medicine's rule nor law's shields that human from consequence. Both hand it to them. Everything else is unsettled: 1,134 AI bills introduced in 2025, 131 enacted, 40 states passing at least one.

The optimistic read is that the buck still stops with a person.

The other read: who do you trust, and how much?

AI Accountability for Leaders Who Own the Number

If you carry NRR, gross retention, growth, or clinical quality metrics, this is not a philosophy problem. It is an accountability, liability, and churn problem with a name on it, and the name is probably yours.

Three exposures worth auditing this quarter:

Document the checkpoint. Every AI-assisted decision that touches a customer or a patient needs a named human who signed off, and a record that says so.
Publish the error rate. Your team cannot exercise judgment over a system whose failure modes nobody has shown them. Neither can your customers.
Write the disclosure plan before you need it. Decide now what you say, to whom, and how fast. Nobody drafts this well in the hour after it breaks.

None of this is expensive. All of it is invisible until the day you need it, and on that day it is the only thing standing between your team and the crumple zone.

When Is It Worth Taking the Bullet?

We have been measuring this wrong. When we assess AI accuracy, or a human's call, we talk about quality. What we are actually buying is trust.

Do I trust that the AI is accurate? Do I trust the person making this call?

With my skin on the line, will they come through for me?

Nobody has time to verify who deserves that trust, so we go on track record and whether the results held up last time.

We trusted it. It failed. Now we are standing in front of a judge deciding whether to blame ourselves or the machine.

Before you decide to take the fall to save the day, here are three principles worth knowing.

Integrity is the currency, and owning it buys more than deflecting. Researchers ran 35,676 quarterly earnings calls: analysts forecast higher for the CEOs who blamed internal factors for bad results instead of external ones.
The blame lands somewhere, and it runs downhill. The crumple zone forms around whoever had the least control, not the most. If you do not claim the failure, it settles on whoever touched the tool last, and that person did not choose it, configure it, or decide it was ready.
None of it works if it does not read as real. Read as performative, the identical act gets called hypocrisy and people pull back.

Saying “my mistake” signals you are willing to own it. Whether it helps depends on two things you cannot fix in the moment: whether your team already reads you as competent, and whether they believe you meant it. Have both, win trust. Miss either, you won nothing.

So when is it worth it? When the control that failed was yours. That is the whole test.

And when it was not yours, the answer still is not to let it roll downhill onto someone who never had a vote.

Code Duello

Alexander Hamilton died following a framework built to defend a man's honor while giving both parties structured off-ramps before anyone fired. He skipped the off-ramps. That is the failure mode: treating the hit as the only honorable move when the process gave him three ways out.

The alternative is a modified After-Action Review, something I learned in the Army and have run ever since whenever a person or a program went sideways. Call it the Misfire Review. When you are standing in front of the judge, whoever your judge happens to be:

Own it. Name the control that failed and claim that control as yours. Owning the internal failure protects how people read your integrity. Pointing outward reads as an excuse.
State the facts, not the defense. Describe what the tool did, with an error rate attached. “The model misses this 4% of the time” lands as competence. “I wasn't sure” does not.
Show your own judgment. Document what you concluded before the tool was in the room, with a timestamp. In malpractice research, clinicians who followed standard-of-care AI advice were judged less harshly than those who overrode it.
Focus forward. State the fix, not the regret. Showing a correction beats promising one, every time.
Do not surprise people. Never run steps 2 through 4 for the first time after a failure. The research is blunt: disclosing AI use costs you trust, and hedges like “a human reviewed it” do not buy it back. That cost shrinks when people already knew AI was in the workflow. Normalize it early or pay for it late.

Own What You Own

I am still boggled by the lawyer's choice.

They would rather proclaim their own incompetence than disclose over-reliance on an algorithm's output.

We are all living in this world now. Automated tools and programs are interchangeable and repairable. Human integrity, authenticity, and competence are not.

As long as the lights stay on, the machines will outlive us all.

I have no intent of spending the time I have left protecting an algorithm's reputation with my own.

I hope the lawyer who took the bullet here has more time than I do to recover.

In the end, it's just not worth it for a tool I can swap out in a week.

What have you built to keep customer and team trust high when AI is making decisions? I'd like to hear what you're seeing.

Work With Me

I spent three tours as a VP of Customer Success before going fractional. Now I help SaaS and healthcare leaders build the workflows and decision disciplines that reduce churn, keep AI-assisted decisions defensible, and protect the people who own the number. If you are standing up AI in a customer-facing or clinical workflow and no one has yet defined who owns the call, that is the conversation to have.

Book 15 minutes

Cheers,

Adam Peddicord

Customer Success by Design