AI essay · Mission Delivery Notes
Perfection is a strawman.
AI agents make mistakes. So do the people doing the work today. Judging an agent against perfection doesn't make it safer. It sets the agent up to fail and leaves the process you already have, with its errors, in place. Compare against the alternative.
I can't drive
I am legally blind. I have never been able to get in a car and simply go where I need to go. Every trip depends on someone else's schedule, a transit route, or a ride I have to arrange.
So when I hear that self-driving cars aren't ready because one made a mistake, I hear something specific. I hear that a technology that could hand me back an ordinary freedom must be perfect before anyone is allowed to use it, while the system it would replace kills about a hundred people a day in the United States.
That is the standard we keep applying to AI: not "is it better than what we have," but "is it flawless." Nothing passes that test. No person, process, or vendor ever has.
The nirvana fallacy has a name
Economists call it the nirvana fallacy: comparing a real option against an ideal one, finding the gaps, and rejecting the real one because it isn't perfect. But the real choice was never between it and perfection. It was between it and whatever you are doing now.
Or put simply: perfect is the enemy of good. Waiting for perfect doesn't just delay the good. It keeps the bad you already have. Every month spent waiting for a flawless agent is another month of the errors, backlog and bias in today's process. Waiting feels free. It isn't.
When an agency evaluates an AI agent, the real alternative is almost never a flawless process. It is a team of capable, busy, tired people doing repetitive work under deadline. Those people make mistakes. Measure the agent against them.
Correctness: measure the error rate you already have
Correctness comes first because it is where the strawman does the most damage. Someone shows an agent getting a case wrong and the evaluation ends. Nobody asks how often the current process gets that same case wrong.
Usually nobody knows. That is the first problem.
Daniel Kahneman and his colleagues ran what they called a noise audit at an insurance company. They gave underwriters identical cases and compared the premiums they set. Executives expected the typical difference between two underwriters to be about 10 percent. The median difference was 55 percent. The same policy, priced by two trained professionals at the same company, came back wildly different, and leadership had no idea.
That pattern is not limited to insurance:
Medicine. A 2014 study estimated an outpatient diagnostic error rate of about 5 percent for US adults, roughly 12 million people a year, or at least 1 in 20.
Federal payments. GAO reported $162 billion in estimated improper payments across 68 programs in fiscal year 2024. Not all of that is human error; much of it is missing documentation and process failure. But it is the process as it runs today, and that is the baseline.
Judgment in general. A 2000 meta-analysis of clinical versus mechanical prediction found mechanical methods were on average about 10 percent more accurate than expert human judgment, and substantially better in a third to nearly half of the studies. Human experts came out clearly ahead in only a small minority.
None of this means people are bad at their jobs. It means people are people. Fatigue, interruptions, inconsistent training, and Friday afternoons are part of every human process. We accept that error rate because we have to, and because we rarely measure it.
So measure it. Before you decide whether an agent is good enough, sample the current work and count the mistakes. Then run the agent on the same cases and count again. That is a fair test. Anything else is a demonstration of the agent's imperfection, and you already knew it was imperfect.
Bias: zero is not the baseline
Bias is the second place the strawman shows up. The concern is real: a model trained on human data can reproduce human prejudice, and it should be tested for that. But the question is still comparative. Biased compared with what?
The human baseline is well documented. In a 2003 field experiment, Marianne Bertrand and Sendhil Mullainathan sent fictitious résumés to employers. Résumés with white-sounding names received 50 percent more callbacks than identical ones with Black-sounding names. A 2017 meta-analysis of field experiments found white applicants receiving on average 36 percent more callbacks than Black applicants since 1989, with no change in the level of hiring discrimination over 25 years.
Twenty-five years of awareness, training, and good intentions did not move that number.
An agent's bias has one large advantage over a person's: you can see it. You can run the same case a thousand times with one attribute changed and measure the difference. You can do it again after every change. You cannot do that with a hiring manager, a claims reviewer, or yourself. Human bias mostly shows up in aggregate, years later, if anyone looks.
So the test is the same as for correctness. Measure the bias in the current process. Measure the agent with the same test. Choose the one that does better, and keep measuring.
"But with a human, someone is accountable"
This is the strongest-sounding objection, so it deserves a direct answer.
The argument goes: when a person makes a mistake, there is someone responsible. When an agent makes a mistake, there is no one.
First, an agent should never be the thing that is responsible. A named person owns the result, the same as with any tool. That part doesn't change.
Second, ask what accountability for an employee looks like in practice. Do you fire someone for a single error? Almost no organization does, and none should. A person who makes a mistake gets feedback, maybe retraining, and goes back to work. That is reasonable. It also means the accountability that is supposed to set humans apart rarely changes the outcome.
When performance problems persist, removal is slow. GAO found that dismissing a federal employee for poor performance can take six months to a year, sometimes longer. When the Merit Systems Protection Board asked federal supervisors whether they could remove an employee who was still deficient after a performance improvement plan, only 26 percent said yes. Fifty-one percent said no.
None of that is a criticism of federal employees. Due process exists for good reasons. But it means "a human is accountable" often comes down to a feeling, not a mechanism that corrects errors.
Compare what happens when an agent gets something wrong:
You can switch it off, that minute. No PIP, no hearing, no hurt feelings, no months of documentation.
You can correct it, and the correction sticks. Change the instructions, the context, the tools, or the checks, and every future run gets the fix. A correction to a person fades with time, turnover, and the next busy week.
You can see what it did. An agent can leave a complete record of its inputs, actions, and outputs. Reconstructing why a person made a call six months ago is guesswork.
Federal policy already assumes this. OMB's 2025 AI memo, M-25-21, says that when high-impact AI "is not performing at an appropriate level, agencies must have a plan to discontinue its use." Try writing that plan for a staff position.
Agents aren't more accountable than people. They are more correctable, and correction is what accountability is supposed to produce.
Self-driving cars show what this costs
Self-driving cars are the clearest public example of the perfection standard and of what it costs.
The human baseline is grim. NHTSA estimates that 36,640 people died in US traffic crashes in 2025. That was an improvement, and it is still about a hundred deaths a day. No one suspends human driving after a fatal crash.
Now look at what happens when an automated vehicle fails. In October 2023, a human driver in San Francisco hit a pedestrian and threw her into the path of a Cruise robotaxi, which then dragged her more than 20 feet. California suspended Cruise's driverless permits, effective immediately. The regulator's grounds included that Cruise had misrepresented safety information, which matters and deserved a hard response. But the scale of the reaction to one incident, compared with the daily human toll, shows the asymmetry.
Meanwhile the vehicles that kept operating built a record. Waymo reports that through June 2026 its cars drove 271 million miles with no one in the driver's seat, with 95 percent fewer serious-injury-or-worse crashes and 93 percent fewer pedestrian-injury crashes than human drivers on the same roads. Those figures come from Waymo, whose researchers have published the method in peer-reviewed work. A separate analysis with the insurer Swiss Re found 88 percent fewer property-damage claims and 92 percent fewer bodily-injury claims than human drivers. Independent replication should continue. But the comparison is the right one: against people driving those same streets, not against zero.
In 2017, RAND researchers Nidhi Kalra and David Groves asked how safe automated vehicles should be before deployment. Their model found that allowing them once they are 10 percent safer than the average human driver saves more lives than waiting until they are 75 or 90 percent safer, in every condition they studied. Over the long term, the difference is often more than half a million lives. Waiting for near-perfection is not caution. It costs lives.
That is the trade. Getting the benefit means accepting some risk. Refusing every risk doesn't eliminate it. It keeps the risk you already have, which you've stopped noticing because it's familiar.
Why we keep doing this
Researchers have a name for it. In 2015, Berkeley Dietvorst, Joseph Simmons, and Cade Massey published a set of experiments on what they called algorithm aversion. People who watched an algorithm forecast, including its mistakes, became less likely to rely on it than on a human forecaster, "even among those who saw the algorithm outperform the human." People "more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake."
Same mistake, different verdict. We forgive people for being human. We take a machine's error as proof that it doesn't work.
The same researchers found a way out. In a 2018 follow-up, people were far more willing to use an imperfect algorithm if they could adjust its output, even slightly. Control made them "more tolerant of errors." That points to how to introduce agents responsibly.
This isn't permission to run wild
None of this means giving an agent the keys on day one. You wouldn't do that with a new employee.
When someone joins a team, you check their work closely. You give them bounded tasks. You review before anything ships. As they prove themselves, you check less and give them more. Trust is earned with evidence, and it can be reduced again when the evidence changes.
Agents should earn trust the same way:
Start narrow. One defined task with a clear right answer.
Review everything at first. Every output gets human eyes, and the people reviewing it can adjust and correct it.
Measure both sides. Track the agent's error rate on the same cases, by the same rules, as the human baseline.
Reduce review as evidence builds. Move from reviewing everything to sampling, and from sampling to exceptions, as the numbers justify each step.
Keep the off switch and a named owner. A person stays responsible for the result, and the agent can be stopped the moment the numbers turn.
That is how we work at Intelligrit. Agents draft and act within defined limits. Named people review the evidence and own the result. What the evidence shows decides how much the agents do, not how confident anyone feels.
What this means for government teams
Measure your current error rate before you evaluate AI. If you don't know how often the current process is wrong, you can't judge whether the agent is better.
Use the same test for both. Same cases, same scoring, same reviewers.
Measure bias against the current process, not against zero. Then keep measuring, because an agent can be retested every time it changes.
Write the off-switch plan up front. M-25-21 requires one for high-impact uses. It's also just good engineering.
Count the cost of waiting. Every month the old process runs, it makes its usual mistakes. Rejecting an agent that would have made fewer isn't the safe choice. It just keeps those mistakes in place.
Back to my screen reader
I live in a world of screen readers, and most of the web isn't built for me. WebAIM's 2026 analysis of the top million home pages found detectable accessibility failures on 95.9 percent of them, with an average of 56 errors per page. Those are only the failures an automated tool can detect.
AI has made that world far more manageable. It reads the page the developer never labeled, describes the image with no alt text, and gets me through the form that was never tested with a keyboard. It is not perfect. It sometimes misdescribes things, and I have learned where to double-check it.
But I don't compare it to a perfectly accessible web. That web doesn't exist. I compare it to the unusable site in front of me, and on that comparison it wins easily, every day.
That is the whole argument. Your agent will make mistakes. Your people make mistakes now. Measure both honestly, build trust with evidence, keep the off switch close, and stop asking a new tool to be perfect when the old process never had to be.
Sources
Harold Demsetz, "Information and Efficiency: Another Viewpoint," Journal of Law & Economics 12, no. 1 (1969): 1–22, doi:10.1086/466657. Summarized by Timothy Taylor, "Harold Demsetz: Dissecting the Nirvana Viewpoint" (2021).
Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein, Noise: A Flaw in Human Judgment (2021); insurance audit figures as reported in "How noisy is your company?", strategy+business (May 2021). See also Kahneman et al., "Noise", Harvard Business Review (October 2016).
Hardeep Singh, Ashley N. D. Meyer, and Eric J. Thomas, "The frequency of diagnostic errors in outpatient care", BMJ Quality & Safety 23 (2014).
GAO, "Improper Payments: Information on Agencies' Fiscal Year 2024 Estimates", GAO-25-107753 (March 2025).
William M. Grove and others, "Clinical versus Mechanical Prediction: A Meta-Analysis", Psychological Assessment 12, no. 1 (2000).
Marianne Bertrand and Sendhil Mullainathan, "Are Emily and Greg More Employable than Lakisha and Jamal?", NBER Working Paper 9873 (2003); American Economic Review (2004).
Lincoln Quillian and others, "Meta-analysis of field experiments shows no change in racial discrimination in hiring over time", PNAS 114, no. 41 (2017).
GAO, "Federal Workforce: Improved Supervision and Better Use of Probationary Periods Are Needed to Address Substandard Employee Performance", GAO-15-191 (February 2015).
U.S. Merit Systems Protection Board, "Remedying Unacceptable Employee Performance in the Federal Civil Service", research brief (June 2019).
OMB, "M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust" (April 2025).
NHTSA, "Early Estimate of Motor Vehicle Traffic Fatalities and Fatality Rate in 2025", DOT HS 813 800 (April 2026).
California DMV, "DMV Statement on Cruise LLC Suspension" (October 24, 2023); CBS News Bay Area, "Cruise admits false report".
Waymo, Safety Impact data hub (data through June 2026); Kristofer D. Kusano and others, "Comparison of Waymo Rider-Only Crash Rates by Crash Type to Human Benchmarks at 56.7 Million Miles," Traffic Injury Prevention 26, sup1 (2025), doi:10.1080/15389588.2025.2499887.
Waymo and Swiss Re, "New Swiss Re study: Waymo is making roads safer" (December 2024).
Nidhi Kalra and David G. Groves, The Enemy of Good: Estimating the Cost of Waiting for Nearly Perfect Automated Vehicles, RAND RR-2150 (2017).
Berkeley J. Dietvorst, Joseph P. Simmons, and Cade Massey, "Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err", Journal of Experimental Psychology: General 144, no. 1 (2015).
Berkeley J. Dietvorst, Joseph P. Simmons, and Cade Massey, "Overcoming Algorithm Aversion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify Them", Management Science 64, no. 3 (2018).
WebAIM, "The WebAIM Million" (2026).