To the people who were told that the system could not make an exception,
I was the person who made exceptions difficult. I did not stand behind a counter or sign the letters that arrived at your homes. I worked in a room without windows, designing the sequence of screens, thresholds, and approvals through which other people had to pass. My name was never printed on a refusal. That absence was part of what made my work so effective.
I was employed by the fictional city of Merrow as a control optimizer. The title sounded technical enough to avoid argument. My assignment was to reduce inconsistency in public services: fewer missing documents, fewer mistaken payments, fewer unauthorized decisions, fewer opportunities for someone to exploit a rule. There were legitimate reasons for the work. A poorly controlled service can lose money, distribute benefits unfairly, or leave vulnerable people exposed to the preferences of whoever happens to be on duty.
I believed that better control would mean better justice. For several years I was rewarded for proving it.
The first system I redesigned handled temporary housing assistance. Before my arrival, applications moved between offices in paper folders. People lost documents. Some decisions took weeks, and similar applicants sometimes received different answers. A resident with a persuasive voice could receive attention that a quiet resident did not. I saw the problem clearly. I thought I understood the solution.
We replaced the folders with a single digital pathway. Each application had to pass through identity verification, address confirmation, income assessment, and a final eligibility check. No stage could be skipped without a recorded authorization. Every action produced a timestamp. Every officer had a measurable workload. The dashboard showed exactly where each case stood.
The first results were impressive. The average processing time fell. Duplicate applications were detected earlier. Supervisors could find missing cases instead of searching cabinets. We had removed some genuine injustices. I still believe those improvements mattered.
Then I began to optimize the delays that remained.
A supervisor told me that too many applications sat in a category marked awaiting clarification. I examined the data and found that many applicants did not answer a request for additional information within seven days. I shortened the request, made its language more direct, and added an automatic reminder. Response rates improved. The remaining cases appeared to be the applicants least willing to cooperate.
I did not meet them. My data did not require me to.
One of those applicants was a woman named Sera, who worked irregular shifts at two kitchens and slept on her sister’s sofa with her young son. The system requested a utility bill in her name to confirm her address. She had no such bill. The reminder arrived while she was working. By the time she found someone who could explain the alternative evidence rule, the deadline had passed.
Her case closed automatically. On my dashboard it became a successful resolution of an inactive application.
I learned of Sera only because an appeals officer sent me a note. The officer had been trying to reopen cases that were closed for missing evidence rather than ineligibility. Our system allowed a new application but not a simple restoration of the old one. The difference mattered because assistance was calculated from the date of a valid submission. Sera lost weeks she could not afford to lose.
I reviewed the workflow and discovered that it was functioning exactly as designed. That was the sentence I had once considered the strongest possible defence of my work.
An optimized system can perform flawlessly while producing a result nobody would defend if required to look at it closely.
I asked my team to investigate. We found that the utility-bill request was not essential to the law. It was one convenient method of checking an address, chosen because it was easy to standardize. Other evidence was permitted, but the alternative route appeared only after an applicant selected a small link labelled exceptional circumstances. People who needed it most often had the least time, confidence, or access to discover it.
We had not prohibited Sera from proving where she lived. We had arranged the process so that the easiest path belonged to someone whose life already fit our assumptions.
I proposed moving the alternative evidence options onto the main screen. A colleague warned that this could increase false claims. He was not inventing a concern. Fraud prevention is a real public obligation, and weak checks can harm people who depend on a service. Yet we had no evidence that showing lawful alternatives would produce more fraud. We were defending a restriction because its absence frightened us.
That was the first time I noticed how often my profession treated a possible misuse as more important than a documented exclusion.
I had learned to describe every control as a barrier against risk. I had not learned to ask who encountered the barrier, how often, and at what cost. The language of risk made the problem sound complete even when half its consequences remained invisible.
My next assignment was to optimize access to the city’s shared workshop. Residents could use its tools to repair furniture, learn trades, and produce small goods. The workshop had suffered thefts, damaged equipment, and arguments over bookings. We introduced identity cards, training records, timed reservations, and penalties for missed sessions. The tools were safer. The disputes diminished. The booking calendar looked beautiful.
Attendance also declined.
The director thought the workshops had become less popular. I reviewed the figures and discovered that experienced members were booking more hours while newcomers booked fewer. A newcomer had to complete an online orientation, wait for an identity check, and reserve a slot several days in advance. Experienced members already knew the system and could reserve the most desirable times within minutes of their release.
Our controls had made the workshop easier to manage and harder to enter. The people who understood its rules benefited most from the efficiency those rules created.
A teenage apprentice called Tomas told a volunteer that he had stopped trying because he could never find a slot after school. The volunteer suggested he should plan ahead. Tomas replied that he did not know his shifts far enough in advance to plan. He was not failing to use the system correctly. The system had been built around the schedule of somebody else.
I added an allocation for first-time users and a small number of walk-in appointments. We kept the safety training and the identity checks, but moved some of the burden from the applicant to the institution. It required more staff time. The dashboard showed a rise in administrative minutes per visit. It did not show that Tomas returned, learned to use a lathe, and eventually helped repair the workshop’s broken shelving.
I began to distrust any metric that could register the cost of a human being while missing the value of their participation.
The trouble with control is not that it always fails. The trouble is that it often succeeds at the narrow thing it was built to do. A locked door can eliminate unauthorized entry. It can also prevent a person from reaching the service the building exists to provide. Whether the door is good depends on more than the number of times it remains shut.
At Merrow, we called our design philosophy friction reduction. Yet we were selective about whose friction counted. We measured the time an employee spent reviewing a case but not the time an applicant spent finding evidence. We counted the number of calls avoided by a new form but not the hours a resident spent searching for a telephone number that no longer appeared on the website. We celebrated fewer exceptions without asking whether people had stopped requesting them.
One quarter, the mayor’s office congratulated us on a record decline in complaints. I knew by then that we had also introduced a new complaint portal requiring users to create an account. A person who could not complete the account setup could not submit the complaint. We had made dissatisfaction harder to observe and then praised ourselves for seeing less of it.
An older engineer in our department used to say, ‘Every measure is a window, not the whole room.’ I repeated the phrase in presentations because it sounded wise. I did not understand it until the window showed me a clean dashboard and a queue of people waiting outside the office.
That queue formed after a winter storm damaged several apartment blocks. Residents needed emergency repairs, temporary accommodation, and confirmation that their buildings were safe. Our system required proof of occupancy before issuing certain forms of assistance. Many residents had left their documents behind in damaged rooms. Some had no electricity to charge a phone. The very event that created their need also made the normal verification process difficult.
Our emergency procedure existed, but it was buried in an administrator’s manual and required two approvals. Officers were reluctant to use it because exceptional cases were reviewed closely in performance meetings. An ordinary denial was easy to explain. An exception required the officer to defend their judgment.
We had created a system in which following the default was professionally safer than responding to the emergency.
A housing officer named Leena called me while residents waited at her desk. She asked whether I could unlock the emergency pathway. I told her that I needed authorization from the service director. She said the director was helping evacuate another building. I suggested waiting for the formal approval. There was silence on the line. Then she asked, ‘Who is this process protecting right now?’
I could list the risks it was designed to control. I could not tell her who was being protected by the delay.
I called the director, documented the emergency, and obtained authorization to release the pathway. The residents received temporary help. I was praised for resolving the incident. The praise disturbed me because Leena had done the important work. She had refused to confuse my process with its purpose.
After the storm, I proposed a different way to evaluate controls. We would continue measuring error, fraud indicators, processing time, and unauthorized access. But we would also measure preventable abandonment, repeat submissions, appeal reversals, applicant effort, and the consequences of delay. Every high-impact rule would have a named owner responsible for reviewing whom it excluded and why.
My proposal did not please everyone. Some managers worried that the new measures would make performance appear worse. They were right. A more complete account of our work would reveal problems that the old dashboard concealed. Others feared that giving staff more discretion would invite favouritism. That fear also deserved attention.
I did not propose replacing rules with personal sympathy. Sympathy can be inconsistent, and discretion without accountability can become another form of power. I proposed documented, reviewable discretion: clear eligibility principles, multiple acceptable forms of evidence, emergency routes that could be activated quickly, and audits that examined both improper approvals and improper refusals.
We also began testing procedures with the people who had to use them. We invited residents with limited digital access, irregular working hours, disabilities, and unfamiliarity with administrative language. We paid them for their time. Their criticism was not always gentle. One participant asked why we had placed the help button after the step at which people most needed help. Another asked whether the application had been tested on a phone with a cracked screen and a weak connection.
We had tested the software. We had not adequately tested the experience of depending on it.
I started reading the published letter An Open Letter on Algorithmic Authority differently. A decision need not be made by a mysterious artificial intelligence to acquire the appearance of machine-made inevitability. A sequence of ordinary rules can also conceal the people who chose its thresholds. And An Open Letter on Efficiency That Excludes reminded me that a faster process is not necessarily a fairer one if its speed depends on pushing difficulty onto those outside the organization.
These insights did not turn me against optimization. I still want systems that prevent theft, protect confidential records, detect mistakes, and distribute scarce resources fairly. I want staff to know which decisions they may make and applicants to understand what evidence is required. Good controls make public promises more reliable.
But reliability must include the ability to recognize when the normal route is wrong for the circumstances in front of us.
Control should be judged by what it protects, not by how completely it prevents deviation. A rule against favouritism is valuable. A rule that hides lawful alternatives from applicants is not made fair by applying it consistently. An audit trail is valuable. An audit that counts every exception as a threat may train staff to ignore urgent human needs.
I now ask five questions whenever someone requests another control. What harm is it intended to prevent? What evidence shows that the harm is likely? Who will bear the work of complying? How will we recognize an unjust refusal? And what route remains open when the ordinary procedure cannot accommodate an ordinary human life?
These questions do not always lead to fewer rules. Sometimes they lead to stronger safeguards, better explanations, or more careful review. They always make it harder to pretend that a control has no cost.
I think about Sera when I see a completed case count. I think about Tomas when a booking calendar shows full capacity. I think about Leena when a manager praises the absence of exceptions. Those people are not anecdotes that disprove every metric. They are reminders that metrics need to be interpreted within the lives they describe.
If you build systems, do not wait for a crisis to discover their emergency doors. If you supervise the people who operate them, do not punish every justified deviation and then demand compassion. If you are offered a perfect dashboard, ask what happened to the people who could not get far enough into the process to become data.
Above all, do not let the elegance of a system become an excuse for refusing to see its consequences. The cleaner the diagram, the more deliberately you must search for the person who does not fit inside it.
I once believed my finest achievement would be a process in which nobody could make an unauthorized choice. I would now prefer one in which every important choice can be explained, challenged, corrected, and traced to someone willing to take responsibility.
That is not a rejection of control. It is a refusal to mistake control for care.
With an apology for the doors I made too difficult to open,
A former control optimizer of Merrow, still learning what a well-designed system owes the people who must live inside it



