One Decision, Replicated a Hundred Times

What the hiring-AI reckoning reveals about every scoring system, including the one running your contact centre.

William Rochelle asked the question that started this, plainly, in a recent piece: what if the job you applied for never existed? It is a good question, and behind it sits a mechanism worth following, first through hiring, where it is finally getting attention, and then onto a contact centre floor, where I have spent twenty-six years watching the identical failure run unnoticed.

Three stories broke in the same window, and they are really one story.

A Texas investigation is asking whether people paid for a better shot at jobs that were never real. A billion-plus rejections, disclosed by a hiring software vendor in its own court filing, are now the subject of a certified nationwide age discrimination action. And the largest independent study of algorithmic hiring ever run, out of Stanford, Chapman, and Northeastern, examined more than four million applications from three million people across 156 large employers, all scored by the same vendor, and found clear racial disparities in who advanced.

The outrage is warranted. It is also not the lesson.

The lesson is that a scoring layer now sits between a human being and a paycheck, almost no one can see it, and the specific way it fails is not a hiring problem. It is an architecture problem. In the contact centre the layer sits between an agent and a customer instead, but it is the same layer failing the same way. Same failure. Different room.

Three properties worth understanding

The interesting part of the hiring story is mechanical and it has three parts.

First, the safety net is gone. Applying to a hundred companies used to mean a hundred roughly independent decisions. Someone has a bad day, someone else likes your background and the variance averages out. Apply widely enough and a qualified person eventually clears somebody's bar. That is the whole implicit logic of volume applying. A single scoring vendor collapses it. Your rejections stop being independent draws and become correlated, because one function computes your rank and computes essentially the same rank everywhere it is deployed. A no from one becomes a no from all. There are no longer a hundred decisions.

There is one decision, replicated a hundred times.

Second, the score is a variable with no feedback loop. You never see it, so you cannot calibrate against it. There is no reason attached, no "no, but," no signal you could act on. And because retention rules keep the record, a low score is not a one-time event. It is a durable state that follows you. Functioning markets self correct through signals: a price, a reason, a channel to contest. Here the signal is suppressed by design.

Third, the harm is invisible in aggregate and visible only at the unit. This is the part that should stop people cold. The Stanford team did not find the disparity by looking at company totals. It vanishes at that altitude. They found it by analysing each of the roughly 1,700 positions separately, which is exactly how discrimination law is written to be applied, through the four-fifths rule. On the totals, every employer looks compliant. Disaggregate to the individual role and nearly forty thousand applications from Black applicants turn out to have gone to positions where the algorithm produced outcomes that meet the federal definition of adverse impact.

Read that structure again. The harm is real at the unit level and disappears the moment you roll it up. That is not, in the first instance, a fairness failure.

It is an attribution failure wearing a fairness failure's clothes.

The same layer, in your operation

Now move the layer from hiring into operations, because the structure travels intact.

Automated quality scoring rates an agent's calls before any human listens and the agent never sees the score that shaped their ranking, their coaching queue, their next roster. The team dashboard looks healthy on the aggregate. Handle time is fine. CSAT is fine. The composite is green. Meanwhile the weak signal that would explain a specific failure, the one call type, the one policy step, the one cohort of interactions, is buried under the average, exactly where the hiring disparity was buried before someone thought to disaggregate.

And when the good people leave, the story writes itself the way it always does.

The problem was them.

They were not resilient enough.

Not engaged enough.

Not a good fit.

Attrition gets booked as a hundred individual failures. It is almost never a hundred individual failures.

It is one orchestration failure,

replicated across a hundred people, and the scoring layer is what made it look like a hundred separate events instead of one systemic one.

That misattribution is why I wrote 'The Teams That Stay', as those who depart typically indicate a flaw in the system, not in themselves.

This is the same move the hiring vendors made. Nobody set out to build a discriminatory system. They deployed a scoring layer without observability, without contestability and without attribution granularity, and the architecture produced the outcome. The AI performed exactly to the level of the system it was dropped into. That is the whole thesis in one line, and it is not unique to AI. AI performs to the level of the architecture it is deployed into, the same as any platform ever has.

You cannot price an outcome you cannot attribute

Here is where it becomes a decision, not a grievance.

Every number a leader puts on one of these outcomes is a price. The cost of attrition. The cost of a bad hire. The return on the AI you just bought. And a price is only as real as the attribution underneath it. If the scoring layer has no observability, no contestability, and no way to trace an outcome back to the specific interaction and the specific cause, then the number you attach to it is not a measurement. It is a black box's output, laundered into a business case.

You cannot price an outcome you cannot attribute. You can only estimate it, defend it, and be surprised by it later, which is how a system quietly rejects a billion applications, or churns a floor of capable agents, while every dashboard above it stays green.

The fix is not "less automation." The volumes are real, some scoring layer is going to exist. The fix is the architecture the layer sits inside. High-reliability organisations earn their reliability by doing the one thing these systems are built to prevent: they surface weak signals and treat them as information, and they hold a preoccupation with failure at the unit level rather than the aggregate. That is the standard. Build the attribution first, at the grain where the harm actually lives, and only then, deploy for scale. Do it in the other order and you have built 'The 85% Trap' on purpose. I named a book after that trap because it is that common: capable technology, dropped into an architecture that cannot attribute what it does, underperforming exactly as designed.

The Australian position is different in form, not in substance

A note before I go further: I am not a lawyer and none of this is legal advice. I am setting out what I currently understand of the legislation already in place and the reforms coming, and any of it should be checked with counsel before you act on it.

Everything above is a United States story. Australia has no equivalent case and no AI-specific statute. The National AI Plan confirmed in late 2025 that the country would rely on its existing laws and sector regulators rather than a standalone AI Act. That absence reads like a gap. It is not one.

AI-assisted hiring decisions already sit under Australian anti discrimination and employment law: the Fair Work Act, the Age Discrimination Act, the Disability Discrimination Act, the Racial Discrimination Act, and their state equivalents. The same protected attributes that drive the American cases, age, race and disability, are covered here. A scoring layer that produces those outcomes is exposed, whether or not anyone has litigated it yet.

Transparency is arriving on a fixed date. From December 2026, the automated decision making reforms in the Privacy and Other Legislation Amendment Act 2024 require organisations to disclose where personal information feeds substantially automated decisions that carry legal or similarly significant effect, and to give people meaningful information about how those decisions are made. That is an observability obligation, written into law, landing in months. A scoring layer no one inside the business can explain stops being only an operational risk and becomes a compliance one.

Then the sharper point, and it is where the Australian frame cuts deeper than the American one. Australian anti discrimination law puts the burden on the person bringing the complaint to establish that discrimination occurred, and it tests for indirect discrimination: a condition or requirement that disadvantages a protected group. To run that argument, an applicant needs position level visibility into how the layer scored people like them. That is precisely the visibility the black box is built to withhold. In the United States, regulators forced that disaggregation with subpoena power. No such lever exists for an Australian applicant, or for the Australian operations leader watching the same layer from the inside. Both have only the attribution the architecture was built to produce.

So the law exists, and the evidence is locked inside the scoring layer. Build the attribution in, and the same instrumentation answers the coming transparency obligation and stands up a defence to a discrimination claim. Skip it, and you can do neither.

The part that should land

Regulators found the hiring version by disaggregating. That is the entire method. They stopped looking at the totals and looked at the roles, and the bias that had been hiding in plain sight stepped forward.

Leaders can find the operational version the same way and they do not need a subpoena to do it.

The hiring story is getting attention because the harm lands on a paycheck. The contact centre version lands on a customer and on an agent, and it is running right now, under the same kind of layer, with the same green dashboards on top.

The question is not whether you have a scoring layer. You do. The question is whether you can attribute what it decides. If you cannot, you are not managing an outcome.

You are wishing for one.

Built on Rigor. Engineered for Scale.

Sources for the figures referenced above: Texas Office of the Attorney General (LinkedIn CID, July 2026; $17.8B FY2025 revenue); Mobley v. Workday, N.D. Cal. (ADEA collective certification, 16 May 2025; ~1.1 billion applications disclosed in filings); Bommasani, Bana, Creel et al., "Algorithmic Monocultures in Hiring" (Stanford HAI / Chapman / Northeastern, 2026). Workday reports are used by more than half the Fortune 500. Australian position: Privacy and Other Legislation Amendment Act 2024 (Cth), automated decision-making transparency provisions commencing December 2026; Fair Work Act 2009, Age Discrimination Act 2004, Disability Discrimination Act 1992, Racial Discrimination Act 1975 and state anti-discrimination Acts; National AI Plan (December 2025); Australian Human Rights Commission guidance on AI and recruitment.