Do AI hiring laws cover promotions, or only external hiring?

Promotion is squarely covered. New York City's Local Law 144 applies to automated employment decision tools used to screen candidates for employment or employees for promotion, and it says so explicitly. Title VII has always treated promotion as an employment action, and state algorithmic discrimination statutes generally define consequential employment decisions to include advancement, compensation, and termination. The idea that these rules are about the applicant funnel is a widespread and expensive misreading.

Why do promotion tools escape bias audits so often?

Because of how they were purchased. Applicant tracking systems arrive through a procurement path that now expects an AI compliance review, while promotion scoring usually arrives as a feature inside an HRIS or talent-management suite the company already owned. Nobody signed a contract for an employment decision tool, so nobody routed it to legal. The capability changes in a release note rather than in a purchase order.

What data makes promotion models especially prone to bias?

Promotion models train on internal history, and internal history encodes whatever the organization has done before. Performance ratings carry documented rater bias, tenure and continuous-service features penalize anyone who took parental or medical leave, high-visibility project assignment reflects manager discretion, and prior promotion velocity simply reproduces the existing shape of the leadership pipeline. The model does not need a protected characteristic in its features to reconstruct one from these proxies.

How do you measure adverse impact in promotions?

The same selection-rate logic used for hiring applies: compare the promotion rate for each group against the rate for the most-selected group, and treat a ratio below roughly four-fifths as a signal that warrants investigation. The practical difficulty is defining the pool. Because eligibility for a promotion is itself often determined by an algorithm or by manager nomination, an analysis that starts from the slate can look clean while the exclusion happened one step earlier.

Does a human approver at the end remove the risk?

Only if the human meaningfully changes outcomes. A committee that approves the model's ranked slate nearly every time is documenting the model's decision, not supervising it. Regulators and plaintiffs both look at override rates, and a rubber-stamp rate near a hundred percent tends to establish that the tool substantially assisted or replaced discretionary decision making — which is the trigger the statutes use.

AI in Promotion and Succession Planning: Discrimination Risk in 2026

The Blind Spot Is Structural, Not Careless

When AI hiring laws started landing, compliance attention followed the word hiring. Teams inventoried resume parsers, video interview scorers, and assessment vendors, because those are the systems that sit in the recruiting stack and have contracts attached. Promotion algorithms live somewhere else entirely — inside the performance module, the internal mobility marketplace, or the succession dashboard that an HR business partner opens twice a year.

That difference in purchasing path, not any difference in legal exposure, is why these tools go unaudited. And the exposure is arguably higher: promotion cases involve identified employees with tenure, documented performance, and colleagues who can testify about the pattern, rather than anonymous rejected applicants who never learn why they were screened out.

The Systems That Count as Employment Decision Tools

Promotion readiness scoring

A model that rates employees as ready-now, ready-in-one-year, or not-ready is screening employees for promotion in the most literal sense the statutes describe. If the score gates who is even considered, it is doing the selecting.

High-potential identification

HiPo designation controls access to sponsorship, stretch assignments, and executive visibility — the inputs to every later advancement decision. An algorithm that assigns the label shapes years of downstream outcomes from a single classification.

Internal mobility matching

A marketplace that surfaces open roles to some employees and not others is performing candidate screening on the internal population. Employees never see the roles they were not shown, which makes the exclusion invisible to the person affected by it.

Succession slate generation

AI-assembled candidate slates for senior roles concentrate the highest-value decisions in the least-audited tool. These slates are also frequently generated from tenure and prior-promotion features that encode historical pipeline composition directly.

Skills inference from work artifacts

Inferring capability from tickets, commits, or CRM activity systematically rewards whoever was assigned the visible work. Assignment discretion belongs to managers, so the model inherits their patterns while presenting the result as objective measurement.

Flight-risk and retention scoring

Retention models influence who receives counteroffers, equity refreshes, and development investment. When compensation follows the score, the model is making a consequential employment decision even though nobody calls it a promotion tool.

Why the Training Data Is the Whole Problem

An external hiring model at least learns from a population the company did not construct. A promotion model learns exclusively from the organization's own history — which employees advanced, how fast, and under which managers. If the leadership bench skews in a particular direction today, a model fitted to reproduce past promotions will reliably identify candidates who resemble the people already there, and will do so with a confidence score attached.

The specific features that cause trouble are mundane. Continuous tenure penalizes parental and medical leave. Performance ratings carry well-documented rater effects. Assignment to high-visibility projects is manager discretion recorded as employee merit. Promotion velocity is the outcome variable smuggled back in as a predictor. None of these are protected characteristics, and together they approximate several of them well enough to produce a disparity a court can measure.

Audit Checklist for Internal Advancement AI

1. Inventory What You Missed

☐Walk the HRIS, performance, and talent modules feature by feature, not vendor by vendor
☐Include capabilities that arrived in a product update rather than a signed contract
☐Flag any score that gates eligibility, visibility, or slate membership

2. Measure at the Right Stage

☐Compute selection rates at eligibility, at slate inclusion, and at final promotion separately
☐Define the pool as everyone who could have been considered, not everyone the tool surfaced
☐Compare against the four-fifths benchmark and investigate anything below it

3. Interrogate the Features

☐Identify proxies: continuous tenure, rating history, project visibility, prior promotion velocity
☐Check whether leave, part-time status, or schedule flexibility depress scores mechanically
☐Confirm the vendor will disclose the feature set — if they will not, that is your finding

4. Prove Human Oversight Is Real

☐Track how often reviewers change the model's ranking or add candidates it excluded
☐Give reviewers the underlying evidence, not just the score, before they decide
☐Notify affected employees where required, and keep the notice records with the audit

The Accessibility Layer Inside Internal HR Tools

Internal talent portals are held to the same accessibility obligations as any other employment system, and they are usually in worse shape than the public careers site because no one outside the company ever sees them. If the internal mobility marketplace renders job cards in a component keyboard navigation cannot reach, or the self-nomination form has unlabeled fields a screen reader cannot announce, then employees with disabilities are structurally excluded from applying — before any model scores anyone. That is a failure to provide equal access to advancement opportunities, and it compounds whatever the algorithm does next.

Your internal portals are probably worse than your public site

Employee-facing HR tools rarely get an accessibility review, and a self-nomination form a screen reader user cannot complete is an advancement barrier before any model runs. RatedWithAI scans your pages and shows exactly which controls and forms fail.

Run a Free Scan →

Frequently Asked Questions

Our vendor says the promotion module is not an AEDT. Is that reliable?

Treat it as a starting point, not a conclusion. Vendors assess their product against a definition without knowing how your organization uses it, and the same feature can be advisory at one company and decisive at another. If the score determines who reaches the slate at your company, your usage is what a regulator will examine.

We only use the model to shortlist. Doesn't the committee make the decision?

Shortlisting is the decision that matters most, because everyone excluded at that step never receives consideration. The statutes reach tools that substantially assist decision making precisely to capture this pattern. A committee choosing among five names the model selected from four hundred is exercising discretion inside boundaries the model drew.

Do we have to notify employees that an algorithm is involved?

In several jurisdictions yes, and the promotion context often carries a longer notice window than external hiring because the employee is already there to be told. Some regimes also require disclosing the job qualifications and characteristics the tool considers, and a few give employees a right to request an alternative process.

What if the audit finds a disparity?

Finding it is the point, and a documented remediation record is far better than an absence of analysis. Practical responses include removing or reweighting proxy features, adjusting the eligibility gate, expanding the pool, and requiring reviewers to consider candidates the model ranked lower. What you should not do is stop measuring — an audit that gets discontinued after a bad result is worse evidence than one that was never started.

Does this apply to compensation decisions too?

Increasingly yes. Statutes that define consequential employment decisions typically list compensation alongside hiring, promotion, and termination, so algorithmic merit-increase allocation and equity refresh targeting sit in the same category. Compensation models also produce disparities that are unusually easy to quantify after the fact.