Essay
Algorithm as Agent: AI Vendor Liability for Title VII Hiring Discrimination After Loper Bright
Cheng-chi (Kirin) Chang *
Introduction
Derek Mobley, a Black man over forty with a disability, applied for more than one hundred jobs with companies that used Workday, Inc.’s applicant-screening platform. Mobley was rejected from every job he applied to, often within minutes and frequently before any humans at the employer had opened his file. 1 See Third Amended Complaint ¶¶ 29-52, Mobley v. Workday, Inc., No. 23-cv-00770 (N.D. Cal. Mar. 27, 2026), ECF No. 287 [hereinafter TAC]; Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 807 (N.D. Cal. 2024) (order granting in part and denying in part motion to dismiss). In February 2023 he sued not the employers but Workday itself. Mobley based the suit on the theory that Workday was an “agent” of the employers under Title VII, the Age Discrimination in Employment Act (ADEA), and the Americans with Disabilities Act (ADA). 2 TAC ¶¶ 9, 130-34; see 42 U.S.C. § 2000e(b); 29 U.S.C. § 630(b); 42 U.S.C. § 12111(5)(A). In July 2024, the district court partially denied Workday’s motion to dismiss, and in May 2025, Judge Rita Lin certified a nationwide collective on the ADEA claim, aggregating approximately 1.1 billion processed applications. 3 Mobley, 740 F. Supp. 3d at 813-14; Mobley v. Workday, Inc., No. 23-cv-00770, 2025 WL 1424347, at *11 (N.D. Cal. May 16, 2025) (order granting preliminary collective certification). That said, nearly a year before that certification order, the Supreme Court’s decision in Loper Bright Enterprises v. Raimondo had handed Workday its most natural defense.
In Loper Bright Enterprises v. Raimondo, the Court overruled Chevron U.S.A. Inc. v. Natural Resources Defense Council, Inc., ending forty years of judicial deference to agency interpretations of ambiguous statutes. 4 Loper Bright Enters. v. Raimondo, 144 S. Ct. 2244, 2273 (2024); Chevron U.S.A. Inc. v. Natural Res. Def. Council, Inc., 467 U.S. 837 (1984). In Chevron’s place, Loper Bright instructed that courts use ordinary tools of statutory interpretation to ascertain a statute’s “single, best meaning.” 5 Loper Bright, 144 S. Ct. at 2266. Six months later, the Trump Administration revoked the Biden artificial intelligence (AI) executive order. 6 Exec. Order No. 14179, 90 Fed. Reg. 8741 (Jan. 31, 2025) (revoking Exec. Order No. 14110, 88 Fed. Reg. 75191 (Nov. 1, 2023)). Executive Order 14110 directed federal agencies to develop AI safety standards and to address algorithmic discrimination, including in hiring. Then, on January 27, 2025, the U.S. Equal Employment Opportunity Commission (EEOC) withdrew its 2023 AI hiring guidance from its website. 7 See Erinn L. Rigney, Ninamarie C. Moore & Isabella F. Sparhawk, The Changing Landscape of AI: Federal Guidance for Employers Reverses Course with New Administration, K&L Gates (Jan. 31, 2025), https://perma.cc/YZA2-Q6BX; see also EEOC, Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964 (May 18, 2023), https://perma.cc/P6JR-FE57. By April 2025, Executive Order 14281 had directed federal agencies to deprioritize disparate-impact enforcement entirely. 8 Exec. Order No. 14281, 90 Fed. Reg. 17537, 17538 (Apr. 28, 2025). Section 2 declares it “the policy of the United States to eliminate the use of disparate-impact liability in all contexts”; Section 4 directs agencies to deprioritize enforcement of disparate-impact provisions, expressly including Title VII’s, 42 U.S.C. § 2000e-2.
Workday’s defense assumes that holdings resting on the EEOC’s interpretive support cannot survive its withdrawal; Workday has already pressed exactly that argument against Mobley’s parallel ADEA holding, and conventional wisdom extends it to the agency theory. 9 See Defendant Workday, Inc.’s Motion to Dismiss Second Amended Complaint and to Strike at 6-7, Mobley v. Workday, Inc., 2026 WL 1711302 (N.D. Cal. Jan. 21, 2026) (No. 23-cv-00770), ECF No. 237, (arguing that Rabin v. PricewaterhouseCoopers LLP, 236 F. Supp. 3d 1126 (N.D. Cal. 2017), on which the court relied to hold that the ADEA reaches disparate-impact claims by job applicants, rested on Chevron deference to the EEOC’s interpretation, and that its “reliance on Chevron deference is thus no longer valid” after Loper Bright); Defendant Workday, Inc.’s Motion for Interlocutory Appeal Pursuant to 28 U.S.C. § 1292(b) at 6, Mobley v. Workday, Inc., No. 23-cv-00770 (N.D. Cal. Mar. 27, 2026), ECF No. 286 (renewing that argument and acknowledging the district court’s response that Rabin was “not contingent on deference to agency interpretation” and that Skidmore deference extends to the EEOC guidance).
This Essay makes three doctrinal moves. First, Mobley’s real foundation is not EEOC interpretation but Restatement (Third) of Agency sections 1.01 and 7.07. 10 Restatement (Third) of Agency §§ 1.01, 7.07 (A.L.I. 2006). The route into Title VII is textualist, because under Nationwide Mutual Insurance Co. v. Darden, the undefined term “agent” in 42 U.S.C. section 2000e(b) carries its common-law meaning. 11 Nationwide Mut. Ins. Co. v. Darden, 503 U.S. 318, 322-23 (1992) (“[W]here Congress uses terms that have accumulated settled meaning under . . . the common law, a court must infer . . . that Congress means to incorporate the established meaning of these terms.” (quoting Cmty. for Creative Non-Violence v. Reid, 490 U.S. 730, 739-40 (1989)). Second, the formal-control problem under sections 1.01 and 7.07(3)(a) is solved by a doctrinal label this Essay calls Functional Control Inversion, which Judge Lin’s reasoning in the July 2024 motion to dismiss order already invokes. 12 Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 806-08 (N.D. Cal. 2024). Third, the boundary between substantive participation (Workday) and mere tool provision (a database vendor without a scoring layer) tracks the decisional-weight axis already implicit in Mobley itself and in Walters v. OpenAI, L.L.C.’s defamation framework.
Part I canvasses the Mobley docket and the 2024 to 2026 administrative-law cascade that followed Loper Bright, that is, the rapid sequence of executive orders and agency reversals the decision set in motion. Part II defends the common-law foundation; Part III draws the boundary between vendors that substantively participate in hiring decisions and those that merely supply a tool. Part IV addresses industry-burden objections. The Essay begins where Mobley’s posture meets Loper Bright’s reversal, questioning whether the doctrine that survived July 2024’s motion to dismiss can survive June 2024’s earthquake.
I. Background
A. The Mobley Docket
Mobley filed in February 2023 under Title VII, section 1981, the ADEA, and the ADA; the complaint named no employer. 13 Complaint, Mobley v. Workday, Inc., No. 23-cv-00770 (N.D. Cal. Feb. 21, 2023), ECF No. 1. Judge Lin granted Workday’s first motion to dismiss with leave to amend in January 2024, then granted in part and denied in part the second in July 2024. 14 Mobley v. Workday, Inc., No. 23-cv-00770, 2024 WL 208529, at *7 (N.D. Cal. Jan. 19, 2024); Mobley, 740 F. Supp. 3d at 813-14. The intentional-discrimination claims fell, while the disparate-impact claims survived on an agency theory. The agency theory holds that, because Title VII reaches not only employers but also “any agent” of an employer, a vendor to which an employer delegates part of the hiring function can itself be liable as the employer’s agent. 15 Mobley, 740 F. Supp. 3d at 804-07. The pivotal sentence in the decision read Workday’s software as “not simply implementing in a rote way the criteria that employers set forth, but . . . instead participating in the decisionmaking process by recommending some candidates to move forward and rejecting others.” 16 Id. at 807. Later, in May 2025, Judge Lin certified a nationwide ADEA collective, finding that “[a]llegedly widespread discrimination is not a basis for denying notice.” 17 Mobley v. Workday, Inc., No. 23-cv-00770, 2025 WL 1424347, at *1 (N.D. Cal. May 16, 2025). Workday’s disclosure of approximately 1.1 billion processed applications during the relevant period made this potentially the largest collective ever certified. 18 Id. at *11; see TAC ¶ 30. A July 2025 expansion order folded HiredScore Spotlight and Fetch features, part of an entity acquired by Workday in 2024, into the certified class. 19 Mobley v. Workday, Inc., No. 23-cv-00770, slip op. at 1-2 (N.D. Cal. July 29, 2025) (order re HiredScore dispute).
B. The Loper Bright Cascade
In Loper Bright, Chief Justice Roberts’s six-justice majority overruled Chevron, holding that the Administrative Procedure Act requires courts to exercise independent judgment in deciding whether an agency has acted within its statutory authority and that courts may not defer to an agency interpretation simply because a statute is ambiguous. 20 Loper Bright Enters. v. Raimondo, 144 S. Ct. 2244, 2273 (2024). An administrative law cascade followed at speed: Executive Order 14179 revoked Biden Executive Order 14110; 21 Exec. Order No. 14179, supra note 6, at 8741. the EEOC removed its AI guidance on January 27, 2025; 22 The withdrawn guidance had explained that Title VII’s adverse-impact framework applies to employers’ use of algorithmic and AI selection tools. See EEOC, supra note 7. Executive Order 14281 directed agencies to deprioritize disparate-impact enforcement; 23 Exec. Order No. 14281, supra note 8. and the Department of Justice rescinded its Title VI disparate-impact regulations on December 10, 2025. 24 Rescinding Portions of Department of Justice Title VI Regulations, 90 Fed. Reg. 57141 (Dec. 10, 2025) (to be codified at 28 C.F.R. pt. 42) (eliminating the disparate-impact provisions of 28 C.F.R. § 42.104 on the ground that Title VI bars only intentional discrimination). The rescission is regulatory only and does not affect Title VII (private-employer) disparate-impact claims, which remain codified at 42 U.S.C. § 2000e-2(k). On the surface, every one of these developments is a headwind for AI-civil-rights litigation. But each operates at a different doctrinal layer, and the doctrinal path that carries Mobley’s agency theory into Title VII (§ 701(b)’s text, Darden, and the Restatement of Agency) does not depend on any of them.
C. The Common-Law Layer
What none of the 2024-2026 developments touched was Title VII’s text. § 701(b) defines “employer” to include “any agent of such a person.” 25 42 U.S.C. § 2000e(b). Parallel agent provisions appear at 29 U.S.C. § 630(b) (ADEA) and 42 U.S.C. § 12111(5)(A) (ADA). Congress did not define “agent” in the statute. The controlling case is Darden, decided unanimously by Justice Souter in 1992, which addressed undefined statutory terms generally rather than Title VII’s “agent” specifically 26 Nationwide Mut. Ins. Co. v. Darden, 503 U.S. 318, 322-23 (1992) (quoting Cmty. for Creative Non-Violence v. Reid, 490 U.S. 730, 739-40 (1989)). : Undefined common-law terms in federal employment statutes carry their common-law meaning unless the statute clearly indicates otherwise. Loper Bright’s instruction that courts use ordinary tools of construction to find a statute’s “single, best meaning” also points federal courts to that common-law meaning, supplied here by the Restatement (Third) of Agency. The doctrinal level on which Mobley’s agency theory operates is therefore not regulatory; it is common law, statutorily incorporated by Title VII § 701(b), and required by Darden’s textualist canon. 27 See 42 U.S.C. § 2000e(b); Darden, 503 U.S. at 322-23. What follows shows why that floor holds.
II. The Common-Law Foundation
A. The Restatement Framework
The agency relationship the Restatement of Agency defines is not a creature of administrative interpretation; it is a doctrinal structure with four discrete elements, 28 Restatement (Third) of Agency § 1.01 (A.L.I. 2006). each of which is either plainly satisfied in the AI-vendor scenario or is the subject of a doctrinal puzzle this Essay must solve. Section 1.01, which defines the principal-agent relationship, requires (1) the principal’s manifestation of assent, (2) that the agent act on the principal’s behalf, (3) subject to the principal’s control, and (4) the agent’s manifestation of assent or performance of the requested service. 29 Id.; id. cmt. d. The first two elements are uncontested. When an employer signs a master services agreement with Workday and uploads job postings, the employer has manifested assent, and Workday’s screening serves the employer’s hiring goals, not Workday’s own. 30 See Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 806-08 (N.D. Cal. 2024). The fourth element is satisfied under section 1.01 comment d, which clarifies that an agent’s assent need not be communicated to the principal. 31 Restatement (Third) of Agency § 1.01 cmt. d (A.L.I. 2006). However, the third element, control, is where black-box AI breaks the traditional analysis, 32 See infra Part II.B. and where Functional Control Inversion enters in Part II.B.
Section 7.07 governs whether vicarious liability flows to the principal. 33 Restatement (Third) of Agency § 7.07(1) (A.L.I. 2006). That is a separate question from the one section 1.01 answers. Section 1.01 asks whether the vendor is an agent at all, which § 701(b) makes directly actionable; section 7.07 asks the further question whether that agent is also an employee the principal answers for vicariously. Section 7.07(3)(a) defines an “employee” as “an agent whose principal controls or has the right to control the manner and means of the agent’s performance of work.” 34 Id. § 7.07(3)(a). Some courts treat this as more demanding than section 1.01’s “subject to control” language. 35 See, e.g., Alfaro-Huitron v. Cervantes Agribusiness, 982 F.3d 1242, 1254 (10th Cir. 2020) (contrasting the “minimal level of control” that establishes an agency relationship with the “much more significant and intrusive right of control that makes an agent an employee” under section 7.07(3)(a)); Quigley v. Rosenthal, 327 F.3d 1044, 1064 (10th Cir. 2003) (deeming control of “the manner of work performed” by the agent “a requirement necessary only if an agent is alleged to be an employee as opposed to an independent contractor”). However, the gap closes once the right-versus-exercise distinction in the comments is taken seriously. 36 See Restatement (Third) of Agency § 1.01 cmts. c, f. That distinction separates the right to control from its actual exercise: Section 7.07(3)(a) is met so long as the principal retains the right to control the manner and means of the agent’s work, even if it rarely or never exercises that right. The Supreme Court already imported common-law agency principles into Title VII analysis for harassment liability in Burlington Industries, Inc. v. Ellerth and Faragher v. City of Boca Raton. 37 Burlington Indus., Inc. v. Ellerth, 524 U.S. 742, 754-55 (1998) (drawing on the Restatement (Second) of Agency to ground Title VII supervisor-harassment liability in common-law agency principles); Faragher v. City of Boca Raton, 524 U.S. 775, 791-92, 808 (1998) (applying the same agency principles to hold an employer vicariously liable under Title VII). Common-law agency is the standard route into Title VII, not a novel one; 38 See Meritor Sav. Bank, FSB v. Vinson, 477 U.S. 57, 72 (1986) (declining to set a definitive rule but directing courts to look to common-law agency principles for guidance on Title VII employer liability); Ellerth, 524 U.S. at 754-55; Faragher, 524 U.S. at 791-92. the Essay’s contribution is showing why this standard route survives Loper Bright while the agency-deference route does not. Plaintiffs need not invoke section 7.07 for vicarious liability; section 1.01 establishes the vendor’s status as agent, which triggers direct liability under § 701(b)’s textual inclusion of “any agent.” 39 42 U.S.C. § 2000e(b) (including “any agent” of an employer within the statutory definition of “employer”).
Workday’s principal Ninth Circuit defense is Miller v. Maxwell’s International Inc., 40 Miller v. Maxwell’s Int’l Inc., 991 F.2d 583, 587-88 (9th Cir. 1993). which Judge Lin distinguished in the July 2024 order. 41 Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 806 (N.D. Cal. 2024). Miller addressed individual-supervisor liability driven by concerns about exposing individual employees with limited resources; that rationale does not extend to corporate algorithmic vendors. 42 See id. (distinguishing Miller). Beyond that distinction, Miller sits inside a broader Ninth Circuit pattern that has long extended Title VII obligations to non-direct-employer entities under common-law principles. In Bates v. Pacific Maritime Ass’n, then-Circuit Judge Kennedy applied the successorship doctrine to bind a non-original-party entity to the obligations of a Title VII consent decree, observing that “[a]n employer is not necessarily relieved of successor obligations simply because it displaces the former employer.” 43 Bates v. Pac. Mar. Ass’n, 744 F.2d 705, 708-10 (9th Cir. 1984) (Kennedy, J.) (extending Title VII obligations to non-original-party entity). Slack v. Havens applied the same principle nearly a decade earlier. 44 Slack v. Havens, 522 F.2d 1091, 1094-95 (9th Cir. 1975). The Supreme Court, along with the First, Second, and Eleventh Circuits, have all applied an “agent of the employer” theory to non-direct employers across decades. 45 See City of L.A., Dep’t of Water & Power v. Manhart, 435 U.S. 702, 718 n.33 (1978); Williams v. City of Montgomery, 742 F.2d 586, 589 (11th Cir. 1984) (per curiam); Spirt v. Tchrs. Ins. & Annuity Ass’n, 691 F.2d 1054, 1063 (2d Cir. 1982); Carparts Distrib. Ctr., Inc. v. Auto. Wholesaler’s Ass’n of New England, Inc., 37 F.3d 12, 17-18 (1st Cir. 1994). Workday’s Miller argument is a regional outlier. A corporate vendor exercising a delegated screening function is not Miller’s middle-manager defendant; it is exactly the kind of intermediary § 701(b) was meant to reach. 46 See 42 U.S.C. § 2000e(b); Slack, 522 F.2d at 1094-95.
B. The Control Inversion
The traditional formulation of agency control, where the principal directs and the agent executes, is inapplicable to the AI-hiring scenario. That failure is not, however, fatal to agency liability. Typically, the principal tells the agent what to do and monitors performance. 47 Restatement (Third) of Agency § 1.01 cmt. f (A.L.I. 2006) (defining principal-agent control). However, in black-box AI hiring, the employer cannot inspect proprietary decision rules 48 See TAC ¶ 101. or supervise their operation in real time. Read formally, the principal-controls-agent element is missing. 49 See Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 808 (N.D. Cal. 2024) (acknowledging the formal-control gap). This formal gap is the source of the independent-contractor argument that a vendor whose decision rules the employer can neither direct nor monitor is an independent contractor rather than an agent, and thus outside § 701(b). 50 See Restatement (Third) of Agency § 1.02 (A.L.I. 2006) (providing that how the parties characterize their relationship is not controlling; agency arises only when the elements stated in section 1.01 are present). Nevertheless, section 1.02 of the Restatement (Third) of Agency forecloses this move. A contractual “independent contractor” label is not controlling, because agency turns on whether section 1.01’s elements are present in fact. 51 Id.
Functional Control Inversion is not a new doctrine. It is the label this Essay coins for the form-versus-function principle Judge Lin invoked in the July 2024 motion-to-dismiss order. 52 See Mobley, 740 F. Supp. 3d at 806-08. This order read Workday’s software as participating in the decisionmaking process rather than implementing employer criteria in a rote way and observed that nothing in the federal anti-discrimination statutes distinguishes between delegating functions to an automated agent and to a live human one. 53 Id. (citing Williams v. City of Montgomery, 742 F.2d 586, 589 (11th Cir. 1984) (per curiam)). The Restatement (Third) of Agency confirms this reading. 54 See Restatement (Third) of Agency § 1.01 cmt. c (A.L.I. 2006). Section 1.01 comment c states that “a person may be an agent although the principal lacks the right to control . . . the agent’s exercise of professional judgment” and that “[a] principal’s failure to exercise the right of control does not eliminate it.” 55 Id. Section 1.01 comment f adds that “a principal has the right to give interim instructions or directions to the agent once their relationship is established.” 56 Id. § 1.01 cmt. f. And section 7.07 comment f, applying these principles to employees specifically, holds that “all employers retain a right of control, however infrequently exercised.” 57 Id. § 7.07 cmt. f. That right is what employers who use Workday retain, through control levers such as terminating the contract, disabling or reconfiguring features, auditing the vendor’s outputs, and overriding individual determinations. 58 See id. § 7.07 cmt. f. What these employers have delegated is the moment-to-moment screening function. Thus, section 1.01’s control element is satisfied. 59 Id. § 1.01. Functional Control Inversion describes how that delegated discretion takes algorithmic form.
Judge Lin did not name the Functional Control Inversion concept, but it is the underlying logic of her opinion. The class-certification order uses two phrases that, read together, articulate the insight precisely. Quoting the July 2024 motion-to-dismiss order, the court observes that “Workday’s customers delegate traditional hiring functions, including rejecting applicants, to the algorithmic decisionmaking tools provided by Workday.” 60 Mobley v. Workday, Inc., No. 23-cv-00770, 2025 WL 1424347, at *3 (N.D. Cal. May 16, 2025) (quoting Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 808 (N.D. Cal. 2024)). And the certification reasoning rests on the finding that collective members “were allegedly required to compete on unequal footing due to Workday’s discriminatory AI recommendations.” 61 Id. at *1. That delegation and that unequal footing are the inversion. The principal has surrendered substantive screening, and the agent’s decisional weight produces the disparate outcome. 62 See id. at *2, *9-10.
C. The Textualist Reinforcement
The conventional reading of Loper Bright, that the AI-civil rights litigation pipeline collapses with EEOC guidance, misreads what Loper Bright actually did. The misreading assumes Mobley’s agency theory leaned on the EEOC’s interpretation, so that withdrawing the guidance pulls the theory down with it. That assumption accords with Workday’s framing and conventional wisdom. Neither Mobley’s complaint nor Judge Lin’s orders invoked agency deference, and the EEOC’s guidance only confirmed the common-law meaning of “agent.” 63 .See TAC ¶¶ 9, 130-134 (alleging Workday “acts as an agent of its client-employers by exercising delegated authority over the hiring process”); Mobley, 740 F. Supp. 3d at 804-07 (construing “any agent” in 42 U.S.C. § 2000e(b) as a matter of statutory text and common-law delegation); Mobley, 2025 WL 1424347, at *7 (certifying the collective on that agency theory). None of these sources relies on Chevron deference or the EEOC’s guidance. Loper Bright relocated statutory interpretation from agency deference to text. 64 Loper Bright Enters. v. Raimondo, 144 S. Ct. 2244, 2266, 2273 (2024). For Title VII’s term “agent” in § 701(b), that relocation is a benefit, not a cost. Darden supplies the controlling canon: “Where Congress uses terms that have accumulated settled meaning under . . . the common law, a court must infer, unless the statute otherwise dictates, that Congress means to incorporate the established meaning of these terms.” 65 Nationwide Mut. Ins. Co. v. Darden, 503 U.S. 318, 322-23 (1992) (quoting Cmty. for Creative Non-Violence v. Reid, 490 U.S. 730, 739-40 (1989)). Title VII’s “agent” is a classic common-law term, undefined in the statute, with centuries of doctrinal content under the Restatements. 66 See Restatement (Third) of Agency § 1.01 (A.L.I. 2006). Under Darden, courts must apply that common-law meaning. 67 Darden, 503 U.S. at 322-23; see Pauline T. Kim, Data-Driven Discrimination at Work, 58 Wm. & Mary L. Rev. 857, 866-67 (2017) (arguing Title VII § 703(a)(2) prohibits classification bias). Loper Bright’s instruction that courts use ordinary tools of construction, including the common-law presumption, directly reinforces this reading. 68 See Loper Bright, 144 S. Ct. at 2273.
There are three additional reasons this theory survives. First, no Chevron deference is invoked: The Essay asks courts to apply Restatement definitions as Darden requires, and Loper Bright is irrelevant to this common-law application. 69 See id.; Darden, 503 U.S. at 322-23. Second, Skidmore deference survives in a supporting role. Loper Bright explicitly preserved Skidmore v. Swift & Co., which gave persuasive weight to agency interpretations with “power to persuade.” 70 Skidmore v. Swift & Co., 323 U.S. 134, 140 (1944); Loper Bright, 144 S. Ct. at 2262, 2267; see Brief for EEOC as Amicus Curiae, Mobley v. Workday, Inc., No. 23-cv-00770 (N.D. Cal. Apr. 9, 2024), ECF No. 60 (Skidmore-eligible EEOC interpretation of Title VII agency). Therefore, the EEOC’s April 2024 amicus brief in Mobley remains Skidmore-eligible, not as binding interpretation but as confirmatory evidence of the common-law analysis under Darden. 71 See Skidmore, 323 U.S. at 140 (according persuasive weight to an agency’s judgment based on “the thoroughness evident in its consideration, the validity of its reasoning, its consistency with earlier and later pronouncements, and all those factors which give it power to persuade”); Brief for EEOC as Amicus Curiae, supra note 70, at 2 (arguing that Mobley plausibly alleged Workday is an “agent of employers because employers have purportedly delegated authority to Workday to make at least some hiring decisions”). Third, the Burlington–Faragher line supplies a proof of concept. In both decisions the Supreme Court resolved Title VII employer liability through common-law agency, drawing the governing standard from the Restatement (Second) of Agency to hold employers vicariously liable for their supervisors’ harassment. 72 See Meritor Sav. Bank, FSB v. Vinson, 477 U.S. 57, 72 (1986); Burlington Indus., Inc. v. Ellerth, 524 U.S. 742, 754-55 (1998);Faragher v. City of Boca Raton, 524 U.S. 775, 791-92, 808 (1998). Importing the Restatement into Title VII is thus nothing new. This Essay asks only that courts take the same step with the Restatement (Third), which has since succeeded the Second. The doctrinal channel is well-worn; only the edition of the Restatement has changed. The Loper Bright earthquake reshaped administrative-law deference, but in doing so it required courts to read federal statutes in the textualist, common-law sense in which Title VII’s “agent” was always meant to be read.
III. The Decisional-Weight Boundary
A. The Archetype: Mobley
Workday autonomously screens, scores, and rejects applicants before any human at the employer reviews their files; within-minutes rejection emails confirm autonomous operation. 73 TAC ¶¶ 50, 62. That timing is an Iqbal marker: Rejection faster than plausible human review is itself a non-conclusory allegation of substantive delegation of decisional weight to the vendor. 74 See id.; Ashcroft v. Iqbal, 556 U.S. 662, 678-79 (2009) (distinguishing well-pleaded factual allegations, which are entitled to the assumption of truth, from conclusory assertions, which are not); Bell Atl. Corp. v. Twombly, 550 U.S. 544, 570 (2007) (requiring factual allegations that state a plausible, rather than merely conceivable, claim to relief). The Third Amended Complaint pleads the agency theory in its operative form: “Workday, Inc. acts as an agent of its client-employers by exercising delegated authority over the hiring process, including screening, ranking, and filtering applicants through its platform. Employers rely on the results of Workday’s selection procedures in making hiring decisions.” 75 Id. ¶ 9; see Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 808 (N.D. Cal. 2024). Where the algorithm operates as a gating mechanism, with rejected applicants never reaching human decisionmakers, the vendor has been delegated substantive decisional authority and is an agent of the employer for Title VII purposes. 76 Mobley, 740 F. Supp. 3d at 808.
B. The Floor: Tool Provision
A vendor that supplies a candidate database without scoring or filtering exercises no decisional weight. 77 See id. (drawing a boundary line). The principal-agent control inquiry never reaches the inversion problem because there is no decision being inverted. Judge Lin applies the same logic, rejecting agency liability for a software vendor whose product would only function as “a spreadsheet software program to sort workers by birthdate and then filtered out all applicants over the age of forty.” 78 Id.; see also Mobley v. Workday, Inc., No. 23-cv-00770, 2025 WL 1424347, at *7 (N.D. Cal. May 16, 2025) (adopting same boundary). The vendor is a tool provider, not an agent; Title VII liability does not attach to the vendor, and the employer remains primarily liable for whatever discriminatory practices it operates downstream. 79 See 42 U.S.C. § 2000e-2(k) (Title VII disparate-impact provision).
C. The Hard Case: Recommendation Layers
HiredScore, acquired by Workday in 2024, generates AI recommendations that an employer’s recruiters later review. 80 Mobley v. Workday, Inc., No. 23-cv-00770, slip op. at 1 (N.D. Cal. July 29, 2025) (describing HiredScore Spotlight and Fetch features). For these products, the vendor does not autonomously reject applicants. Consequently, the inversion question is fact-specific: Does the recruiter substantively follow the recommendation, or treat it as one input among several? If recruiters functionally rubber-stamp the AI output, the vendor is an agent; if not, the vendor is not. Decisional weight is a matter of degree. A recommendation that recruiters remain formally free to reject, yet in practice almost never override, carries the decisional weight of an autonomous rejection, so the vendor that supplies it is an agent. And where the recommendation layer predictably steers a protected group toward rejection, the recruiter’s formal final say does not defeat agency, because a biased input that proximately drives the adverse decision remains actionable even when a human ratifies it, and a discriminatory component of a selection process is not cured by a balanced overall result. 81 See Connecticut v. Teal, 457 U.S. 440, 455-56 (1982) (holding that a discriminatory component of a selection process is actionable even where the overall “bottom line” is balanced); 29 C.F.R. § 1607.4(C) (2024) (providing that where the total selection process shows adverse impact, the individual components are evaluated for adverse impact); cf. Staub v. Proctor Hosp., 562 U.S. 411, 420-21 (2011) (holding that a decisionmaker’s exercise of judgment does not prevent a biased subordinate’s act from being the proximate cause of the adverse employment action). At the pleading stage, plaintiffs can plausibly allege functional rubber-stamping by pointing to objective public markers, including vendor marketing claims that the product reduces recruiter screening time by a stated percentage, or statistical evidence that candidates below an algorithmic threshold are uniformly bypassed without human interview. 82 See generally Ifeoma Ajunwa & Cheng-chi (Kirin) Chang, Provenance, 73 UCLA L. Rev. (forthcoming 2026) (proposing an input- and output-provenance framework for verifying and auditing A.I.-generated content, including automated hiring systems that encode historical discrimination). Judge Lin’s July 2025 expansion order folded HiredScore Spotlight and Fetch features into the certified class, expressly rejecting Workday’s arguments that HiredScore was acquired post-FAC and that it was “a separate product, built on a wholly separate technology platform,” and deferring to the decertification stage Workday’s argument that the HiredScore features used materially different scoring algorithms. 83 Mobley, slip op. at 1-2 (N.D. Cal. July 29, 2025) (quoting Letter Brief – Parties Joint Letter Regarding HiredScore at 2, Mobley v. Workday, Inc., No. 23-cv-00770 (N.D. Cal. July 16, 2025), ECF No. 157). Vendor liability follows the product-line family, not a single stock-keeping unit (SKU); the agency analysis travels with the integration. 84 See id. at 2 (rejecting the product-separation defense). The mechanism is the unit of analysis the court itself used. The expansion order defines the collective by the “unified policy” of using Workday’s AI recommendation system “to score, sort, rank, or screen applicants,” a definition that turns on what the software does within the employer’s hiring process, not on which acquisition, code base, or brand supplies it. 85 Id. at 1 (holding that the collective’s “unified policy” includes the HiredScore AI features). The order polices the same functional line at the margins: Workday may exclude a customer only by showing that the customer received no scores or rankings, or never scored or screened candidates based on the AI features; exclusion negates the delegated function, not the product label. 86 Id. at 2 (setting out the customer-exclusion mechanism).
D. The Anchors: Open-Source, the Fair Credit Reporting Act, and Walters
Three boundary cases anchor the framework, fixing its outer edges by showing where agency liability attaches and where it does not. First, under open-source models, where a publisher releases model weights and an employer self-deploys, no principal-agent relationship arises between the model author and the deploying employer. 87 See Restatement (Third) of Agency § 1.01 (A.L.I. 2006) (requiring the principal’s manifestation of assent that the agent act on the principal’s behalf and subject to the principal’s control, and the agent’s consent so to act). Liability attaches to the deploying employer; the framework does not extend to model publishers because there is no contractual delegation. Second, the Eightfold pivot: In Kistler v. Eightfold AI, Inc., filed January 20, 2026 by counsel including former EEOC Chair Jenny Yang, plaintiffs chose the Fair Credit Reporting Act 88 15 U.S.C. §§ 1681–1681x. and California’s Investigative Consumer Reporting Agencies Act (ICRAA) 89 Cal. Civ. Code § 1786. over Title VII agency theory. 90 Complaint at 28, 31, 33, Kistler v. Eightfold AI, Inc., No. C26-00214 (Cal. Super. Ct. Contra Costa Cnty. Jan. 20, 2026). Eightfold’s recommendation-only profile placed it lower on the decisional-weight axis than Workday, since Eightfold only surfaces recommendations for a human recruiter to review rather than autonomously rejecting applicants, and counsel routed through a different statutory framework. 91 Id. at 3. Third, Walters v. OpenAI: 92 Walters v. OpenAI, L.L.C., No. 23-A-04860-2, slip op. at 4, 7-8 (Ga. Super. Ct. Gwinnett Cnty. May 19, 2025) (granting summary judgment to OpenAI and finding that Riehl, a journalist, recognized the output’s falsity “within about an hour and a half”). Here, the Georgia court’s grant of summary judgment to OpenAI illustrates the framework’s outer edge. ChatGPT was queried by a journalist who chose the prompt, read the output, and recognized its falsity within ninety minutes; 93 Id.; cf. Cheng-chi (Kirin) Chang, Hallucinated Authority: AI Citations as Reckless Misrepresentation, 115 Geo. L.J. Online (forthcoming Sept. 2026) (manuscript at Part III) (on file with author) (locating responsibility for AI-generated falsehoods in the attorney who adopts and submits them, rather than in the tool that generated them). on the framework advanced here, OpenAI was not exercising any function the user had traditionally exercised, and the user remained the decisional actor. 94 See Walters, slip op. at 6-8. The framework distinguishes the two not by technology but by who has been delegated decisional authority. 95 Cf. Sandra G. Mayson, Bias In, Bias Out, 128 Yale L.J. 2218, 2221 (2019) (arguing that algorithmic disparity reflects structural inequality). For the canonical analysis of algorithmic disparate impact under Title VII, see Solon Barocas & Andrew D. Selbst, Big Data’s Disparate Impact, 104 Calif. L. Rev. 671, 694 (2016). In Walters, the user chose the prompt and judged the output, so the user remained the decisionmaker; in Mobley, the employer handed the screening decision itself to the vendor. The controlling question is which actor holds the power to decide; whether that actor is an algorithm or a human is beside the point. Thus, the boundary between substantive participation and tool provision tracks decisional weight directly and gives federal courts a workable test at the pleading stage.
IV. The Industry-Burden Objection
The strongest counterargument is pragmatic, not doctrinal. Defendants will argue holding AI vendors directly liable as agents will paralyze the Software as a Service industry. Such a complaint fails. First, agency liability has applied to staffing agencies, professional employer organizations, and traditional recruiters for decades, and those industries have not collapsed. 96 See Spirt v. Tchrs. Ins. & Annuity Ass’n, 691 F.2d 1054, 1063 (2d Cir. 1982); Williams v. City of Montgomery, 742 F.2d 586, 589 (11th Cir. 1984) (per curiam); Carparts Distrib. Ctr., Inc. v. Auto. Wholesaler’s Ass’n of New England, Inc., 37 F.3d 12, 17-18 (1st Cir. 1994). Second, the framework’s boundary test in Part III explicitly limits agency status to substantive-participation vendors; pure tool providers are excluded. 97 See supra Part III.B (describing database vendor as floor). Third, vicarious liability has historically been a feature of mature service industries, not a bug. 98 See Restatement (Third) of Agency § 7.07 (A.L.I. 2006).
Contractual risk allocation is the actual mechanism, not a problem to be avoided. Workday already negotiates indemnity provisions with its more than ten thousand enterprise customers; 99 See TAC ¶ 30. standard indemnity language allocates exactly this kind of legal risk between vendor and employer. The question is not whether the risk is allocated, because it always is, but which allocation the law endorses. Agency liability under common-law doctrine endorses an allocation that holds the substantive decisional actor accountable; 100 Cf. Restatement (Third) of Agency § 7.07(1) (A.L.I. 2006) (vicarious liability for tortious agent conduct). contractual indemnity then reallocates between vendor and employer through arms-length negotiation.
California’s October 2025 Fair Employment and Housing Act (FEHA) regulations codify the agency theory at the state regulatory level. 101 Cal. Code Regs. tit. 2, § 11008(a) (2025). The operative definition treats as an “agent” “any person acting on behalf of an employer, directly or indirectly, to exercise a function traditionally exercised by the employer or any other FEHA-regulated activity . . . including when such activities and decisions are conducted in whole or in part through the use of an automated decision system.” 102 Id. That language closely tracks Judge Lin’s Functional Control Inversion formulation in the July 2024 motion-to-dismiss order, and was finalized after that order was issued. The implication of this new regulatory language transforms the Essay’s posture: This is no longer a piece of doctrinal advocacy asking courts to recognize a new theory of vendor liability. It is doctrinal mapping, describing a functional-control principle that has now been codified into California regulatory text and adopted by the Mobley plaintiffs’ counsel in their operative complaint, the Third Amended Complaint. 103 TAC ¶ 9. Federal common-law analysis under the Restatement (Third) of Agency, the Mobley line of orders, and California’s regulatory text now point in the same direction. 104 Compare Restatement (Third) of Agency § 1.01 (A.L.I. 2006) (defining an agent as one who acts on the principal’s behalf and subject to the principal’s control), with Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 806-08 (N.D. Cal. 2024) (treating a vendor whose software participates in the hiring decision as the employer’s agent), and Cal. Code Regs. tit. 2, § 11008(a) (2025) (defining “agent” to include any person acting on behalf of an employer, directly or indirectly, to exercise a function traditionally exercised by the employer, including through an automated decision system). Convergence across three independent drafters, a federal judge, the California Civil Rights Council, and the Mobley plaintiffs’ counsel, is itself strong evidence that the doctrinal architecture is sound. The framework does not paralyze the industry, nor does it depend on any single agency’s enforcement priorities. Instead, it places liability where decisional weight rests, mirrors what California has already codified, and lets contractual indemnity handle the rest. 105 The pathway here is complementary to regulatory and ex ante alternatives. See Ifeoma Ajunwa, An Auditing Imperative for Automated Hiring Systems, 34 Harv. J.L. & Tech. 621, 630-31 (2021) (mandated audits); Aziz Z. Huq, Constitutional Rights in the Machine-Learning State, 105 Cornell L. Rev. 1875, 1884 (2020) (ex ante regulation plus aggregate litigation).
Conclusion
Loper Bright did not destroy AI-civil-rights litigation; it relocated the doctrinal discussion from administrative-deference to the older, sturdier ground of common-law agency, exactly the ground Darden marked thirty-three years ago. 106 Nationwide Mut. Ins. Co. v. Darden, 503 U.S. 318, 322-23 (1992). The route runs through Title VII § 701(b)’s “agent,” Darden’s textualist common-law canon, section 1.01 of the Restatement (Third) of Agency, and the Functional Control Inversion principle. 107 Id.; 42 U.S.C. § 2000e(b); Restatement (Third) of Agency § 1.01 (A.L.I. 2006); see Mobley, 740 F. Supp. 3d at 806-08. The boundary between substantive participation and tool provision tracks the decisional-weight spectrum, not a metaphysical line between AI and non-AI. Put differently, what triggers liability is the employer’s handoff of the decision to someone else. A vendor that decides who advances is an agent, whether it decides by algorithm or by human judgment. And federal-state convergence, with California’s FEHA regulations tracking Judge Lin’s reasoning almost verbatim, confirms that the doctrinal analysis holds. 108 See Cal. Code Regs. tit. 2, § 11008(a) (2025).
The Mobley litigation will likely reach summary judgment within the next few years, and the courts that decide it will be working on common-law agency ground that Loper Bright and Darden left firmer than the conventional reading assumes. The algorithm was always the agent; Loper Bright and Darden together simply force us to remember why.
*Cheng-chi (Kirin) Chang (張正麒), Research Fellow, U.S.-Asia Law Institute, NYU Law; Affiliate Research Fellow, AI and the Future of Work Program, Emory Law (formerly Associate Director and Academic Fellow). I am grateful to Katherine A. Wilhelm, Ifeoma Ajunwa, and the many friends and colleagues whose generosity and encouragement have meant so much to me, in this work and well beyond it. Thanks to Lawrence B. Solum for his encouraging “Highly Recommended” of a draft of this Essay on the Legal Theory Blog. I also thank Sandra Chen, Hannah Lynn Dahleen, Raj Gambhir, Elise Faith Miner, Jane Alice O’Connell, William Summers, Megan Tran, and Jessica Jiaxi Zhu, and the other editors of the Stanford Law Review for their care and dedication in bringing this Essay to publication. Any errors or oversights are my sole responsibility.