Institutional procurement refusal of AI vendors

Pressure is brought to bear on a large institutional buyer — a school district, a university system, a hospital network, a city government, a pension fund — to refuse a contract with a specific AI vendor, to cancel an existing contract, or to write a procurement standard that no vendor in a given category currently meets. The campaign is shaped accordingly: a public letter from faculty or students, a budget-hearing turnout, an investigative piece naming the contract, a board-meeting motion, sometimes a tied divestment demand. The immediate output is a single procurement decision, but the decision is built to be cited by the next institution facing the same vendor.

An actor chooses this strategy because institutional buyers are the demand side AI companies actually depend on, and a refusal at a hospital network or a state university system is a revenue loss the vendor sees in the next quarter — a feedback loop regulation does not produce on the same timescale. The strategy also exploits a governance asymmetry: many institutional buyers have public-facing trustees, faculty governance, or open budget processes that are far more responsive to organised constituents than a state legislature is. Each won refusal narrows the vendor's customer base and provides precedent and procurement language the next campaign can lift verbatim.

It trades scope for that asymmetry. A single procurement refusal protects only one institution's users and rarely outlasts a friendly administration; vendors counter by offering the next administration a discount, packaging the AI inside a bundle the buyer cannot refuse, or re-entering via a sub-contractor. The strategy depends on a coalition inside the institution willing to absorb the political cost — faculty, students, union members, board allies — a constituency campaigns can build only where one already exists in latent form.

Verdict — the position-statement layer ships fast; the buyer-side refusal layer is the gap

The strategy works on the propagation of a procurement-refusal posture across the sectoral-association layer. Within five years it has moved from no organised sectoral-pivot adoption to four professional associations carrying institutional procurement refusal as a standing position: the NEA's July 2024 representative-assembly five-principle policy with model contract language delivered to 2.8-million-member affiliates, the ALA's October 2025 ACRL competency framework opening with the proposition that "adoption of AI technologies is neither necessary nor beneficial in all cases", the April 2026 ALA AI Policy Working Group draft forbidding library staff from connecting patron records to AI tools, and the AMA's 2018 H-480.940 framework establishing clinical-validation gates on healthcare AI deployment. These are professionally-credentialled instruments delivered to the working-level institutional buyer (a school district, an academic library, a hospital); their force is not the document itself but the social-and-legal pressure the document supplies for a sympathetic procurement officer to lean on, and the speed with which the sectoral-association layer has filled in is the strategy's clearest near-term claim to working.

The strategy is strong as a vehicle for installing a procurement-refusal posture at the sectoral-organisation layer, and weaker as a tracer of actual buyer-side refusal decisions.

Strong on the corpus's clearest concrete procurement-refusal cluster — the 2019 music-festivals track, in which Fight for the Future's artist-pressure mobilisation produced ≥40 major US festival commitments not to deploy facial recognition at events alongside LiveNation and AEG publicly reversing posture inside the same four-month window. The festivals case is the strategy's clearest worked example at private-institutional-buyer scale and the four-month cycle time is the strategy's clearest tempo advantage over the legislative arms running on the same harm. It is also a single cluster in one sector with a campaign-side anchor outside the sectoral-association layer the rest of the strategy's adopters occupy — a worked instance, not a documented pattern.

Weaker on the position-to-procurement conversion. The professional-association layer is upstream of any actual institutional refusal — a school board adopting an NEA model resolution, an academic library implementing the ALA AI Policy Working Group draft, a hospital network gating AI procurement on H-480.940. The strategy's working theory is that the upstream position legitimates the downstream refusal, and the downstream refusal is the strategy's actual output. The corpus's record of downstream refusal driven by the upstream position is structurally thin: no school district is yet documented in the corpus to have refused a specific AI vendor citing the NEA's model resolution; no academic library is yet documented to have refused a specific vendor on the ACRL Framework's authority; no AMA-member hospital is documented to have declined an AI clinical-decision tool on H-480.940's procurement-validation grounds. The strategy's organisational substrate is real and growing; the procurement-record substrate is the part the next decade's evidence will say whether the strategy actually delivers.

Weaker on the constituency-strength constraint. The strategy works in K-12 education because the NEA has 2.8 million members and a representative-assembly democratic instrument; it works in academic libraries because the ALA has 47,000 institutional and professional members organised across eight divisions; it works in healthcare because the AMA has 271,000 physicians and a 190-plus specialty-society infrastructure. Where the sectoral organisational substrate is thinner — public-sector social services, retail and warehouse work, journalism, the gig-economy workforce — the strategy has no analogue working at scale in the corpus, and the AI-deployment harms that those workforces face (automated benefits-eligibility decisions, algorithmic worker management, generative-AI displacement of editorial labour) are not currently held by a sectoral-association adopter the strategy could carry. The strategy is structurally pegged to a pre-existing professional-community infrastructure able to carry the public-affairs work and to discipline its member institutions toward the position; that infrastructure exists in the US in a small number of mature sectors and is absent or thin in many others.

Weakest on the co-optation problem. The professional association whose AI position is "AI must be carefully validated" is simultaneously the association whose AI position is "AI must be deployed responsibly", and the same governance instrument that gates one procurement decision legitimates another inside the same institutional voice. The Library Copyright Alliance's October 2025 statement that existing copyright law is adequate for generative-AI training — and the ALA-and-ARL-joined amicus brief in Bartz v. Anthropic carrying the same fair-use posture — is procurement-favorable framing for the LLM industry's training-data acquisition layer, sitting alongside the ALA's deployment-side procurement-refusal posture inside the same association. The AMA's "augmented intelligence" terminological commitment is a frame that can be read as procurement-favorable as easily as procurement-gating; the NEA's federal-advocacy push for AI literacy curriculum and educator-embedded research hubs produces procurement budgets for AI training and tools the procurement-refusal posture is supposed to gate. The strategy's reach inside the association does not predict the association's net effect on member procurement; an honest reading must trace which side of the association's institutional voice carries the procurement decision in any given case, and the corpus does not yet contain that evidentiary trace.

Ecology

The strategy is paired with municipal affirmative ban on a class of AI use as the corpus's two-arm buyer-side denial cluster. Both strategies operate by denying a buyer market to the vendor rather than by regulating use; the difference is the refusing actor — the city government on one side, the university, school district, library, hospital, or private firm on the other. The municipal arm has produced the denser concrete-refusal record (17 US municipal facial-recognition bans by May 2022); the institutional arm has produced the deeper sectoral-organisation substrate. Run together they cover the universe of non-legislative buyer actors; run alone either arm has structural ceiling problems the other does not — the municipal arm is preempted at the state level, the institutional arm is bypassed by the institution's federal partner.

It is paired with tech worker refusal inside AI labs as the supply-side mirror. The tech-worker arm refuses to build the vendor's product; the institutional-buyer arm refuses to purchase it. Together they apply pressure at both ends of the procurement transaction. The 2019 #NoTechForICE campus pledge programme — computer-science majors refusing Palantir employment until the ICE contract ended — and the 2019 UC Berkeley Privacy Law Scholars Conference dropping Palantir sponsorship are paired pressure surfaces on the same vendor: one operates inside Palantir's labour market, the other inside Palantir's institutional reputation market. The two arms are most powerful when run on the same vendor in the same political moment; run apart they fragment the campaign's narrative and the vendor weathers each arm separately.

It is paired with creator-class collective bargaining on generative AI as the sectoral-organisation route to enforceable refusal. The bargaining arm produces contractually enforceable restrictions on AI use inside a bargaining unit — the SAG-AFTRA five-category digital-replica regime, the WGA Article 5 training-data restrictions — while the institutional-procurement-refusal arm produces position statements that supply procurement language to the bargaining team and to the institutional procurement officer. The NEA's model collective-bargaining contract language is the corpus's clearest single instance of the two arms running through the same sectoral organisation in the same vehicle.

It is fed by counter-narrative framing of AI harm at the public-facing register. The "AI must not replace the educator / librarian / physician" frame is the procurement-refusal posture's working rhetorical anchor — Becky Pringle's "the student-educator connection must always be the center", the ACRL Framework's "Skepticism" mindset, the AMA's "augmented intelligence rather than artificial intelligence" terminological commitment. Without the framing the position-statement defaults to "what safeguards should accompany the technology", which is a procurement-conditioning frame, not a procurement-refusing one.

It is fed by empirical audit and expose at the substrate. The NEA policy statement explicitly cites racially-biased AI grading systems, facial-recognition failures against Black students, and object-recognition tools wired to alert police as the operational-harm anchors that produced the political pressure for the five-principle vote. The Algorithmic Justice League's Gender Shades audit findings carry into the school-FRT procurement-refusal lineage; the AMA's prior-authorisation campaign is anchored on the 1.2-seconds-per-claim Cigna case as its operational-harm number. The audit number is what converts an aspirational position into a procurement-decision argument the institution's general counsel can stand behind.

It is structurally adjacent to civil society inside technical standards bodies — both shape the vendor-acceptance terms before the procurement decision rather than the procurement decision itself. Standards bodies set the technical baseline a vendor must meet; sectoral associations set the organisational-policy baseline a buyer institution should apply. Together they form the corpus's two upstream-of-procurement arms — the technical standard and the institutional position — and a vendor that fails either is gated. The two arms are partly complementary and partly substitutable: where a binding standard exists, the procurement-refusal position rides on the standard's authority; where the standard is weak or industry-captured, the procurement-refusal position is the only floor.

The strongest competing strategy on the AI-developer side is the public-sector AI procurement push — the AI industry's organised effort to insert AI deployment as default-on inside large institutional contracting frameworks (federal-government AI procurement directives, state-government AI procurement standards, large EdTech and EHR vendor bundles where the AI is a sub-component the buyer is not asked to separately approve). The institutional-procurement-refusal strategy's hedge against the counter-offer is the line-item disclosure requirement — that any AI sub-component inside a bundle must be separately disclosed and separately consented to at the procurement layer — which the ALA AI Policy Working Group draft, the AMA's seven-principle disclosure requirement, and the NEA's transparency-into-AI-algorithms federal-policy ask each carry in different sectoral vocabularies. Where the disclosure requirement holds, the refusal can attach to the AI sub-component without refusing the whole bundle; where it doesn't, the AI rides into the institution as a default the procurement-refusal posture cannot reach.