24AUG

Intuit cuts 3,000, licenses its data

4 min read

16:17UTC

Intuit cut 17% of staff and, in the same announcement, agreed to feed its tax and finance data into Anthropic and OpenAI, turning displaced workers' expertise into training signal for their replacement.

← Back to Four methods, one answer on AI and jobs Jump to analysis ↓

EconomicDeveloping

Key takeaway

Each licensing-plus-layoff round widens the data moat of the firm supplying the replacement.

Intuit cut about 3,000 staff, 17% of its workforce, on 20 May 2026, with chief executive Sasan Goodarzi citing a refocus on AI across TurboTax, QuickBooks and Credit Karma.⁠¹ Affected US employees leave on 31 July with 16 weeks' severance, and the company is closing its Reno, Nevada operation.

The same memo signed the loop shut. Intuit struck multi-year deals to feed its specialised tax and financial data into Anthropic's Claude and OpenAI's models.⁠² The institutional knowledge those workers built, years of processing returns and reconciling accounts, becomes the training signal for the systems taking over the work. A worker who lands a new job has helped train the model that closed the door behind them.

This is the sharpest version of a pattern the topic has tracked loosely. Anthropic's Project Glasswing established the template of frontier firms accumulating proprietary sector data under privileged-access deals; the Intuit arrangement extends that hoard to consumer tax and financial records. Separately, Anthropic owes the proposed $1.5bn Bartz class fund, in which HarperCollins expects a share. That settlement is not a News Corp data-licensing deal.

A general model cannot file a US tax return well; one trained on Intuit's reconciled filings can, which is why the data is worth more than the headcount. Each licensing-plus-layoff round narrows the set of firms that hold sector-specific expertise and widens the lead of the one supplying the replacement, so displacement and data concentration are the same movement seen from two sides.

Deep Analysis

In plain English

Intuit makes TurboTax (for filing personal tax returns), QuickBooks (for small-business accounting), and Credit Karma (a personal finance app). These products handle sensitive, specialised financial data: tax returns, business cash flows, credit histories. Anthropic and OpenAI pay premium prices for this data because it is exactly what an AI model needs to answer financial and tax questions competently. On 20 May 2026, Intuit's chief executive Sasan Goodarzi announced the company was cutting about 3,000 workers, 17% of everyone it employs. On the same day, he announced multi-year licensing deals to feed Intuit's proprietary data into Anthropic's Claude (an AI assistant made by Anthropic, an AI safety company) and OpenAI's models (the company behind ChatGPT). This matters because the workers being cut spent years building and refining the knowledge base that makes Intuit's products useful. A tax specialist at Intuit knows which edge cases trip people up, which return errors are common, how to explain complex rules in plain English. That expertise, encoded into Intuit's historical data and product interactions, is now being licensed to train the AI systems that will handle those same tasks. The displaced workers helped create the training material for the systems replacing them. Anthropic's proposed $1.5bn Bartz settlement concerns past book copying. HarperCollins expects a share through the class. That payment is separate from commercial data partnerships and does not establish a new supply of News Corp content.

Deep Analysis

Root Causes

Two structural incentives intersected when Intuit closed the deal: AI model companies pay premium prices for proprietary domain-specific datasets, and Intuit's institutional knowledge was at peak transferability before the human workflow closed.

First, AI model companies are paying premium prices for proprietary, high-quality domain-specific datasets that cannot be assembled from public internet data; tax and financial records are the highest-value category. Second, Intuit's workforce displacement was planned on a timeline that coincides with the AI refocus, meaning the institutional knowledge those workers carry is at peak transferability before the human workflow is replaced.

The absence of any legal obligation for Intuit to share licensing proceeds with affected workers reflects a gap in employment and intellectual property law that predates the AI transition: US employment contracts routinely vest all work-product IP in the employer, and collective bargaining agreements have historically not addressed data-licensing proceeds.

Dario Amodei's January 2026 essay called on AI companies to 'steer customers away from firing workers' , but Anthropic then signed the Intuit deal, suggesting the voluntary framing is insufficient against commercial incentives.

What could happen next?

Consequence
Anthropic and OpenAI gain proprietary tax and financial training data that strengthens their models' financial-advisory capability, widening the competitive gap with model companies that lack such domain-specific datasets.
Short term · Assessed
Risk
The Intuit structure, simultaneous displacement and data licensing, may become a template for financial services, healthcare and legal-sector employers seeking to convert workforce transitions into model-training assets.
Medium term · Assessed
Precedent
No legal challenge to the data-licensing component has been filed; without a WGA-style union claim or regulatory intervention, the Intuit structure sets a legally unchallenged precedent for labour-knowledge transfer.
Long term · Assessed

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: Standard Oil's 1879 acquisition of the Cleveland refineries provides the closest structural analogue to the Intuit data-licensing loop. Rockefeller bought out competing refiners, absorbed their operational knowledge into Standard Oil's process engineering, and then reduced employment at the acquired sites.

The workers' accumulated refining expertise became a Standard Oil asset without any transfer payment to the workforce. Federal antitrust action 30 years later broke the accumulation but did not redistribute the captured expertise value to workers.

Counter-parallel: Hollywood's 1960 Screen Writers Guild agreement with the major studios established residual payments for reuse of writers' work in syndication and later home video. The WGA used a property-rights argument: the writer's contribution was embedded in the licensed product and should attract a royalty on reuse.

That precedent failed to transfer to AI training in the 2026 WGA settlement, where studios rejected payment for AI training use of writers' work. The counter-parallel shows that the residual-rights mechanism exists but has already been tested and defeated at the exact legal frontier where Intuit's data deal sits.

Intuit's 17% workforce reduction against multi-year AI data deals suggests a permanent operating-model shift, not a one-time efficiency exercise. TurboTax alone processes tens of millions of US tax returns annually, each a structured data record with high training value for financial AI. If Anthropic's Claude can answer TurboTax-level questions competently on Intuit's data, the addressable market for TurboTax's human-assisted tier contracts substantially.

The 16-week severance package for 3,000 US workers represents approximately $360-480m in one-time costs at median Intuit US salaries, which will be more than offset by the annual payroll reduction and the revenue from the multi-year data licensing deals.

Intuit shareholders gain from the combined payroll reduction and licensing revenue. The net effect for the US tax-preparation labour market is a permanent reduction in demand for human tax specialists at scale, as the AI-trained replacement improves precisely because Intuit has supplied it with Intuit-quality training data.

Consensus view: Elettra Bietti (Northeastern University School of Law) and Salomé Viljoen (Michigan Law) have argued that data-licensing deals negotiated simultaneously with workforce displacement announcements raise a structural property-rights question distinct from standard IP licensing: the workers who generated the institutional knowledge have no legal claim to a share of the licensing proceeds, because employment-for-hire doctrine vests that knowledge in the employer.

Viljoen's data-relations framework, published in 2021, anticipates this structure precisely and identifies it as a wealth-transfer mechanism from labour to the data-accumulating platform, operating through legally unremarkable transactions.

Counter-view: Kai-Fu Lee (Sinovation Ventures), drawing on parallel AI industry data deals in China where domestic regulators imposed revenue-sharing requirements on platform-to-model licensing, argues that the US framing of this as a property-rights failure misses the market dynamic: Anthropic and OpenAI have significant leverage to refuse the deal or offer lower terms, meaning displaced workers should seek negotiation power through their unions rather than through property-rights claims that courts will not recognise.

Lee notes that Baidu's 2024 licensing deals with Chinese state-media archives included 3% revenue sharing with the originating institution, a negotiated outcome that US unions have not yet sought.

Key tension: Whether the wealth generated by Intuit-quality proprietary training data can be captured by the workers who built it through any available legal or collective mechanism, or whether it structurally accrues to the model companies regardless of labour negotiation strategy.

Sources:NASSCOM

Mentions:Reno →Anthropic →OpenAI →Michigan →ChatGPT →China →Nevada →News Corp →Claude →Project Glasswing →Intuit →Dario Amodei →

First Reported In

Update #10 · Rival studies split on AI's hit to jobs

NASSCOM· 24 May 2026

Read original →

Causes and effects

Caused by

AISI: GPT-5.5 matches Mythos on 32-step attack

Anthropic's Project Glasswing established the model of frontier-AI firms accumulating proprietary sector data; Intuit's multi-year deal extends that accumulation to consumer tax and financial data.

Occurred 1 May 2026

Read story →

This Event

Intuit cuts 3,000, licenses its data

The displaced lose their jobs and then see their expertise licensed to the system that replaced them.

Different Perspectives

TSMC

The chipmaker raised its 2026 revenue growth forecast to about 44% on 15 October while holding capital spending at $60-64bn. High-performance computing, mostly AI data-centre chips, made up 68% of quarterly revenue, so demand for the hardware behind automation is still rising.

South Korean government

Its statistics office published the first youth figures on the 15-to-34 band set by a revised law, showing a fall of 98,000 against 147,000 on the old band. Including people in their early thirties, whose employment rose, makes the official youth decline look gentler.

Office for National Statistics

The ONS put UK vacancies at 691,000 for July to September and professional, scientific and technical vacancies 11.6% lower on the year. In September it had shown the same sector still adding jobs, so the fall appears in adverts before it reaches headcount.

Chinese government

Officials told a State Council Information Office briefing on 10 October that 10.52 million new urban jobs had been created by September, 87.7% of the annual target. Official figures show investment falling overall but rising in networks, data services and robot output, with urban unemployment at 5.2%.

Tharman Shanmugaratnam

Singapore's President, co-chair of the World Bank Jobs Council, urged governments at the Bangkok meetings to make crisis-era youth employment support a standing policy tool. That treats young workers' AI exposure as a lasting budget line rather than a shock to ride out.

Srini Pallia, Wipro

Wipro's chief executive called AI a long-term growth driver as the company cut 6,716 staff on flat revenue. HCLTech and TCS added staff in the same quarter, so Wipro's cut sits beside its own weak revenue rather than an industry-wide fall.

Written with AI from Lowdown's published reporting. Not individually reviewed by an editor. Editorial standards.