Iceberg ahead! We’re only seeing the tip of changes from AI

On a quarterly spend report, $10,000 that used to flow to a freelancer marketplace now shows up as just a few hundred dollars of model usage, tied to an expense platform dataset and a handful of vendors. The shift looks small in a chart, until you translate it into unit economics. 

In a Ramp research paper, Ryan Stevens uses payments data from thousands of firms to track spending from the third quarter of 2021 through the third quarter of 2025, straddling the October 2022 release of ChatGPT. The result delivers a lesson that leaders can apply to real budgets.

The headline result deserves to lead every policy conversation about generative AI and work. Among the most exposed firms, each $1 decline in online labor marketplace spending aligns with about $0.03 of additional spending on AI model providers. That implies cost savings of roughly 97 percent when companies replace outsourced task labor with model usage. 

The roughly 25-to-one expense ratio explains why adoption keeps expanding, even as executives debate hallucinations and brand risk. This gap changes the budget logic for writing drafts, first-pass research, customer support macros, lightweight coding scaffolds, and structured summaries. Contract labor typically scales linearly: more output requires more hours, more invoices, and more coordination time. Model usage scales via throughput: prompts, retrieval, evaluation, and review create a workflow where marginal output costs stay low. 

The paper also shows the substitution in spending shares. Online labor marketplaces fall from 0.66 percent of spend in Q4 2021 to 0.14 percent by Q3 2025, while AI model providers rise to 2.85 percent by Q3 2025. Ramp’s public write-up of these patterns in its AI spend report matches what many finance teams already see: a vendor line item that looks like software yet displaces labor budgets. 

Stevens captures the visible tip of the iceberg because his measurement focuses on external online labor marketplaces and model providers inside a payments platform. The larger mass sits underwater in payroll budgets, internal teams and embedded contractors who never show up as marketplace merchants. In companies I work with, teams often realize comparable savings inside that submerged portion, where the shift shows up as slower hiring, fewer backfills and shorter project timelines rather than a clean vendor swap. 

That iceberg dynamic matters for national labor statistics. Official measures lag because they track headcount and wages, but substitution often begins as task compression. One analyst with a strong review loop can produce what used to require several contractor hours, and the output ships faster. The company keeps the same roles on paper, yet the demand for marginal labor hours declines across departments.

Research on task exposure explains why the submerged portion grows quickly. OpenAI and the University of Pennsylvania estimate that about 80 percent of the U.S. workforce has at least 10 percent of tasks exposed to large language model capabilities. Exposure rarely means full replacement of a job. It means faster first drafts, faster synthesis and a heavier premium on judgment, context and accountability.

That shift pushes new management work. Procurement teams negotiate usage controls, data handling, and access governance with model vendors, a theme echoed in the Organization for Economic Co-operation and Development’s recent guidance on firms’ AI adoption. Human resources teams redesign entry roles around evaluation, source-checking and customer nuance. Managers rewrite playbooks so that speed gains translate into better service rather than sloppy volume.

A 95 percent unit cost advantage rewrites labor demand across layers of the economy, starting where tasks are modular and buyers already treat labor as on-demand. Platform research reports falling demand for automation-prone categories, including a 21 percent decline in job posts for automation-prone work. That aligns with Stevens’s substitution ratio because both reflect the same buyer decision: purchase acceptable output at a far lower unit price. 

The next layer hits early-career pathways that historically served as training grounds for higher-skill judgment. Stanford researchers using ADP payroll data find early-career workers ages 22 to 25 in highly exposed roles faced a 16 percent relative employment decline after generative AI adoption. When firms can buy first drafts for three cents on the dollar, they reserve remaining work for people with experience. 

Policymakers must understand that the future is in roles that supervise, verify and integrate AI output. Workforce development must emphasize AI literacy as a baseline professional competency.

The labor market already responds to the signal sent by AI’s 97 percent reduction in costs. The visible shifts in marketplace spending represent a small, measurable slice of a larger repricing wave. The iceberg below the surface will determine whether the next decade produces widespread productivity gains with stronger careers, or a hollowed-out middle where entry roles vanish and trust in institutions falls.

Gleb Tsipursky, Ph.D., is CEO of the future-of-work consultancy Disaster Avoidance Experts. He is the author of “The Psychology of Generative AI Adoption” and Returning to the Office and Leading Hybrid and Remote Teams.

Copyright 2026 Nexstar Media Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.

More Opinions - Technology News

Senate GOP says Trump signals shift to backing DHS compromise with Democrats

GOP cracks in Senate begin to show in DHS shutdown fight

Van Hollen: Trump is ‘lying’ about talks with Iranians

US issues ‘worldwide caution’ alert to travelers

Democrats release details behind spending on Noem’s $220M ad campaign

Republicans optimistic on path to end DHS shutdown

The Movement: Kent, Carlson, Greene create Shadow MAGA faction defying Trump

Iran war takes mounting toll on America’s military

These are the best places to live in the US for 2026, study finds

Bannon says ICE at airports ‘test run’ for 2026 elections

Homan: ICE officers will not assist with airport security scanning amid TSA ...

USS Gerald R. Ford aircraft carrier leaves Middle East amid Iran fight

Democrats’ quest for relatable white dudes finds new candidates

NTSB chair: LaGuardia crash investigator was stuck in TSA line for hours

Key things to watch for in Florida special election on Trump’s turf 

Pentagon closes office space for journalists after judge’s ruling on building ...

Manchin urges Democrats to ‘stand up’ to Senate GOP amid push to nix ...

Netanyahu says Trump sees opportunity for deal with Iran

The Hill Podcasts – Morning Report

© The Hill