Cents versus dollars
A simple agentic support task is estimated at roughly $0.004 to $0.10 in token cost. Five minutes of customer service or IT support labour is estimated at $2.15 to $3.02 after the paper's 25% overhead assumption.
Project 04 | Independent research
The Economics of Agentic AI
Token Cost, Human Labour, and the Future of Work
A scenario-based study of when agentic AI remains cheaper than human labour after accounting for model calls, tools, retries, supervision, error correction, infrastructure, privacy, and compliance.
Read the full paper- Author
- Nathan Gomes
- Published
- May 24, 2026
- Scope
- 22 pages | 11 tables
- Method
- Scenario-based cost model
Key findings
Figures below are scenario estimates from the paper, based on published model pricing and labour wage data.
Tools remain economical at light usage
One modeled premium task with 8,000 input tokens, 2,000 output tokens, one web search, and one file search totals about $0.1125 before human review.
Review cost changes the equation
A complex workflow using 100,000 input tokens, 30,000 output tokens, five searches, and two minutes of IT review is estimated at $2.66. Heavy review, retries, liability, or correction can narrow or eliminate the advantage.
Tasks change before jobs disappear
Replacement risk is highest for repetitive, digital, low-risk work that is easy to verify. The paper argues that human judgment, accountability, trust, and oversight remain central in complex and high-stakes work.
Cost model
The comparison includes more than tokens
Total AI cost = input tokens + output tokens + tool calls + retries + human review + error correction.
Total human cost = time spent on task multiplied by hourly wage + estimated overhead.
This framing explains why a cheap model call can still become an expensive business workflow, and why model routing, caching, usage limits, monitoring, and targeted approval are economic controls rather than only engineering details.
Long-term pressure
Unit costs can fall while total spending rises
The paper cites a reported 280-fold decline in GPT-3.5-level inference cost from late 2022 to late 2024, alongside projected data-centre electricity growth from 415 TWh in 2024 to about 945 TWh by 2030.
Central conclusion
The answer is conditional
Agentic AI is most economical for repetitive, high-volume, low-risk digital tasks. Humans remain necessary where work depends on judgment, accountability, physical presence, emotional intelligence, legal responsibility, or difficult decisions.
Original paper