Agent Economics

Why Your Single-Agent Demo Costs $9 and Production Costs

A mid-market SaaS company built an AI-powered contract review agent. Legal teams would upload vendor contracts, and the agent would identify non-standard clauses, flag missing provisions, and suggest redline edits against the company's standard terms.

The prototype worked beautifully. The CTO demonstrated it to the board. The CEO authorized a full production deployment.

The architecture team designed it properly — three agents coordinating: a reader agent that parsed contract structure, an analysis agent that compared clauses against the company's legal standards, and a drafting agent that produced redline recommendations. Each agent had its own context, tools, and system prompt. The orchestrator managed handoffs. The output quality was genuinely good.

Then the finance team got the first monthly cloud bill.

Total token consumption per contract averaged 280,000 tokens. At the company's negotiated rate of $8 per million tokens, each contract cost approximately $2.24 in inference alone. Add tool costs: $3.40 per contract. Total: $5.64 per contract. Monthly for 340 contracts: $1,918.

That sounds reasonable. But the finance team wasn't comparing it to zero — they were comparing it to the external legal review service at $75 per contract. The agent won on cost AND speed. The deployment was a success story.

Until the quality audit. A random sample of 50 agent-reviewed contracts found that 7 had missed non-standard indemnification clauses — the kind where a miss could cost $100K+ in a vendor dispute. The miss rate: 14%. The external service's historical miss rate: 3%.

The 14% miss rate on high-consequence clauses didn't mean the agent was bad. It meant the agent needed a different architecture for high-stakes clause types — likely a multi-pass review with an evaluator agent specifically tuned for indemnification, limitation of liability, and termination provisions. That architecture would cost approximately $18 per contract. Still cheaper than $75. But 3x the cost of the simple approach.

The lesson: agent economics isn't just "how much does inference cost?" It's "what does a wrong answer cost?" — and the answer to that second question determines which architecture you can afford.

The Five Cost Centers of a Production Agent

Agent economics is the discipline of modeling the total cost of operating AI agents in production — and comparing that cost against the value delivered and the risk absorbed. It's where every architectural decision in this deep dive meets financial reality.

Think of it like shipping options. Standard ground ($9) gets most packages there in good shape. Express with signature ($25) is warranted for important deliveries. Armored courier ($200) is for things where loss is unacceptable. The package contents determine the shipping tier — not the sender's budget or the courier's capability.

The five cost components. Token spend is the most visible but often the smallest piece of total cost:

1. Inference cost — tokens consumed by the model. Input tokens + output tokens x number of model calls per task. At $3-15 per million tokens, a typical agent task consuming 50,000 tokens costs $0.15-0.75.

2. Tool cost — API calls, database queries, external service charges. A legal database query might cost $0.50. A document processing API might charge $0.10 per page. For an agent making 15 tool calls per task, tool costs can exceed inference costs.

3. Error cost — the component most teams undermodel. A customer service agent that gives incorrect billing information creates a $15-50 support ticket. A contract review agent that misses a clause creates $100K+ exposure. Error cost = error rate x average cost per error x volume.

4. Oversight cost — human time spent reviewing, correcting, and supervising. A reviewer at $50/hour spending 5 minutes per output costs $4.17 per task. Even at Level 5 autonomy, someone monitors.

5. Infrastructure cost — the harness, vector databases, logging, monitoring dashboards, governance layer. A typical enterprise agent infrastructure costs $5K-20K per month before a single task runs.

Where the Bill Shows Up (and Who Sees It First)

The cloud bill surprise. Agent costs are variable in a way traditional software costs are not. A power user who asks complex questions costs 10-50x more than a casual user. This makes cost prediction harder and per-user pricing models more fragile. The PM needs usage-based cost modeling from day one.

The "it replaces humans" fallacy. Agent ROI is commonly justified as "it replaces N humans at $X/hour." This calculation is almost always wrong because agents handle the routine 60-80% while humans handle exceptions, edge cases, and quality oversight. The agent doesn't replace the human — it changes the human's job.

The depreciation curve. Agent architectures depreciate. A harness optimized for one model's cost profile becomes suboptimal when the next generation arrives at half the price. The PM builds economic models that include a 6-month refresh cycle.

Connecting Cost to Reliability and Autonomy

Agent economics reveals the true relationship between the three levels of the Agentic Stack. Level 1 (context engineering) is the cheapest investment with the highest leverage — improving context quality costs engineering time but reduces both inference costs and error costs. Level 2 (harness architecture) is the middle investment. Level 3 (agentic systems) is the most expensive to operate but the most valuable for the right tasks.

The cheapest way to improve agent economics isn't cheaper models — it's better context engineering. Better context reduces retries at every level. Better caching reduces token costs at every level. Better tool definitions reduce error rates at every level.

The connection to reliability is direct. Per-step improvement from 90% to 95% reliability doesn't just improve user experience — it reduces error costs. If each error costs $50 and you process 1,000 tasks daily, the difference between 35% end-to-end success (90%/step) and 60% end-to-end success (95%/step) is 250 fewer errors per day: $12,500 daily savings from a 5-point per-step improvement.

Remember This

1. The 22x range ($9 to $200 per task) is real, and the right architecture depends on what a wrong answer costs. If errors cost $10, the $200 harness is waste. If errors cost $100,000, it's cheap insurance. Match cost tier to consequence tier.

2. Total cost of ownership = inference + tools + (error rate x error cost) + oversight + infrastructure. Token spend is often the smallest component. Error cost frequently dominates.

3. Tiered routing is how sustainable agent economics works at scale: simple harness for routine (70-80% of volume), enhanced harness for moderate (15-20%), full harness for high-stakes (5-10%). The routing between tiers costs pennies and saves millions.