The Cost Curve No One Tells You About

TL;DR

The agent cost curve has three phases:

1. Pilot phase — costs are dominated by people (engineers building, supervisors testing). Per-transaction cost looks high.
2. Scale phase — per-transaction cost drops 5–20x as volume grows and people cost amortizes. Most companies see this and project the curve continuing.
3. Saturation phase — per-transaction cost flattens; supervision, governance, and second-order costs (re-evaluation, drift management, audit) become the dominant cost component.

The strategic mistake: budgeting for pilot or scale economics, not saturation. The cost curve flattens earlier than CFOs expect.

---

The agent cost curve has three phases. Most companies budget for the second and get surprised by the third.

The “AI is getting cheaper” narrative is true at the per-token level and partially true at the per-transaction level. It’s not true at the program-level total cost level once supervision, governance, and second-order costs are included. This piece is the cost curve dynamics CFOs and CTOs need to plan against.

The three phases

Phase 1: Pilot

What it looks like: a team of 3–5 engineers builds the first agent. A few hundred to few thousand transactions per week. Per-transaction cost includes the engineers’ time amortized across small volumes — looks like $50–$500 per transaction.

What CFOs see: pilot looks expensive on a per-transaction basis. Conclusion: scale will fix this.

What’s missing: the people cost is a fixed cost; it amortizes well as volume grows. The pilot phase isn’t representative of steady-state.

Duration: typically 3–6 months per agent.

Phase 2: Scale

What it looks like: agent in production at meaningful volume (10K–1M+ transactions per week). Per-transaction cost drops dramatically. Looks like $0.50–$5 per transaction at typical scale.

What CFOs see: the cost curve looks favorable. Conclusion: keep adding agents and keep growing volume.

What’s missing: the per-transaction cost is dropping because of volume amortization, but the second-order costs (governance, supervision, re-evaluation) haven’t fully kicked in yet. The visible cost is the model + infrastructure + amortized people; the invisible cost is governance, IR, drift management.

Duration: 6–18 months per agent.

Phase 3: Saturation

What it looks like: agent at steady-state. Volume growth flattens. Now: governance, supervision, re-evaluation, drift management, audit, regulatory work, vendor management — these become the dominant cost components.

What CFOs see: the per-transaction cost decline stops. Sometimes reverses. Conclusion: AI economics weren’t what was projected.

What’s actually happening: the second-order cost structure is fully visible now. It’s substantial — typically 30–60% of the program’s total cost.

This is where most enterprises are by 2026, and where the cost-curve disappointment originates.

What gets under-budgeted

Five second-order cost components.

1. Supervision. Per The Cost Economics of Autonomous Agents at Scale, 10–25% of a senior operator per agent in production. For a portfolio of 20 agents: 2–5 FTEs.

2. Governance. AI program lead, governance committee time, policy and audit work. For mid-large enterprise: 3–8 FTEs.

3. Re-evaluation. As models change, re-evaluating each agent against the new model. Quarterly to annual depending on stability. Each re-eval: ~1 month of engineering time per agent.

4. Drift management. Eval set maintenance, drift monitoring, periodic recalibration. Continuous; ~5–15% of an engineer’s time per agent.

5. Audit, regulatory, legal. As scale grows, the regulatory surface grows. Compliance work, audit cooperation, customer security questionnaires. Typically 0.5–2 FTEs in dedicated AI compliance work for mid-large enterprise.

These add up to a meaningful portion of the program’s total cost. Programs that budgeted only for licensing + infrastructure + initial people get surprised.

The flat curve

Across the three phases, here’s what the program’s per-transaction cost typically looks like (illustrative):

Phase	Per-transaction cost	Time
Pilot	$50–$500	Months 1–6
Scale	$0.50–$5	Months 6–18
Saturation	$1–$10	Month 18+

The saturation cost is sometimes higher than the scale cost because the second-order costs catch up. Programs that planned for “scale costs continuing forever” get surprised.

What to plan for

Five planning principles.

1. Budget the saturation cost, not the scale cost. Use saturation-phase economics for any agent likely to be in production for 2+ years.

2. Plan headcount for the second-order work. Supervision, governance, re-evaluation. Scale linearly with portfolio size.

3. Anticipate model-change cycles. Major model changes (every 6–18 months) trigger re-evaluation work; budget for it.

4. Build the regulatory work into base cost. Don’t treat compliance as a one-time project; it’s an ongoing operating cost.

5. Reassess unit economics annually. What looked profitable in scale phase might not pencil in saturation. Be willing to retire agents whose unit economics don’t survive.

Counter: aren’t models getting cheaper?

Yes — model and inference cost continues to drop ~50% / year. But program total cost depends on more than model cost, and the components that matter most in saturation aren’t model cost. The per-token cost may continue dropping while the per-transaction cost flattens because supervision and governance scale differently from inference.

What to do this quarter

1. Map your AI portfolio by phase. Which agents are in pilot, scale, saturation?
2. Calculate steady-state total cost for the saturation-phase agents. Include all five second-order components.
3. Update the AI budget for next year using saturation economics, not scale economics.
4. Identify agents whose unit economics won’t survive saturation. Plan for retirement or restructuring.

FAQ

Does the cost curve flatten for all use cases? Mostly yes, but at different absolute levels. Internal-only agents typically have lower governance costs than customer-facing or regulated agents. The curve shape is similar; the saturation cost differs.

What about open-source / self-hosted models? Self-hosted lowers the model cost; doesn’t affect supervision, governance, or compliance cost. Saturation economics are similar; the floor is slightly lower.

Will continued model improvements reduce supervision needs? Marginally over time. Better models have fewer failures, which reduces supervision intensity. But governance, audit, regulatory work doesn’t decrease much; that’s process cost, not model cost.

How does this affect build vs. buy? The per-transaction cost in saturation is similar for built and bought agents. The difference: buy lowers the build cost in pilot/scale phases; build can produce lower marginal cost in saturation if you can amortize the platform.

What’s a sustainable per-transaction cost target? Depends entirely on use case. Customer support: $0.50–$2 per transaction. Specialized domain agents: $5–$20. Sales/marketing tools: varies. Set a target for the specific use case based on the value created, not on a generic “AI cost” number.

---

Working with JAIN on AI cost economics? We help executive teams plan for saturation-phase costs, not just pilot or scale costs. Book a 30-minute call.

Related reading:

• AI Strategy for CTOs and CEOs
• The AI ROI Question
• The Cost Economics of Autonomous Agents at Scale
• The Build vs Buy Decision for AI