The economics of agentic systems
Stop pricing agents per token. The right unit is per outcome — and the cost curve looks completely different.
Why finance teams are anxious about AI costs
Finance and platform teams are watching their AI bills climb and asking the obvious question: 'is this worth it?' The honest answer is that the question is almost always asked at the wrong unit of analysis. Per-token costs feel large because tokens are visible. Outcome costs — the actual cost of completing a customer onboarding, processing a claim, closing a ticket — are usually invisible, and they're the only number that matters.
Per-token vs. per-outcome vs. per-customer
Mixing these up leads to bad decisions in both directions.
Per-token cost
What the model provider charges. Visible, granular, and the wrong frame for enterprise decisions. Optimizing per-token alone often makes per-outcome costs worse.
Per-outcome cost
The total cost to complete a unit of work — including model calls, tool calls, downstream API costs, retries and human approvals. This is the number to manage.
Per-customer cost
Customer-lifetime cost of AI-supported service. Useful for retention and CAC modelling; the right frame for B2C and high-touch B2B.
Where to focus
Not every optimisation is worth the engineering time. Here's our experience with what actually moves the needle.
- Model routing — sending cheap tasks to cheap models. Often a 5-10x cost reduction with no quality loss.
- Tool-call efficiency — agents that retrieve the right context the first time, instead of fishing. Compounds enormously.
- Memory & context reuse — short, well-structured prompts beat long, hopeful ones.
- Self-healing workflows — avoiding the cost of escalating to humans for transient failures.
- Caching & precomputation — pre-grounding common knowledge, caching deterministic responses.
Agentic vs. RPA vs. status quo
When you compare per-outcome costs, agentic systems tend to be cheaper than RPA at scale (because adapting to UI changes doesn't require redevelopment), and dramatically cheaper than human-only for routine work. The honest comparison is rarely 'agentic vs. nothing' — it's 'agentic vs. the current operating model', and the latter usually includes hidden costs (rework, attrition, exception handling) that don't show up on any invoice.
Build the cost model. Include the hidden costs. Make the comparison honestly. The numbers usually work.
Ready to put autonomous agents to work?
See xyner in your environment with a guided executive demo.