Ask a CFO what an agentic system costs and they will reach for the model invoice. It is the wrong line. Our system ran 44,000+ runs in 63 days on under $50 of model spend. The tokens were rounding error. The money went somewhere else.
Every brand system has four cost lines. Model spend. Integration. Review time. Memory upkeep. Only the first one shows up on a vendor quote. The other three decide whether the system pays for itself or quietly consumes the team's week.
Integration is paid once. Or forever.
Connecting the system to your brand's sources is real work. Brand guidelines, product data, past campaigns, the channels that publish. Done well, it is paid once. Done badly, it is paid every time a source changes shape. The difference is the connector design. We use read-only connectors with a defined scope, so a schema change breaks one agent's input, not your live data. Model-agnostic routing means a model swap does not trigger a rebuild. Integration stays a one-time line rather than a monthly one. The architecture is laid out on /system.
Review time is the real bill.
Every output that reaches a human costs a minute of a senior person's attention. Multiply by output volume and this line dwarfs everything else. Throughput that goes 3–5× while review capacity stays flat does not create leverage. It creates a queue. The approval gate has to be designed, not assumed.
Three mechanisms keep it cheap. First, agents grade their own work against enforceable brand rules before a human sees it, so the reviewer sees fewer, better drafts. Second, low-stakes work routes through a lighter gate than high-stakes work. A social caption and a press statement should not cost the same minute. Third, audit-grade logs mean a reviewer reads the reasoning rather than reconstructing it. Review moves from re-doing to signing off. That is where the economics turn.
Memory upkeep is the line nobody budgets.
A brand system is only as good as what it remembers. Voice rules, positioning decisions, what was rejected last quarter and why. This memory decays if nobody owns it. The efficient version folds upkeep into the approval gate itself. Each approval or rejection writes back to the rule set. The system learns the correction once and does not repeat the error. Upkeep stops being a separate task and becomes a by-product of review. The 90-day path on /journey sequences this deliberately: rules first, then volume.
What this means for the buying decision.
When you evaluate a system, ignore the model rate card. Ask four questions instead. How many sources does integration touch, and what happens when one changes? What fraction of outputs reach a human, and how is that fraction expected to fall? Where do corrections go, and does the system stop repeating them? What does the exit cost if any of those answers disappoint? Our answers are on /posture: read-only access, human approval on anything that ships, logs you can replay, and no lock-in on models or data.
The model bill will keep falling. It already rounds to zero at our scale. The three other lines are where a brand system earns its keep or loses it. Design for them and the operating leverage is real. Ignore them and you have bought a cheap engine attached to an expensive queue.
If you want the four-line cost stack drawn against your own team, book the strategy blueprint call.
