Showback

Showback is the practice of reporting infrastructure and software costs back to the teams that generate them, without actually transferring budget. Each group sees what its services, environments and workloads cost, broken down in terms it recognises, and the finance ledger stays exactly where it was. It is the reporting half of cost accountability, separated from the accounting half. The reason it works is visibility rather than incentive. Most wasteful spending is not a decision anyone made; it is an oversized instance nobody resized, an environment nobody switched off, a log retention setting nobody revisited, or a model chosen during a prototype and never reconsidered. None of that survives being displayed next to a team name every month, and almost none of it required a budget transfer to fix. Showback is usually the right first step because it is reversible and cheap. Chargeback, where the cost genuinely moves onto a team budget, produces stronger incentives and considerably more argument: about the fairness of allocation, about shared costs, about whether a platform team is a service or a tax. Running showback first surfaces those disputes while the stakes are still low, and it also reveals whether your cost data is accurate enough to survive being used for something consequential. The quality of the report determines whether anyone acts on it. Costs need to be attributed by tags or accounts that map to how the organisation is actually structured, presented in units engineers recognise — per service, per environment, per customer, per feature — and delivered often enough to connect to a change someone made. A monthly PDF arriving three weeks after the period ends teaches nobody anything, because the change that caused the number is long forgotten. AI workloads make showback more valuable and harder to do. Cost per request varies with prompt length, retrieval size, model choice and how many tool calls an agent decides to make, so two features with similar traffic can differ enormously in spend. Attributing that requires the application to record the cost of each call along with what it was for — feature, tenant, workflow — because the provider invoice knows only tokens, not purposes. The metric to build towards is not total spend but cost per unit of value: per resolved ticket, per generated document, per active customer. Total spend rising is not a problem if usage is rising faster; total spend flat is not a success if the work being done has quietly halved. Showback that reports only absolute numbers tends to produce indiscriminate cuts, while showback expressed in unit terms produces engineering.