Data Engineering · 6 min read
Data and AI costs: measure the value, not just the bill
A practical guide to data and AI unit economics: measure accepted outcomes, review effort, quality, and operating costs together.
Updated
A lower technology bill is useful only when the business still gets the outcome it needs. Refreshing a report less often can reduce warehouse cost while delaying a replenishment decision. A cheaper language model can lower the price of a request while increasing the time people spend correcting its output. Cost and value need to be reviewed together.
This is becoming a more visible leadership question as data platforms, AI services, and software subscriptions overlap. The FinOps Foundation's State of FinOps 2026 survey identifies managing AI spending as a leading forward-looking priority among its respondents. That is a signal from the FinOps community, not a census of all businesses. It supports a practical question: can the organization connect technology spending to useful work?
The answer does not require a large finance program before the first pilot. It starts with a clear cost boundary, a meaningful unit of work, and an owner who can explain the tradeoffs.
Choose a unit the business recognizes
Infrastructure measures such as storage, compute time, and tokens explain consumption. Business measures explain why the consumption matters. Depending on the workflow, useful units might be a reconciled reporting cycle, an accepted case summary, a completed forecast review, or an authorized customer-service interaction.
Define "completed" carefully. A generated draft is not necessarily an accepted result. A report refreshed successfully may still be unusable because source data is incomplete. Exclude neither failures nor retries from the cost numerator simply because they did not produce a useful output. Otherwise, the apparent unit cost improves by hiding the work that went wrong.
Use a simple formula: cost per accepted outcome = relevant period cost / accepted outcomes in that period. The formula is straightforward; agreeing on the scope is the important work. Record the quality conditions that make an outcome acceptable.
Put the whole workflow inside the boundary
For a data workflow, include ingestion, storage, transformation, querying, orchestration, monitoring, and the applicable support effort. For AI, consider model calls, retrieval, embeddings, tool calls, evaluation, logging, and human review. Some costs are shared or fixed rather than directly metered. Show those separately before deciding how to allocate them.
Do not hide major exclusions. A pilot may report variable service cost while excluding implementation and existing staff time. That can be a useful early measure when labeled clearly, but it is not the same as total cost of ownership. Show both views when deciding whether to expand.
The FinOps unit-economics guidance provides a useful framework for linking technology costs with business activity. The FOCUS specification can help normalize cost and usage records from supported providers. Neither chooses the right business denominator or proves a return on investment; those remain organizational decisions.
A worked example of a misleading saving
Consider an illustrative monthly case-summary workflow. Option A costs $600 in metered services and produces 1,000 drafts. Reviewers accept 800 after spending twenty hours checking and correcting them. At an assumed $50 per review hour, the combined metered and review cost is $1,600, or $2 per accepted summary. Shared platform and development costs are excluded in this example.
Option B reduces the service bill to $400 but requires thirty review hours, with only 700 accepted summaries. Its comparable cost is $1,900, or about $2.71 per accepted summary. The service bill fell, but the cost of useful output rose. The quality difference may also make the options unsuitable for a simple price comparison.
These figures are not market prices or expected results. They illustrate why procurement and technical teams need a shared definition of value. Staff time released is capacity, not automatically cash saved. The business case should say whether that capacity improves service, increases coverage, or avoids additional hiring under a realistic plan.
Separate efficiency from demand growth
A rising bill can mean waste, growth, improved adoption, or a more demanding service level. Review volume and unit cost together. If accepted workload doubles while unit cost declines, total spending may rise for a good reason. If spending rises while usage and quality remain flat, investigate the difference.
Break the variance into understandable drivers: volume, price, workload mix, retries, refresh frequency, and changes in service requirements. For AI, more complex cases or longer context can increase consumption. For analytics, an expensive repeated query may be a modeling issue rather than a user behavior issue.
Avoid comparing unlike populations. A customer-facing workflow with extensive checks should not be judged against an internal draft assistant using a much narrower control boundary. Segment the cost view by meaningful workload and risk differences before setting targets.
Design controls before the bill arrives
Cost controls are most useful in the workflow itself. Set sensible limits on query size, repeated attempts, agent steps, and external tool calls. Monitor unusual consumption and make the escalation owner clear. A budget alert is a notification, not necessarily a hard spending cap; confirm the actual behavior of each service.
Use the lowest-cost design that meets the required quality and reliability. That may mean a rules-based check before a model call, a cached result with an acceptable freshness window, or an incremental transformation instead of rebuilding unchanged history. Test each change against the same acceptance cases.
Caching and reduced refresh frequency have tradeoffs. Cached answers must respect permissions and data changes. Less frequent updates must still arrive before the business decision. Never remove security checks or necessary evidence to produce an attractive cost reduction. The optimization should preserve the agreed service, not quietly redefine it.
Make shared costs visible and fair
A warehouse or AI platform can serve several teams. Assign directly attributable usage where possible, then explain how shared costs are handled. Allocation by query volume, active users, or business activity may each be reasonable in a particular context, but none is universally fair. Keep an unallocated category visible while attribution improves.
Initially, a showback report that informs owners may be more useful than charging departments for every request. The purpose is to improve decisions, not create arguments over small amounts. Review material exceptions and give teams enough detail to change the behavior causing them.
Include ownership in new work. A pipeline, dashboard, or agent without a business owner can keep consuming resources after its original purpose has ended. Retirement is an engineering and business decision: confirm dependencies and retention needs before switching it off.
Use a compact monthly review
A useful review brings together accepted volume, total scoped cost, unit cost, quality, timeliness, and the largest changes. Include the current forecast and its assumptions. An executive should be able to distinguish healthy adoption from an uncontrolled retry loop without reading infrastructure logs.
For each significant issue, record the decision, expected effect, owner, and review date. Then check whether the change actually improved the relevant measure. A theoretical saving in a design document is not realized value until the workload behaves as expected.
Start with one recurring reporting process or one AI pilot. Establish a baseline before optimizing, and use the same definitions afterward. This turns cost management into a practical part of delivery rather than a separate exercise after launch.
VeriVal helps teams connect data architecture, governed metrics, and AI workflow evaluation to business priorities. If the bill is visible but the value is difficult to explain, start a conversation.