What Did Six Months of AI Spend Actually Buy?
By Sam M. Sweilem, CEO of LockedIn Labs ·
Connect workflow costs to accepted work and business results, so finance and technology leaders can review what an AI investment delivered.
Six months after an AI investment, the CFO and CIO should be able to sit down with the same evidence and explain what the business bought. A provider bill can show consumption. A team can show adoption. Neither, by itself, tells them which workflows improved, what it cost to deliver the improvement, or where that change reached the profit and loss statement.
The connection has to be designed while the work is happening. Waiting until the review leaves people reconstructing costs from invoices and benefits from anecdotes. I would start with a small number of material workflows and give each one a business owner, a defined outcome, and a cost record that survives changes in models, applications, and teams.
Start with the work the business accepts
An accepted outcome might be a support case resolved without reopening, an invoice processed correctly, or a software change released after review. Define the quality threshold, acceptance event, and observation window before counting results. A generated answer or completed model call is an intermediate step. The business owner should decide what finished work means.
Give that work a persistent identifier. Connect it to the workflow and version, application, accountable team, cost center, run, and individual attempts. Keep acceptance, rejection, escalation, and later rework attached to the same record. Use references to controlled systems rather than copying customer records or sensitive content into telemetry. Those identifiers let engineering explain a run while finance groups the same costs by the relevant business unit.
The FinOps allocation capability describes organizational mappings, metadata, and shared-cost strategies. My proposed extension is to carry those mappings through the work itself, so a cost can be traced from an application event to its accountable team and business purpose.
Include the attempts that did not work
The cost of an accepted outcome includes the failed attempts needed to reach it. Capture model usage, retrieval, tools, runtime infrastructure, retries, review time, and rework. Retain abandoned work in the period’s cost pool. Otherwise, a report can make the operation look cheaper simply by dropping the expensive cases that never finished.
Shared infrastructure, licenses, and engineering support also need an explicit treatment. Separate directly measured costs from allocated costs and document the allocation rule. A usage-based share, a fixed share, and centrally funded capacity answer different management questions. Change the rule deliberately and preserve its version so a six-month trend does not quietly compare different definitions.
OpenTelemetry’s GenAI semantic conventions provide a vocabulary for model operations and usage. They help connect technical measurements; an organization still has to define its business acceptance events and allocation mapping. Check the convention version supported by your instrumentation. Missing usage should remain visible as missing evidence, with any estimate clearly labeled.
Reconcile the operating view with the bill
Usage multiplied by a rate is useful for an early estimate. It can differ from the invoice because of contractual rates, credits, commitments, timing, and usage categories. Keep estimated run costs distinguishable from reconciled charges. Record the pricing period, currency, source, and unresolved variance instead of quietly adjusting every run to make the totals agree.
The working measure is the period’s attributable operating cost divided by accepted outcomes under the same definition and observation window. Show the components alongside it, including review effort and unresolved allocations. The LockedIn Labs token-economics guide develops that operational calculation. The executive review takes the next step: determining whether the resulting work changed a business result.
Compare a baseline that still means something
Preserve the starting workflow’s volume, quality, cycle time, staffing effort, and cost. Compare similar case types and difficulty, and identify changes in demand, staffing, process, or policy. A staged rollout or comparable cohort can help distinguish the intervention from broader changes. Where the comparison is weak, explain the uncertainty rather than assigning every improvement to AI.
Released capacity and realized savings belong in separate columns. Time freed for other work can be valuable, but it reaches a financial result only through a documented consequence: less overtime, reduced external spend, avoided hiring, or additional productive output. Finance should validate the connection to the relevant expense or contribution measure. Counting hours at an assumed salary rate does not establish that an expense disappeared.
Revenue needs the same discipline. Faster responses or higher conversion may accompany an AI rollout, but pricing, sales activity, seasonality, and customer mix may also have changed. Preserve the attribution method and distinguish observed revenue from the portion supported by the evidence. FinOps unit economics supports connecting technology costs to business value and maintaining consistent metric definitions. The review method here is my proposed way to make that connection inspectable.
Make the six-month decision from the record
For each material workflow, bring the original expectation, reconciled cost, accepted volume, quality and rework, baseline comparison, observed business change, and confidence in attribution. Include the share of spend that remains unallocated and the outcomes without acceptance evidence. A complete report can contain uncertainty; concealing it makes the next investment decision harder.
The practical test is whether the CFO can follow the claimed business effect back to its evidence and the CIO can follow the allocated cost back to the systems and work that incurred it. That record should support a specific decision to expand, redesign, continue measuring, or stop a workflow. Six months later, the organization should know what the investment bought and what it still cannot prove.