AI forecast failure isn't a modeling problem. It's an attribution problem. Finance teams are forecasting from invoice totals rather than source-level consumption data, so the forecast is always trailing spend by weeks. Until that changes, the number you put up on the board is a guess.
The number in your model is already wrong
Most finance teams building an AI cost forecast start with last month's invoice. They take the total, apply a growth rate, and project forward. The model looks clean. The assumptions look reasonable. The output looks like a forecast. It is not a forecast. It is an extrapolation from a lagging summary of aggregated spend that tells you almost nothing about what the next 30 days will cost.
The evidence for this is not anecdotal. According to Larridin's January 2026 analysis of enterprise AI infrastructure spending, 85% of organizations miss their AI cost forecasts by more than 10%. Eighty percent miss by more than 25%. These are not outliers. This is the baseline.
85% of enterprises miss AI infrastructure forecasts by more than 10%. 80% miss by more than 25%. - Larridin, January 2026
The CFO who produced those forecasts did not make a modeling error. The model was fine. The data feeding it was structurally inadequate for the task.
What CFOs are actually forecasting from
Ask a finance team where their AI cost data comes from, and you will typically hear a version of the same answer: cloud billing summaries, vendor invoices, and seat-based SaaS line items.
Each of these data sources has the same problem: they aggregate spend before the finance team sees it. By the time the number reaches the FP&A model, it has lost the context that would make it useful.
You know the total.
You do not know which team drove it, which workflow consumed it, or which model generated it. A software company with 300 employees and an active AI-powered support chatbot recently ran a spend audit. Visible token costs came to $6,000 per month across three providers. The finance team had low confidence the number was complete. The API calls powering the product's chatbot may have been invoiced differently, or not captured at all. The team could not tell from the billing data alone.
This is not a small company problem. It happens at $500K annual AI spend and it happens at $5M. The data structure is the same.
Why the engineering view doesn't help
Observability platforms exist for a reason. Datadog, and tools like it, give engineering teams real-time visibility into API call volume, latency, error rates, and cost per request. For infrastructure performance management, they are the right tool. They are not the right tool for finance. And the reason is not a dashboard problem.
Observability data is event-level and time-series. It tells you what happened at the API call level in the last hour. Finance data is period-aggregated and dimensionally structured. It tells you what the engineering team spent in Q1, mapped to cost centres, so it can be reported against a budget.
These are different data shapes, built for different consumers, answering different questions. Datadog itself identifies its AI cost monitoring audience as "AI engineers and FinOps personnel." Not finance leadership. The distinction is accurate.
CB Insights reports that fewer than 20% of enterprise leaders currently track AI ROI. The majority track productivity proxies: hours saved, tasks completed, code generated. These are useful operational metrics. They do not belong on a P&L or in a board presentation. When the CFO is asked whether the AI investment is working, a productivity proxy is not an answer.
Fewer than 20% of enterprise leaders track AI ROI. The majority measure productivity proxies only. — CB Insights, Tech Trends 2026 **The attribution gap defined
The specific thing that is missing is spend attributed at the workload level. Workload-level attribution means every unit of AI spend is tagged to: the team that consumed it, the workflow or application that generated it, the model and provider that served it, and the time period in which it occurred. Without this, a CFO has an aggregate number. With it, a CFO has a data set. The difference is not cosmetic. A forecast built from workload-level data can separate predictable from variable workloads, identify which teams are growing spend fastest, and catch model-level cost changes before they appear on an invoice. A forecast built from an invoice total can do none of these things. It is extrapolation from a number that was already out of date when it arrived.
The forecast isn't wrong because your model is bad. It's wrong because the data feeding it stops at the invoice.
What the miss costs
A 25% forecast miss on a $1M AI spend budget is a $250,000 variance. At a $5M budget it is $1.25M. These are not rounding errors. They land as unplanned overruns, emergency reforecasts, and credibility damage in front of the board. The downstream consequences are specific.
Finance teams that cannot forecast AI spend accurately cannot confidently commit to capacity reservation agreements, even when those agreements would significantly reduce per-unit costs. They cannot set meaningful team-level budgets. They cannot produce the gross margin analysis that their investors are increasingly asking for. Deloitte, writing in January 2026, concluded that "traditional total cost of ownership models may need a refresh" as AI tokens become a primary unit of value.
The polite version of the same conclusion: the finance toolkit built for SaaS subscription spend is the wrong instrument for token-based infrastructure.
What changes when attribution exists?
Organizations with workload-level attribution in place operate from a materially different position. They can produce a 30-day rolling forecast that separates predictable, workflow-driven spend from variable, agentic spend. They can set team-level token budgets and receive alerts when those budgets drift. They can make commitment decisions with a utilization forecast behind them, not a gut check.
More directly: when the CFO is asked by the board what the AI spend contributed to gross margin last quarter, the answer is a number, not a narrative. The data layer that makes this possible is attribution at source, pulled from billing APIs before aggregation destroys the workload signal.
This is a data infrastructure problem. It has a solution. The gap between organizations that have solved it and those that haven't is growing every quarter.
REQUEST EARLY ACCESS
Taiken is built for finance teams who need workload-level AI spend attribution, not engineering dashboards. Early access for finance leaders at organizations spending $500K+ annually on AI APIs.
Taiken is building financial governance tooling for enterprise AI spend. If you're a CFO or VP of Finance managing material AI budgets, apply for design partnership.