Deloitte finds 74% of CFOs would need days or weeks to pause AI use that ran over budget. The fix: manage AI by cost per completed workflow, with a pause control built in.
On October 21, Deloitte’s CFO Signals survey for the third quarter of 2026 put numbers on something many finance teams suspected. Only 20% of CFOs track AI costs by token or workload, and 53% review AI usage costs after the fact rather than in real time. Seventy-four percent said they would need “several business days to several weeks” to pause AI use that went over budget. As TechTarget reported, 80% rely on less advanced controls: 56% on basic AI-specific controls and 24% on general IT controls alone.
The same week, Gartner predicted that “by 2029, 60% of organizations deploying AI will establish a dedicated function responsible for mapping AI total cost to value.” It also found that only 4% of 1,010 non-technology S&P 1200 companies reported AI cost reductions material enough to disclose in the first half of 2026. And on October 23, a major model provider made spending caps and pooled budgets generally available for enterprise plans, because agent work consumes more tokens than chat.
Spend is tracked in the wrong unit
Read the Deloitte numbers fairly. They measure controls, not whether AI is paying off, so they do not prove that AI budgets are out of control. What they do show is that most finance chiefs see AI cost the way they see a utility bill: after it arrives, in aggregate, with no clear link to the work it bought.
That is a structural problem, not a discipline problem. Seats, tokens and cloud invoices are how vendors bill, and none of them tells a CFO what one reconciled invoice or one closed exception cost, or what it returned. The new vendor caps help, but they limit spend per person or group. Capping an analyst’s assistant is not the same as stopping one workflow that has started looping on bad data.
Our position: manage AI as a cost per completed workflow, measured against net business benefit, with a pause control at the workflow level. If you cannot stop a single workflow in minutes, you do not control its spend. With Gartner’s survey of 2,002 CIOs finding that 51% expect AI to raise total cost of ownership and 58% face more pressure to deliver cost improvements, this belongs in the budgets being written now.
What this looks like in invoice reconciliation
Take invoice reconciliation in accounts payable. An agent receives an invoice, finds the purchase order and goods receipt, checks quantities and prices against tolerances, and either posts the match or raises an exception for a person to resolve.
Most of that work does not need a language model. Looking up the purchase order, comparing amounts, applying tolerance rules and posting the entry are deterministic steps that conventional software does faster, cheaper and the same way every time. The model earns its cost in a narrower set of steps: reading an invoice in an unfamiliar layout, interpreting a free-text note about a partial shipment, or drafting the explanation a reviewer needs.
Built that way, model spend follows the exceptions, not the volume. Built the other way, with an agent reasoning through every match, the token bill grows with every invoice, and finance learns about it a month later.
Four numbers to track for every workflow
- Cost per completed workflow. Model calls, compute, integration and the minutes of human review, divided by the items that reached a correct, finished state. Count completed work, not attempts.
- Net business benefit. Hours returned to the team, a faster close, fewer duplicate payments or late fees, set against the full cost above. Agree how it is measured before the build starts.
- Share of steps that call a model. If that share rises while the work stays the same, something has drifted: a prompt change, a new data source or a loop.
- Exception and rework rate. A cheap run that sends a large share of items back to people is not cheap.
Tracked daily at this level, AI cost stops being a quarterly surprise and becomes an operating number the process owner can manage.
Design the pause before go-live
A pause control is a finance control as much as a technical one. Gartner’s CIO survey found that only 31% have rules for stopping pilots and 17% have exit rules for projects already in production. Decide these four points before the workflow goes live, not after the first overrun:
- The trigger. A spending threshold per workflow per day, plus anomaly signals such as a spike in model calls per item.
- The owner. A named person, usually the process owner, who can pause the workflow without filing a ticket.
- Work in flight. Open items route to a human queue or wait in place; nothing is left half-posted.
- The restart. Who reviews the cause, and what evidence they need before turning the workflow back on.
The same design answers a security question. Gartner expects that “by 2030, 80 percent of organizations with public-facing AI will have experienced a cost exhaustion attack,” and a workflow-level stop limits that damage too. We covered the wider set of production controls in Reliable AI needs outstanding governance.
How RapidCanvas approaches it
We start from the outcome and the economics together. Before anything is built, the business owner, our field AI engineers and finance agree on the expected value and success criteria, so cost per completed workflow has a benefit to be measured against.
In the Development Studio, which builds on reusable skills, calculations, validations and rules are built as conventional software, and the model is called only for steps that need interpretation. The Enterprise Context Engine™ selects the context each step needs: relevant context, not simply more context. The Operational Harness adds permissions, human approval for critical actions, monitoring, audit trails, versioning and rollback, which is where workflow-level thresholds and pause decisions belong.
On the commercial side, when RapidCanvas provides the model, token consumption is included in the annual subscription. When you bring your own model, we optimize token consumption and give you visibility into workflow costs.
The question to ask before the budget closes
Ask your team one question about every AI workflow in production: what did the last completed item cost, what did it return, and who can stop it in the next ten minutes? If the answer takes days, start there.
Book an AI workshop to map cost per completed workflow for one of your finance processes.
