Financial institutions face an unexpected operational strain as spending on the computing tokens that power autonomous systems threatens to devour corporate budgets. According to McKinsey's Global Banking Annual Review, spending on computing tokens could account for as much as 20% of total labour costs. This financial overhang is forcing banks worldwide to reassess automated workloads as token burn escalates across enterprise deployments.
The math behind multi-agent architecture is directly undermining the economics of early automated software rollouts. Instead of executing isolated prompts, interconnected software agents trigger cascading dialogue loops that drain infrastructure allowances, forcing leadership to reconsider autonomy limits to preserve operational margins. What began as an efficiency experiment is now developing into an infrastructure cost center running on raw compute.
The Breakdown of Query Routing
Banks initially attempted to reduce costs by directing employee queries to different AI models based on their complexity. That routing framework quickly encountered architectural dead ends once multi-agent environments went live.
"Spending on the computing tokens that power these agents could account for as much as 20% of total labour costs."
When autonomous agents communicate with one another to resolve workflows, they consume tokens at volumes that manual routing tiers cannot constrain. How are institutions responding to this unchecked consumption?
Financial institutions have established dedicated teams, sometimes called AI FinOps, to monitor and manage technology spending. The creation of these specialized units reflects a transition from indiscriminate enterprise adoption to defensive cost control. Rather than deploying autonomous systems across every department, banks are prioritizing applications that offer measurable productivity improvements or generate additional revenue.
Consumer Trust and the Dollar Limit
Proponents of autonomous architecture argue that agent token costs will be offset quickly by massive operational efficiencies and new income streams, justifying outsized compute outlays. Operational savings are comparatively easy to calculate, and institutions are actively examining whether agents can attract clients, elevate customer service, and unlock fresh revenue streams.
"Among respondents who use AI, 72% said they would allow a banking agent to complete one financial task on their behalf."
Yet the practical ceiling on agent autonomy is already set by consumers themselves. A McKinsey survey covering 46,000 retail banking customers worldwide revealed that while users rely on software to compare financial services, assess savings rates, and oversee debt, their willingness to delegate real control stops short of enterprise expectations. Among survey respondents who use AI, 72% said they would allow a banking agent to complete one financial task on their behalf, but retail banking customers would limit the amount the system could approve independently to USD 500.
McKinsey expects AI agents to become widely used within the next three years, outpacing the adoption curve of mobile internet, which took between five and 10 years to reach widespread use. Yet the balance sheet implications run counter to corporate expectations: agents could erode profitability by systematically moving deposits into higher-yielding competitor accounts. For now, that balance-sheet erosion remains capped simply because most banks prevent external AI agents from independently authorizing transactions within their systems.
What doesn’t add up
Where exactly does the unit economics of autonomous banking systems break down? The answer lies in the friction between cascading agent-to-agent dialogue loops and strict limits on client trust. While by late October major banks were already blocking autonomous agents from directly authorizing transactions, internal multi-agent environments proved equally capable of straining operational margins.
Our stories: Major Banks Block Autonomous AI Agents Over Security Risks (October 22).