Short answer: Enterprise AI agent programs overspend because leaders optimize the wrong cost line. Token spend gets the attention, but in real deployments, human oversight accounts for 70 to 75 percent of variable run costs, not the model itself, according to a McKinsey QuantumBlack analysis published in August 2026. The fix isn’t a cheaper model. It’s reducing exception rates and building governance directly into how agents execute, not bolting it on after the fact.
The cost problem nobody budgeted for
As companies rush to put AI agents into production, they’re discovering that the agent economics are harder to predict than the pilot suggested. Token prices keep falling, yet enterprise AI spend keeps climbing. McKinsey’s analysis puts real numbers on why: customer-facing AI agents at some banks cost $20,000 to $30,000 to run as a single-agent workflow, and $100,000 to $200,000 for a multiagent team, before multiplying by the number of runs a team makes across every workflow that gets automated.
Natural instinct when costs run high is to optimize the visible expense: pick a cheaper model, tighten the prompt, cut tokens. McKinsey’s data suggests that’s the wrong lever entirely. In banking customer service workflows, token costs represent just 20 to 25 percent of variable run costs. Human oversight, the risk and functional experts reviewing agent output before it reaches production, accounts for 70 to 75 percent. In customer onboarding specifically, that means 10 to 20 percent of every agentic run still needs a human review pass.
That reframes the whole cost conversation. Reducing token spend on a workflow where oversight dominates the cost structure barely moves the total bill. Redesigning the workflow to reduce exception rates and simplify what needs human review does.
The workflow is more complex than the demo
There’s a second finding worth sitting with. A single “AI agent” completing a real job, like onboarding a new bank customer, actually requires five to seven specialized agents, three or more deterministic rules engines, and two to four teams of humans providing oversight and risk review. That’s not a failure state. That’s what a fully loaded, production-grade agentic workflow looks like once you account for document intake, KYC checks, routing, exception triage, and output validation.
This matters because it means governance can’t be an afterthought layered onto one agent’s output. If five to seven agents are coordinating a single business process, each handoff, each tool call, each data retrieval is a point where something can go wrong, and a point where oversight cost accumulates. McKinsey’s own framing is direct: the metric that actually matters is the fully-loaded cost to finish the job, not the cost of any individual agent running in isolation.
Where the oversight cost actually comes from, and where it can be reduced
Two things drive down that oversight burden without cutting corners on risk.
Grounding reduces exceptions. An agent reasoning from incomplete or outdated operational context produces more exceptions, more escalations, and more work for the humans reviewing its output. This is exactly the problem the meshIQ AI Agent is built to solve inside enterprise middleware environments. Rather than reasoning from generic model knowledge, it queries live operational telemetry, queue depths, consumer lag, channel status, and configuration state across IBM MQ, Apache Kafka, RabbitMQ, ActiveMQ, Solace, TIBCO, and IBM App Connect, and correlates that evidence before recommending a resolution. When an agent’s diagnosis is grounded in what’s actually happening in the environment rather than a best guess, there’s less for a human reviewer to second-guess, and fewer runs that need escalation in the first place.
Governance at the point of execution reduces risk without slowing throughput. The other lever is what happens when an agent decides to act, not just what it recommends. Most of today’s AI tooling, LLM gateways, observability platforms, and output guardrails sit outside the execution path. They can trace what an agent did after the fact, but they can’t intercept a tool call before it runs, pause an action for approval, or redact sensitive data before it reaches a model’s context window. By the time an event is observed, the action has already happened.
This is the architectural gap AgentGuard is designed to close. It’s an in-flow governance control plane, embedded directly into agent execution through a lightweight SDK, that evaluates every tool call against organizational policy before it runs. Each action resolves to one of four outcomes: allow it immediately for low-risk operations, require explicit confirmation for medium-risk actions, escalate high-risk actions to a designated approver while execution state is preserved, or deny it outright if it violates policy. Every prompt, tool invocation, and governance decision becomes part of a tamper-evident audit trail, each event incorporating the previous event’s SHA-256 hash, producing evidence that can be independently verified and mapped to frameworks like SOC 2, the EU AI Act, HIPAA, and GDPR.
That distinction matters for agent economics specifically. Oversight is expensive; not because review itself is inherently costly, but because today’s tooling makes every review a manual, after-the-fact investigation. Governance that operates in line with policy decisions and audit evidence generated automatically as part of execution is what turns “70 to 75 percent of variable cost” into a number that can actually shrink over time.
Reuse is the other lever McKinsey points to, and it applies to migration too
McKinsey’s analysis also found that agents of AI economics favor two characteristics: high volume, since fixed costs amortize as runs increase, and reuse, building an agent once and applying it across multiple high-value workflows rather than building bespoke agents for each one.
That principle extends past day-to-day operations into modernization work itself. The meshIQ AI Agent Factory applies the same discipline to middleware migrations, using a coordinated set of specialized agents, an Orchestrator, Surveyor, Analyst, Planner, Builders, and Inspector, to discover, plan, transform, and independently validate changes at scale, while humans retain explicit control over every irreversible production decision. Modernizing one application is a project. Modernizing hundreds of applications one at a time doesn’t scale economically, which is exactly the volume-and-reuse problem McKinsey’s framework describes.
The takeaway for anyone budgeting an agentic AI program
Don’t start the ROI conversation with model selection. Start with two questions: how much of this workflow’s cost sits in human oversight rather than tokens, and does the governance model catch problems before an action executes or only after? Programs that get grounded operational context and inline governance right from the start are the ones positioned to actually capture the economics McKinsey describes, rather than losing control of costs the way so many early agentic deployments already have.
Sources: Where AI agents pay off: A practical guide to the economics of agentic workflows, QuantumBlack, AI by McKinsey, August 2026; meshIQ AgentGuard and meshIQ AI Agent factsheets, 2026.