Chapter Five

Shadow AI, Governance, and the Risk of Cost Explosion

^{Shadow AI, Governance, and the Risk of Cost Explosion}

Shadow AI is best understood as a fact of life, not an exception to be solved once and for all. It appears when employees can get better results faster with consumer-grade tools than with official channels.

It also appears when policies are ambiguous, sanctioned tools are hard to access, and experimentation feels safer outside formal oversight.

KPMG describes shadow AI as a symptom of deeper friction between employee needs and enterprise-provided AI capabilities.

Its analysis points to outdated tools, poor integration into daily workflows, rigid provisioning and overreaching controls as causes of unmanaged adoption. [4]

From Shadow to Structured AI

A mature response does not rely on prohibition.

It creates structured freedom: clear zones where experimentation is encouraged, production pathways where risk is reviewed, and telemetry that allows leaders to see usage before it becomes unmanaged scale.

Sandbox Zone: Employees can experiment with approved tools and synthetic or low-risk data, with training and visible guardrails
Pilot Zone: Use cases are evaluated for data sensitivity, cost behaviour, control requirements, user impact and potential production value
Production Zone: Deployment requires named ownership, security/risk review, audit logging, cost allocation, evaluation criteria and incident response paths
Retirement Zone: Low-value tools, duplicate agents and obsolete automations are removed so the AI estate does not accumulate hidden operational debt.

Each zone should be overlaid with a governance layer to prevent prompt duplication while encouraging reuse of prompts, agents, sub-agents, and skills.

Therefore, maintaining data integrity, preventing data leakage and, more importantly, controlling cost.

^{The Cost Explosion Risk}

AI cost is hard to manage because it arrives through multiple channels: cloud infrastructure, GPUs, vector databases, data pipelines, model APIs, SaaS copilots, agent orchestration tools, security monitoring, observability, testing, and human review.

The unit of consumption may be opaque - tokens, queries, seats, workflow runs, document pages, embeddings, API calls, or bundled platform credits.

The FinOps Foundation reports that AI is expanding the scope of technology cost management and that practitioners are prioritizing capabilities such as allocation, data ingestion, reporting, anomaly detection, planning and forecasting to understand AI spending.

IDC’s infrastructure tracker further reinforces that AI capacity is becoming a structural cost category, with worldwide AI infrastructure spending reaching $89.9 billion in Q4 2025, up 62% year over year, and a forecast to exceed $1 trillion by 2029. [2][5]

_{One of the biggest surprises in AI economics is that model costs are often not the dominant cost category.}
_{Enterprise integration, governance, observability, security and change management frequently outweigh infrastructure expenses.}

Sumita Dutta, Global IT and Digital Head & CIO, Shell Downstream, Customer Operations, Shell

A Worked Example: The True Unit Cost of One Agentic Workflow

Headline token prices make AI look almost free, which is precisely why per-token thinking is dangerous.

Consider a single customer-service triage agent that reads a case, retrieves relevant context, drafts a resolution, and routes anything ambiguous to a human.

The visible cost is the model call. The real cost is the full chain, and it is the chain that scales.

Model Calls: A single resolution is rarely one call. Reasoning, tool selection, a retrieval step and a drafting step might consume four to six calls. A workflow that appears to cost a fraction of a cent per call can quietly cost several cents per completed case
Retrieval and Embeddings: Each case touches a vector database and may re-embed fresh content, adding query cost, storage cost and pipeline cost that never appears on the model bill
Orchestration and Observability: The agent framework, logging, evaluation runs and tracing all consume compute and licensed capacity per execution
Human Review: The escalation path is usually the largest line item by far. If even one case in five reaches a human, the loaded cost of that review can dwarf the entire automated portion of the workflow

The takeaway is not the precise figures, which vary by model, provider and design. It is the shape of the economics: a workflow with a trivial per-call price can carry a material per-completed-task cost once retrieval, orchestration, and human review are included, and that cost scales linearly with volume in a way classic software licensing does not.

Enterprises that price only the visible model call will systematically underestimate AI cost and over-commit before the unit economics are understood.

Every priority use case should be costed end-to-end, per completed business outcome, before it is approved for scale.

The Other Side of the Ledger: The Cost of Not Acting

Discipline is not the same as caution. The risk of overspending on poorly understood AI is real; so is the risk of underspending on readiness while competitors compound small advantages into structural ones.

Both are failures of the same discipline – deciding without unit economics – and a board should weigh them together rather than treating cost control as the only danger.

The discipline the board should demand is not "spend less" but "never commit at scale without a costed return and a way to reverse."

Three Framing Principles Help Keep the Two Sides in Balance

Size the investment as a portfolio, not a single bet: Foundation work (data quality, FinOps tooling, governance plumbing) is a fixed enabling cost that should be justified across many future use cases, while individual agents and workflows are variable bets each carrying their own business case

Treat readiness foundations and use cases differently in governance: It is rational to fund the foundation ahead of proven return because it de-risks everything downstream, but irrational to scale an individual agent on the same faith.

Require a per-use-case return hypothesis before scale: Expressed in the currency the workflow actually moves, whether cycle time, cost-to-serve, error rate, revenue conversion or risk reduction, and fund only those where the modelled return clears the fully-loaded unit cost, not just the visible model call

Controls That Should Exist Before Scaling

Usage-based budgets by product, team, agent and environment
Cost anomaly detection and automated throttling for runaway consumption
Model and tool catalogs with approved use cases, data categories, and owners
Procurement clauses for transparent metering, data processing, audit rights, and exit paths
Environmental impact analysis to measure corporate ESG goals
Business-value tracking that connects AI consumption to cycle time, quality, revenue, risk reduction, or cost avoidance

The Decision Rule

Fund readiness foundations on the strength of optionality; fund individual agents and workflows on the strength of a costed, reversible business case. The board’s question is not "is AI expensive?" but "for this specific use case, do we know the fully loaded unit cost, the expected return and how we would unwind it if we are wrong?"

Executive Risk

The danger is not only that AI is expensive. The danger is committing to AI economics that are poorly understood, difficult to allocate and hard to reverse once embedded in everyday workflows.

^{Organizations will need to analyze not only the monetary costs of cloud vs the cost of AI, but the environmental impact on these decisions and the effects on their ESG commitments.}

^{With respect to AI, they will need to further analyze which LLM model to use for specific business processes, rather than adopting a “one size fits all” model.}

^{LLM’s specialize in different disciplines/uses such as creative, mathematical, software coding, and high-volume data analysis and processing}

Steve Crombie, CIO Bulgaria, ING Bank