Agentic AI cost management exposes the fragmentation tax

At a glance

What this is: This is a Kong analysis of agentic AI cost management, arguing that fragmented tooling, untracked consumption, and duplicate spending create a hidden fragmentation tax that erodes margin and blocks monetisation.

Why it matters: For IAM and identity practitioners, the implication is that AI governance now has to include cost attribution, usage visibility, and control over who and what can consume models, APIs, and agentic workflows at scale.

By the numbers:

• 84% of companies report more than 6% gross margin erosion from AI costs.
• Only 15% of companies can forecast AI costs within ±10% accuracy.

Context

Agentic AI cost management is the discipline of tracking, attributing, and governing spend across models, prompts, agents, APIs, data movement, and infrastructure. In this article, Kong argues that the real problem is not AI adoption alone but the fragmentation of consumption across teams, tools, and environments.

That fragmentation creates both margin erosion and monetisation blind spots. For identity and access teams, the governance question is no longer only who can call which model or API, but how usage, privilege, and billing accountability are tied together across the AI stack.

The article frames this as a finance and operating-model issue, not a narrow cloud-cost problem. That is typical of agentic AI programmes as they scale, because the same delegated paths that power automation also multiply unowned consumption if they are not instrumented end to end.

Key questions

Q: What breaks when agentic AI spending is not measured end to end?

A: When agentic AI spend is not measured end to end, organisations lose the ability to attribute cost to customers, features, or teams. That breaks unit economics, weakens forecasting, and allows duplicate tooling or idle infrastructure to persist unnoticed. The result is margin erosion that only becomes visible after it has already shaped product and finance decisions.

Q: Why do agentic AI workloads make cost forecasting so difficult?

A: Agentic workloads are harder to forecast because the execution path is not fixed in advance. One task may trigger retries, tool calls, verification steps, and multiple model calls, each with different cost implications. Traditional forecasting assumes stable unit costs and predictable execution, which agent behaviour breaks at runtime.

Q: How can security teams tell whether AI lifecycle controls are working?

A: They should look for evidence that access requests, policy enforcement, and usage visibility are centrally recorded and current. If those signals are fragmented across platforms, the programme may be documenting governance rather than enforcing it. Continuous traceability is the practical test.

Q: Should organisations treat AI cost visibility as a governance requirement or a finance task?

A: Organisations should treat AI cost visibility as both, because ungoverned consumption becomes a governance failure before it becomes a line item. Without shared visibility, security cannot understand who is using what, and finance cannot understand what the business is buying. The controls need to be joint, not sequential.

Technical breakdown

Why agentic AI costs fragment across tools and teams

Agentic AI cost drift happens when consumption is distributed across LLM calls, retrieval, egress, GPU capacity, and agent-to-agent traffic that different teams measure separately or not at all. In that model, token spend is only one visible slice of a much larger cost path. The operational issue is attribution: if no system ties usage back to a customer, feature, or business unit, finance sees aggregate burn but cannot explain it. That is why fragmented observability turns routine AI experimentation into persistent cost leakage.

Practical implication: Instrument the full AI request path so every cost-bearing dependency is attributable to a business owner or use case.

How AI FinOps differs from traditional cloud cost control

Traditional FinOps assumes relatively deterministic resources such as instances, storage, and reserved capacity. Agentic AI is different because the same request can produce different reasoning paths, tool calls, and downstream spend depending on context and model selection. That makes forecasting weaker and policy enforcement more important at runtime. In practice, AI FinOps has to manage probabilistic consumption, not just static infrastructure, which means metering at the prompt, agent, and data-access layers.

Practical implication: Adopt governance controls that meter and route AI usage at execution time, not only after invoices arrive.

Why monetisation fails when consumption is invisible

You cannot price, package, or bill for AI capabilities that nobody can measure consistently. The article’s core point is that hidden consumption does not just waste money, it blocks revenue models such as usage-based billing, tiered AI features, and customer-specific attribution. That is especially relevant in agentic systems, where multiple automated steps may occur before a user sees value. If those steps are invisible, the business cannot prove unit economics or defend margins.

Practical implication: Build billing and attribution logic into AI services before broad rollout, so product and finance teams can align on unit economics.

NHI Mgmt Group analysis

Fragmentation tax is now an identity governance problem, not just a finance problem. The article shows that AI consumption is spreading across MCP servers, APIs, model calls, and agent workflows that different teams own separately. Once usage becomes distributed that way, accountability also fragments, and no one can reliably answer who consumed what, under which privileges, or for which business purpose. The practitioner implication is that AI governance has to join cost attribution to access governance, not treat them as separate disciplines.

Agentic AI cost control exposes the limits of static entitlement thinking. Traditional controls assume you can define access once and observe its effects later. That assumption weakens when agentic workflows generate variable paths, variable spend, and variable data movement from the same initial grant. The governance challenge is not only volume, but decision unpredictability at runtime. Practitioners need a model where access, routing, and billing accountability are all evaluated together.

Cost visibility is becoming a prerequisite for AI operating discipline. The article correctly frames visibility as enabling both investment and monetisation, which means finance and security teams are converging on the same control plane. When usage is observable, organisations can distinguish productive automation from waste and shadow AI. That makes telemetry a governance asset, not just an operational report. The implication is that AI programmes without trustworthy usage data will struggle to scale responsibly.

Fragmented AI consumption creates a new kind of shadow AI exposure. The same conditions that hide cost also hide unauthorised or unapproved deployment paths. When teams can spin up overlapping tooling, model endpoints, and agent workflows without shared visibility, the organisation loses both financial and security control. The named concept here is AI fragmentation tax: the compounding cost of duplicated, unmetered, and ungoverned AI consumption across teams and environments. Practitioners should treat that tax as a governance signal, not an accounting nuisance.

From our research library:

• Organisations maintain an average of 6 distinct secrets manager instances, creating fragmentation that undermines centralised control, according to the State of Secrets in AppSec.
• 62% of organisations are at least experimenting with AI agents, but only 23% have scaled one into a single business function, according to McKinsey's 2025 State of AI survey.
• Read next: Agentic AI Identity Maturity Model

What this signals

AI fragmentation tax: the compounding cost of duplicated, unmetered, and ungoverned AI consumption across teams and environments. As agentic systems spread, organisations need a single view of model calls, MCP interactions, and downstream data movement or they will keep paying for the same capability multiple times.

Cost visibility changes governance from retrospective explanation to runtime control. When teams can see spend by agent, feature, and tenant, they can decide whether to route a request, cap it, or reprice it before the cost becomes structural margin erosion.

For practitioners

• Map the full AI cost path Inventory token usage, egress, retrieval, GPU time, MCP interactions, and agent-to-agent traffic so finance can attribute spend to the correct service or business unit.
• Establish cross-functional AI FinOps ownership Assign finance, platform engineering, product, and security shared responsibility for AI consumption policy, attribution, and escalation decisions.
• Meter consumption at execution time Set per-feature or per-tenant usage caps, anomaly alerts, and model-routing controls before bills are generated, so runaway usage is contained early.
• Tie access grants to billable accountability Ensure every agent, workload, or service identity that can invoke models also carries an owner, a cost centre, and an approval trail.
• Separate experimentation from production spend Create distinct budgets and controls for proof-of-concept work versus customer-facing AI services so abandoned experiments do not become zombie infrastructure.

Key takeaways

• Fragmented AI consumption is no longer just an efficiency issue, because it directly shapes margin, pricing power, and governance accountability.
• The article’s figures point to a clear operating problem: most organisations cannot forecast AI spend accurately, which makes scaling agentic systems financially fragile.
• The practical response is to unify attribution, metering, and ownership so AI usage can be governed as a business capability, not a hidden cost centre.

Key terms

• AI FinOps: AI FinOps is the discipline of measuring, attributing, and controlling the cost of AI systems at the level of usage and business value. It extends FinOps into probabilistic workloads where spend depends on model choice, prompt shape, retries, and tool execution rather than fixed infrastructure alone.
• Fragmentation Tax: The hidden operational and governance cost created when identity, access, and security tasks are split across too many tools. It shows up as manual reconciliation, duplicated administration, inconsistent policy enforcement, and slower response to change. The tax grows when teams rely on bespoke integrations instead of a coherent control plane.
• Unit Economics: Unit economics describes the revenue and cost profile of serving one customer or one transaction. In banking, it helps show whether growth creates sustainable value or simply scales losses. Strong unit economics usually depend on customer retention, efficient acquisition, and a business model that can support ongoing service delivery.
• Zombie Infrastructure: Zombie infrastructure is idle or abandoned compute that continues to consume budget after the original experiment or workload has ended. In AI programmes, it often appears as forgotten GPU instances, pinned memory, or unclaimed services that still generate cost without delivering output.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an IAM programme, it is worth exploring.