After months of pushing employees to use more AI, enterprises are learning the bill is only one part of the problem. Many organizations also lack a clear view of which AI agents, models, tools, prompts, and custom workflows are actually running across the business.
ServiceNow made that argument in a July 2 blog on the hidden cost of “invisible AI.” The company warned enterprises are now dealing with agents in one department, models in another, third-party tools no one inventoried, and custom skills built by teams that may no longer own them.
The first AI cost shock came from tokens. That wave of AI cost discipline focused on usage: how many tokens employees burned, which models they used, and whether those costs could be reduced. The next wave is more structural—whether the organization can see, govern, measure, and retire the AI it has already deployed.
Gartner has put numbers behind the risk, predicting that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls.
Analysis
What this means: AI sprawl may be the new shadow IT. The first cost shock came from token usage; the next one is coming from agents and tools no one can inventory, govern, or connect to measurable value. ERP leaders have already seen this pattern with SaaS sprawl and cloud overruns, but AI is compressing the same governance problem into a much shorter cycle.
From Token Spend to AI Sprawl
Token discipline tells a company where money is going. It does not tell the company what is running.
ServiceNow’s 2026 Enterprise AI Maturity Index found 59% of organizations are using agentic AI. But only 20% have basic testing, auditing, and risk assessment in place.
That gap explains why AI ROI is so hard to prove. Many companies can see adoption. Fewer can show whether AI improved response time, reduced case volume, lowered error rates, shortened development cycles, improved close quality, or removed manual work from a business process.
ServiceNow Chief Analytics Officer Vijay Kotu outlined a four-part AI ROI framework built around adoption, utility, impact, and value. The missing piece for many enterprises is impact: whether the business metric the AI was supposed to move actually changed.
If no one set the baseline before deployment, ROI becomes a guessing exercise after the fact.
ROI Needs an Inventory
The visibility problem turns AI governance into an infrastructure issue.
ERP teams already manage systems, integrations, roles, workflows, master data, controls, customizations, and change histories. AI adds another layer on top of that. Agents need owners. Models need approved use cases. Prompts and skills need lifecycle management. Outputs need evidence trails. Data access needs policy enforcement. Business value needs to be measured against process outcomes, not enthusiasm.
The danger is companies only discover the sprawl when something breaks—a model changes behavior, a vendor updates terms, an agent touches data it should not, a workflow produces an error, or a board asks why AI spending rose without measurable return.
Analysis
What this means: AI ROI starts with knowing what exists. A company cannot defend the next AI budget if it cannot say which tools are live, who uses them, what they touch, and which process metric they improved. The measurement problem is not just financial, it is operational.
Control Before the Blast Radius
The risk rises when agents stop answering questions and start taking action.
An AI agent that summarizes a policy document creates one level of risk. An agent that updates a supplier record, grants access, approves a workflow, changes a ticket priority, or triggers another agent creates a different one.
At that point, governance cannot be an after-the-fact review. The system needs to know who initiated the action, what authority they had, which data the agent could access, which system it touched, what policy applied, and whether a human approval was required. It also needs a record that can be audited after the action is complete.
ServiceNow frames that as the difference between managing isolated AI projects and orchestrating AI as enterprise infrastructure. Agentic AI creates chains of action, and chains create blast radius. A misconfigured permission, bad model response, stale context, or poorly governed tool call can move across workflows before anyone sees the damage.
The same issue will appear inside ERP environments. Finance, procurement, HR, supply chain, and IT workflows are full of permissions, thresholds, approvals, segregation-of-duties rules, audit evidence, and exception paths. AI agents operating in those workflows need to inherit those controls, not route around them.
Get Our Free Weekly Newsletter
Governance as Architecture
The tokenmaxxing debate asked whether companies were spending too much on AI. The invisible AI problem asks whether companies know enough to spend wisely.
A company can cap usage and still have unmanaged AI sprawl. It can reduce token spend and still have agents without owners. It can move to cheaper models and still lack audit trails. It can show adoption dashboards and still fail to prove process impact.
The more mature answer is not simply more restrictions. It is an AI operating model that treats inventory, ownership, access, monitoring, evidence, and value measurement as part of the architecture from the start.
That is where ERP leaders should pay attention. As enterprise software vendors embed agents into core platforms, buyers will need to ask more detailed questions.
- What AI assets are visible to administrators?
- Can agents be mapped to business processes?
- How are prompts, skills, models, and tool calls governed?
- What gets logged?
- Who can revoke access?
- How does the system show whether AI improved a process rather than just increased activity?
The next AI buying cycle will be won by the vendor that can prove its AI is visible, governable, measurable, and safe enough to run inside daily operations.
Analysis
What this means: Governance cannot wait until AI is already everywhere. Once agents are embedded in finance, HR, procurement, supply chain, and IT workflows, retrofitting controls becomes expensive and politically messy. The companies that move fastest will be the ones that make AI visible before it becomes unmanageable.
Sponsor Industry-Grade Research
The ERP Governance Test
ERP vendors are already moving AI agents closer to transactions, workflows, and decisions.
That makes visibility a necessity. Customers will need more than a list of AI features. They will need to understand how AI agents are registered, monitored, tested, audited, and retired. They will need to know whether agents respect existing ERP permissions, whether third-party agents can be controlled consistently, and whether AI activity can be tied to process outcomes.
This is also where AI sprawl becomes a finance problem. If a CFO asks what the company received for higher AI spending, the answer cannot be “more usage.” It has to be fewer exceptions, faster cycle times, better service levels, lower manual effort, reduced rework, stronger compliance, or faster modernization.
The uncomfortable lesson from tokenmaxxing is that usage was too easy to mistake for progress. The lesson from invisible AI is that even progress can become hard to prove if no one can see the system clearly.
