How decentralized AI creates chaos debt inside companies

Successful AI adoption can create risks leaders discover later

A finance employee builds an AI reconciliation workflow, a recruiter creates a project that drafts job descriptions, and two support employees assemble a system for first-pass responses. The support system works well enough that another team starts routing tickets through it. None of these projects may have appeared on the company’s initial AI adoption roadmap, yet each solves a real problem before a formal adoption process recognizes the need.

That apparent success can itself create risk. Analiese Brown, chief people officer at Campminder, calls the liabilities that accumulate from decentralized AI without enough ownership or visibility “chaos debt.” Adoption rises, employees report saving time, and managers and senior staff have good reasons to let useful experiments continue. Yet those same signals can mean undocumented dependencies are starting to accumulate.

Those dependencies are hard to see because employees in finance, recruiting, support, and other teams can now build useful workflows without waiting for a formal technology project. Their experiments expose workflow gaps early, and successful ones can spread before central processes catch up. Once colleagues start relying on an experiment, the governance question changes: the organization needs to understand what depends on it and who is responsible for it.

The real liability begins when a personal workflow becomes infrastructure

That need for ownership emerges because people extend systems that already work. One employee’s workflow can supply data for another employee’s report, and that report can become an input into a team’s decisions. Nobody has to explicitly decide that the first workflow is business infrastructure. Dependence on its output gives it that role.

Once that chain exists, reconstructing it becomes difficult because the work may never have been documented and nobody may be responsible for keeping it running. The employee who built it can move to another team while the operating logic remains in that person’s notes. What began as an individual convenience then carries continuity risk for people who may not even know how the workflow works.

That continuity risk makes ownership more fundamental than simply classifying a tool as approved or unapproved. Approval can establish which products employees may use, but executives still need to know who is responsible for a workflow, which reports rely on it, and which decisions consume those reports. A useful governance model has to follow those dependencies because they determine the business consequence of failure.

Those consequences grow as AI makes small-scale building easier. A personal workflow stays relatively contained while its output affects only its creator; once other processes consume that output, the organization needs an owner who can explain the logic, maintain it through personnel changes, and understand what breaks when it changes. Chaos debt is the accumulated exposure created by those unresolved questions.

Shutting down shadow AI can erase the map of what needs fixing

The opposite liability appears when organizations prevent employees from redesigning inefficient work. Bradford Wilkins, vice president of people and organization at Cognite, describes “talent debt,” a term he attributes to Boston Consulting Group (BCG) and EY, as unrealized potential trapped in workflows that have yet to be redesigned. Leaders therefore face liabilities in both directions: moving quickly without ownership accumulates chaos debt, while leaving inefficient work untouched accumulates talent debt.

Discovering AI sprawl often leads leaders to limit employees to approved tools, require requests to pass through IT, and set policies for acceptable use. Those controls can address genuine concerns about unmanaged technology, but they can also erase evidence of why employees created their own systems. The unofficial workflow contains information about demand.

The finance example shows what that information can reveal. The employee’s reconciliation workflow exists because the finance system has a reconciliation gap. Likewise, the support system reveals that the help desk software cannot draft the first-pass responses its users need. Read operationally, each workaround is a requirement discovered through actual work: it identifies a missing function and shows that the gap matters enough for an employee to build around it.

That makes shadow AI useful for discovering requirements even when the implementation is unsuitable for continued unofficial use. A workflow may have unclear ownership, questionable data access, or dependencies that need formal control. Its existence still tells technology and operations leaders where an official workflow is failing employees. Closing the experiment without capturing that information discards evidence needed for redesign.

Losing that evidence can leave the underlying fragmentation in place because employees still face the reconciliation, drafting, or support problem. The work can then continue elsewhere in forms that are harder for leaders to observe. A better environment lets experimentation reveal unmet needs and makes the resulting dependencies visible before the business quietly starts relying on them.

That goal creates the central governance tension. The same workflow can simultaneously be useful infrastructure, an unmanaged dependency, and the clearest available signal that an official process needs redesign. Executives need all three views because each reveals a different decision about ownership, control, or redesign.

Visibility and guardrails offer a third option

That tension explains Brown’s governance principle that guardrails make speed sustainable in an AI-augmented organization. Under that view, governance creates conditions where experimentation can continue while ownership and risk remain visible. Guardrails create visibility and points for intervention, allowing decentralized work to develop within boundaries that leaders can understand.

Campminder provides one example through a career-development tool built by Brown’s director of people. The people team can see the questions employees submit to the tool, giving it enough visibility to identify cases where a person should become involved in the interaction. Human oversight is therefore tied to the situations employees actually bring to the system, allowing the team to respond as needs appear.

The timing of that visibility matters because Campminder designed it into the career-development tool from the start. An inventory can tell leaders what existed on the day they counted it, but workflows, users, and dependencies keep changing. Persistent observability, meaning the continuing ability to see what a system is doing, keeps the operating environment understandable as those conditions change.

That operating model is still developing as Campminder builds and socializes its AI governance model. Brown treats employee participation as part of the governance mechanism because participation affects whether employees take ownership of the resulting rules. Brown puts the ownership point this way: “You don’t bolt a policy onto a culture and call it done, because people need a hand in building it to feel any ownership over it,” with employee participation helping shape how those rules develop.

The same need for continuing visibility led Zapier to a different mechanism. Individual AI usage at the company was strong, while organization-wide visibility was inadequate. Brandon Sammut, chief people officer at Zapier, described a system that collects usage data from the places where employees use AI, including Zapier’s own product and products from Anthropic and OpenAI. Because Zapier’s own product is among the tools being measured, the company has a commercial interest in wider AI use; the system consolidates those records in a single data lake and runs automated reporting over the combined data.

Those consolidated records feed reports delivered through private direct messages at the beginning of every month. Each report covers the previous month’s activity and makes recommendations based on how the employee has been using AI. Zapier deliberately excludes usage leaderboards and rankings among colleagues, so the reporting focuses on improving an individual’s choices.

A benefits employee, for example, might routinely use the most expensive available model for work that a cheaper model could handle. The report can identify that pattern and recommend the cheaper option. It also stores earlier recommendations and changes later ones accordingly, which means oversight responds to an employee’s usage history instead of repeating a static rule every month. The company gains information it can act on while employees receive guidance tied to their actual work.

That guidance operates alongside one spending cap that Zapier retains. Sammut described its purpose as catching exceptional runaway costs from a misconfigured agent or stolen key, while ordinary employee usage can continue within the system. If someone needs the cap raised, they can request it by sending a message in a Slack channel. The control targets a specific failure mode and keeps the path for legitimate activity lightweight.

That narrow purpose becomes clearer from mechanisms Zapier considered and rejected. Usage leaderboards were an option, as were hard spending caps. Sammut’s account is that both would encourage compliance behavior and produce less useful information, so Zapier chose private recommendations and a narrowly targeted emergency limit. The governance system can then learn how AI is being used while keeping its strongest control focused on exceptional spending risk.

The Campminder and Zapier examples connect visibility to different operating decisions. Campminder can see interactions that may require human intervention and is involving employees while shaping governance; Zapier consolidates usage data and turns it into individualized guidance while reserving a hard limit for exceptional spending risk. In both cases, visibility lets the organization observe activity, apply judgment, and intervene when a defined risk appears.

Audit the dependencies before writing another restriction

Those visibility mechanisms point to the immediate task for executives who discover decentralized AI: establish what the organization already has. A useful lookback is the past six months because apparently personal experiments may already have entered reports, decisions, and operating processes during that period. The audit needs to trace builders, owners, dependencies, and accessible data so leaders can understand the workflows behind the AI products employees have opened.

Four questions make that investigation concrete:

• “What has been built in the last six months, and by whom?”
• “Who owns each of those workflows once the person who made it changes roles?”
• “Which of them are sitting inside a process the business now depends on?”
• “What data can these workflows reach?”

Those answers establish legibility, meaning leaders can identify the workflows, the people responsible for them, the business processes relying on them, and their data reach. With that information, leaders can decide which experiments should be formalized, which underlying workflows need redesign, and where controls are warranted. A finance workaround that has become a reporting dependency, for example, presents a different governance problem from an isolated drafting experiment even if both use the same approved AI provider.

That audit establishes a starting state because the environment continues to change after the review finishes. Employees create new workflows, existing ones acquire new users, builders move roles, and dependencies spread. Campminder’s career-development tool provides the earlier example of the stronger design principle: continuing visibility lets the responsible people see emerging situations and apply human judgment as they occur. A current inventory establishes what exists today, while persistent observability keeps that operational picture current as the workflows change.

Key takeaways for leaders

• Track success before it becomes chaos debt: Useful employee-built AI can spread into business processes before formal governance catches up. Management teams need visibility into ownership and dependencies as adoption grows.
• Assign owners to critical workflows: An AI workflow becomes infrastructure when other reports, processes, or decisions depend on its output. Business and technology owners should document who maintains each critical workflow and what breaks when it changes.
• Treat shadow AI as a requirements signal: Employee workarounds reveal gaps in finance, support, recruiting, and other official workflows. Technology and operations teams can capture those unmet needs before restricting or replacing unofficial implementations.
• Build visibility into AI governance: Persistent observability gives organizations a current view of usage, dependencies, and situations requiring intervention. Campminder and Zapier show how targeted oversight can preserve experimentation while addressing defined risks.
• Audit dependencies before adding restrictions: Review AI workflows built in the past six months, including their builders, owners, business dependencies, and data access. Use that inventory to decide which experiments need formal ownership, workflow redesign, or stronger controls.