AI has moved from novelty to necessity inside most organizations, including our own. Teams now lean on it daily to innovate faster, summarize information, and get more value to clients. But in a recent LinkedIn Live session, Dataracity's Luke laid out a candid, practical case for a different kind of urgency: as AI usage scales, the absence of structure around it is becoming one of the more significant operational risks facing modern businesses. Below is a summary for leaders deciding how to scale AI responsibly, built on the frameworks shared in the session.
The Familiar Problem, Wearing a New Face
Anyone who has managed a data or BI team will recognize this pattern: duplicated spreadsheets, conflicting reports, business logic buried in someone's head, and a maintenance headache every time a rule changes. Long-standing, ungoverned busiwork already produces sprawl, brittle reconciliations, and knowledge that walks out the door with the person who built it.
The session's central warning is that unstructured AI adoption reproduces this same problem, only faster and less visibly. AI happily generates new code, new reports, and new "answers," each shaped by whatever context an individual user happened to feed it. Every employee effectively becomes their own silo of business logic. And because AI is confident, agreeable, and rarely flags its own blind spots, teams can develop blind trust in outputs that were built on incomplete information, unnoticed assumptions, or a context window that quietly dropped relevant detail.
“Treat AI as a junior team member: give it small, accurate context, and always review its work before it ships.”
Four Pillars for Managing AI Risk
The session organized the risk landscape into four practical areas leaders should be actively managing:
| Pillars | Core Risks |
|---|---|
| Security | “Shadow AI” (employees using personal AI accounts), prompt injection attacks that can exfiltrate stored credentials, and AI agents accessing more of a company's files — SharePoint, email, shared drives — than intended, all of which weaken audit trails and expose sensitive data. |
| Context | AI pays closer attention to the beginning and end of a conversation than the middle, so as context windows grow with added files, tools, and integrations, accuracy quietly drifts — a risk that compounds as more “skills” and connected systems get layered in. |
| Accuracy | AI can filter, infer, or silently work around missing information without flagging it, making errors hard to catch — especially in vague requests or complex automations — and hard to trace back once discovered, particularly if the person who built the process has since left. |
| Cost | Token usage scales with the size and complexity of a request, and duplicate questions across a team, daily automations, and API access fees can push usage well past standard licensing, often without anyone noticing until the bill arrives. |
The Fix: A Human-in-the-Loop Operating Model
Rather than restricting AI use, the session proposes redefining where AI fits in a workflow. The recommended split: AI is excellent at the innovation phase — quickly testing an idea, drafting an approach, breaking the ice on a new process. That accounts for roughly 80% of the initial effort. But the remaining 20% — validating the logic, documenting the decisions, and converting a working prototype into a governed, automated process — has to remain owned by a person.
In practice, that means AI should rarely be the thing running a process day after day. Once an approach is proven, it gets automated through conventional tools — a Power BI dashboard, a scheduled Python job, a workflow platform — rather than re-querying an AI model every morning. This keeps ongoing token costs low and, just as importantly, keeps the underlying business logic documented and owned by the organization, not locked inside someone's chat history.
Turning the Model Into a Strategy
The webinar translated this thinking into four building blocks any organization can adopt:
- A governed sandbox: A locked-down, enterprise-licensed environment where employees can experiment with AI on approved data, with IT controlling exactly which sources, folders, and connectors are accessible. No personal AI accounts, no ungoverned access.
- AI champions: Business-savvy super-users who validate a prototype before it moves toward production, acting as the bridge between what a business user proved out and what IT ultimately builds and supports.
- A central platform: One place where automated processes, dashboards, and AI-assisted apps live, backed by a single, cleaned dataset. Centralizing reduces duplication, simplifies governance and auditing, and gives AI a smaller, more accurate context to work from, which also lowers cost.
- ● A knowledge base: Documented business rules, definitions, and decision logic that AI can draw on directly, rather than requiring every user to re-explain company context from scratch. This is also what preserves institutional knowledge as people change roles.
Together, these elements shrink the context AI needs for any given task — which simultaneously improves accuracy and brings token costs down. It's a virtuous cycle: better governance produces cheaper, more reliable AI, not just safer AI.
The Takeaway for Leadership
AI is not the problem; unmanaged AI is. The organizations that get the most durable value from AI will be the ones that treat it like any powerful but inexperienced new hire: valuable for accelerating ideas, but never left unsupervised, never trusted blindly, and never allowed to become the sole keeper of business logic. A sandbox for safe experimentation, clear ownership through AI champions, a centralized platform and dataset, and a living knowledge base are not bureaucratic overhead — based on this session, they are what make AI adoption both safer and, over time, cheaper.