Establish a Human-AI Agents Alliance. Ravi Shankar of Denodo explains why agentic AI requires a shared data foundation to scale autonomy.
Artificial intelligence is moving from analysis to action. The newest class of systems—agentic AI—does not merely suggest next steps; it can take those steps, adjust to feedback, and coordinate across tools without waiting for a human to click a button. For business leaders, the opportunity is not framed as humans versus machines, but as two complementary workforces learning to share the same sandbox. The organizations that pull ahead will be the ones that intentionally design how people and AI agents collaborate, grounded in a common view of data and a clear division of responsibilities.
From Scripts to Goal‑Seeking Systems
Each major wave of automation has redefined how work flows. Mechanical automation took on repetitive physical motion. Software automation codified repeatable tasks with deterministic rules. Agentic AI advances the frontier again by pursuing goals rather than just executing scripts. These agents interpret changing inputs, propose a plan, test a step, evaluate results, and iterate until an outcome is achieved.
What makes this shift so consequential is endurance and scale. Agents do not tire, request sick days, or lose focus. They can work overnight, across time zones, and in parallel swarms. That reliability invites leaders to re‑architect processes that currently depend on handoffs and queue times—think customer follow‑ups, invoice reconciliation, compliance evidence collection, field‑service triage, or first‑line IT operations. Humans are not removed from the loop; humans are moved up the loop, toward work that composes strategy, sets priorities, and handles ambiguity where judgment is essential.
There is also a quality dimension. When standard operating procedures are encoded as agent behaviors, organizations gain an audit trail of what happened and why. Consistency improves because the “how” is executed the same way each time, while the “when” can expand to 24×7. The result is not just cost efficiency but operational resilience.
A Coordination Playbook for Humans and Agents
Autonomy does not imply isolation. Even the most capable agent needs intent, boundaries, and a way to confer with humans. Begin by defining objectives in business terms and translating them into measurable success criteria. Attach policy guardrails — what the agent may do automatically, what actions require pre‑approval, and which decisions must always be escalated. Expect the unexpected: agents will encounter edge cases that were not envisioned during design. Build deliberate tripwires that pause execution and route the question back to a human when risk, novelty, or conflicting signals are detected.
Check‑ins should be part of the design, not an afterthought. Agents ought to report status at natural milestones: when a task is completed, when assumptions have changed, or when confidence drops below a threshold. These moments maintain transparency without forcing humans to micromanage every move. Over time, as demonstrated performance increases, the cadence of approvals can relax — just as a new colleague earns greater autonomy with experience.
Trust cuts both ways. Agents must demonstrate reliability through explainability, consistent results, and adherence to policy. Humans, for their part, must set boundaries that enable rather than stifle. Over‑constraining an agent with excessive approvals undermines the very advantage of autonomy; under‑constraining it introduces risk. The art is to reduce friction while preserving accountability. Practical patterns include human‑in‑the‑loop review for high‑impact actions, human‑on‑the‑loop monitoring dashboards for ongoing work, and service‑level objectives that quantify timeliness, accuracy, and rework.
One Data Plane, Many Perspectives
True coexistence requires a shared lens on reality. If people and agents see different data or operate on different refresh cycles, collaboration devolves into conflict.A common data plane provides a synchronized, governed source of truth so that decisions are contextual, traceable, and aligned. It also shortens feedback loops: when humans adjust a parameter or correct a data point, agents incorporate the change without losing continuity.Crucially, this shared data environment should be implemented as a logical data layer. Rather than copying data into yet another store, a logical layer connects to source systems and exposes business‑ready views with common semantics, governance, and security. For agentic workflows, this model is ideal: agents can fetch the freshest information directly from authoritative systems in real time, with minimal latency and zero copy lag. The shortest path from question to answer is often through data virtualization and semantic abstraction, not another round of extract‑transform‑load. The logical layer masks physical complexity while centralizing policy, enabling agents to act quickly without bypassing controls.
Different agents need different data at different moments. A planning agent benefits from historical trends and forecasts; an execution agent needs current state, permissions, and cost constraints; a monitoring agent consumes streaming events and exceptions; an optimization agent pulls performance metrics and user feedback. The common data plane should accommodate all of these modes with fit‑for‑purpose access patterns — streaming subscriptions for events, low‑latency queries for operational snapshots, and analytic views for pattern detection. Add fine‑grained governance so that sensitive attributes are masked and actions are logged, and you get speed with stewardship.
How to Start — Without Stalling
When organizations first consider introducing AI agents into their operations, the instinct is often to over-engineer the perfect framework before taking the first step, but waiting for perfect conditions is the fastest path to falling behind. The better approach is to begin deliberately, with the kinds of tasks that are already well understood and measurable.
Start by looking for high-volume, rules-driven activities — the ones your teams execute every day without much variation. Things like ticket routing, renewal reminders, or standard compliance checks. These are ideal proving grounds because success can be observed quickly, and the impact is immediately tangible.
From there, articulate what “good” looks like. In the same way you would define a role for a new team member, specify the objective, the boundaries, and which decisions require human sign-off. These constraints give the agent clarity and provide your organization with confidence that autonomy will not drift into unpredictability.
Next, ensure the foundation is ready. That means establishing a logical data layer to supply agents with consistent, governed, real-time access to the information they need. Without this data backbone, agents will operate with partial context, and their performance will reflect it.
Once the foundation is in place, design a check-in rhythm. Agents should know when to proceed independently, when to surface a clarification, and when to pause entirely. These touchpoints reduce friction and prevent unnecessary human oversight while still keeping humans in the loop where it matters.
Then, establish the metrics for trust. Track precision, turnaround time, exception frequency, and the percentage of tasks completed without human help. These signals tell you where to widen autonomy, where to tune behavior, and where human judgment is still indispensable.
Finally, make the learning visible. Share early wins, unexpected findings, and near misses. When teams see how agents behave — and how quickly they improve — momentum builds naturally. Behaviors spread across the organization faster when they are demonstrated rather than mandated.
What Changes Next
As adoption expands, the relationship between humans and agents will deepen. Agents will not only execute tasks but orchestrate other agents, negotiate resource constraints, and schedule work around cost or carbon intensity. Humans will grow more comfortable delegating, and new roles will emerge — agent product managers, prompt engineers for operational contexts, and governance stewards who balance speed with compliance. Etiquette will evolve too: when should an agent interrupt a meeting, how should an agent summarize a day’s work, and what tone should it use when asking for approval?
The destination is not a fully autonomous enterprise; it is a collaborative enterprise where judgment and computation reinforce one another. When people set direction and define constraints, and agents execute with speed and consistency on a shared logical data layer, outcomes compound: faster cycles, fewer errors, clearer accountability, and room for human creativity to flourish.
The sandbox is already big enough for both. The next move is to architect coexistence on purpose — one data plane, many agents, and a workforce where humans do what humans do best.
A quote or advice from the author: “The real breakthrough of agentic AI isn’t autonomy — it’s the ability for humans and intelligent agents to operate on the same data foundation, each amplifying the strengths of the other. That’s where organizations will unlock transformative value.”
Discover the latest trends and insights—explore the Business Insight Journal for up-to-date strategies and industry breakthroughs!
