Agentic AI arrived in the enterprise the way most powerful things do, faster than the ground was ready for it. Nearly two-thirds of enterprises worldwide have now experimented with agents. Fewer than one in ten have scaled them to deliver tangible value. That is McKinsey's June-2026 reading, and the shape of it will be familiar to any data leader who watched a dazzling agent demo turn into a stalled program the moment it met the real estate.
The reflex is to blame the model, or the framework, or the integration. The evidence points somewhere else. Eight in ten companies name data limitations as the roadblock to scaling agentic AI (McKinsey, 2026). Agents are unusually unforgiving of a weak foundation, because a single agent chains many steps and tools together, and a multi-agent system passes work between specialists. Feed either one fragmented, inconsistent, ungoverned data and the errors do not stay put. They compound.
This is the uncomfortable inheritance of the agentic era for a Chief Data Officer. The organization wants autonomy, and autonomy raises the stakes on every data decision that used to be quietly survivable. A schema drift that once broke a nightly report can now drive thousands of wrong actions before anyone notices. So the work is not to chase the newest agent framework. It is to build the foundation that lets agents act safely, and to do it in the right order. This piece lays out five moves, each with an interactive exhibit you can pull apart, and each with the place RealAI can help.
Force one: agentify a few workflows, not everything
The first mistake is scope. Faced with an agentic mandate, teams try to sprinkle autonomy across dozens of processes at once, and end up with dozens of half-working pilots and nothing in production. The path that works starts narrow. Pick a small number of high-value, end-to-end workflows where more autonomy would genuinely change the outcome, prove them, then widen (McKinsey, 2026).
The selection is not a popularity contest. Each candidate workflow carries a different value potential, and, just as important, a different data-readiness. A workflow can be worth a fortune and still be un-agentifiable this quarter because the data it depends on is fragmented across systems nobody has reconciled. Rank by value and by feasibility at the same time, and the list usually collapses to a handful worth starting now, a few worth funding a data effort for, and a long tail to leave alone.
For a data leader this reframing is a relief, because it turns an impossible mandate into a fundable plan. Instead of "make everything agentic," the sentence becomes "here are the three workflows that carry most of the value and whose data is ready, and the one where a short data effort unlocks the next tranche." In our own engagements a value-first assessment produces exactly that ranking in 4 to 6 weeks, and it is far cheaper than discovering the same thing through a year of stalled pilots.
The exhibit makes the discipline visible. Value concentrates in the first few workflows, and readiness decides which of those you can act on today. Chase the crimson bars first and you join the stalled majority. Start with the amber ones and you ship something that holds, then earn the right to widen.
Force two: agents amplify whatever they stand on
Two agent shapes are emerging, and both are exacting about data. A single-agent workflow has one agent using many tools and data sources in sequence. A multi-agent workflow has specialised agents collaborating through shared context (McKinsey, 2026). Give a single agent fragmented data and it makes inconsistent decisions from one run to the next. Give a multi-agent system the same fragmented data and it does something worse: the agents lose their shared picture, and an error introduced by one propagates through the others before any human sees it.
This is the mechanism behind so many stalled pilots. In a reporting world, a data inconsistency produced a wrong number that someone eventually caught. In an agentic world, the same inconsistency drives actions at machine speed, and by the time an alert fires the system has already acted on it many times over. The old habit of cleaning data on a schedule cannot keep up. Quality has to be continuous: automated validation, anomaly detection, and enrichment running in the pipelines, with the same standards applied to the data agents generate as to the data they consume (McKinsey, 2026).
The fix is not glamorous and it is exactly the CDO's home ground. Give both agent shapes a consistent, interoperable view of the data, maintained continuously, and the failure modes recede. Withhold it and no amount of model quality will save the pilot.
The point the exhibit argues is that the archetype is not the risk. The data is. A governed, shared context is what lets either shape hold together at scale, which is why continuous data quality is a scaling decision, not a hygiene chore.
Force three: modernise the stack, do not rebuild it
Ask how to make data agent-ready and the wrong answer is "rebuild everything." The right one is to strengthen the architecture layer by layer, because the strongest organisations build modular, evolutionary stacks whose parts can be replaced as the technology moves (McKinsey, 2026). The agent-ready architecture is a stack: a source layer where data is ingested with governance travelling alongside it; a platform layer of vector stores and embeddings; a semantic layer of ontologies and knowledge graphs that turns data into meaning an agent can use; a data-products layer of reusable, owned, observable assets; a consumption layer where orchestration and governed retrieval decide what to fetch; and governance running through all of it, from a medallion progression up to an AI gateway that controls how models reach unstructured data.
The reason to care about the whole stack rather than a favourite layer is that agents are only as reliable as the weakest one. Brilliant embeddings on top of a missing semantic layer means agents that retrieve fragments they cannot interpret consistently. A pristine source layer with no governed retrieval means sensitive content surfacing where it should not. Each gap becomes an agent acting on conflicting data, confidently and fast.
The move for a data leader is to assess the stack honestly, find the layers that are storage-era rather than agent-era, and modernise those. Not a rebuild. An evolution, component by component, with the interfaces held stable so the work compounds rather than churns.
Pull a layer out in the exhibit and the whole thing below it goes dark. That is the honest picture of an agentic architecture. It rewards completeness, not brilliance in one place, which is why the sequence is to modernise the weakest layers rather than to gold-plate the strongest.
Force four: governance is the control
There is a point in every agentic program where the question stops being "can it work" and becomes "can we let it." As agents move from suggesting to acting, governance becomes the primary mechanism of control, not a compliance layer bolted on afterward (McKinsey, 2026). Clear policies have to define what each agent may do, what data it may touch, and when a human has to approve. Every action needs to be logged, traceable and auditable, and high-impact actions need a human in the loop with the ability to pause or override.
The tension a CDO has to manage is between autonomy and oversight. Autonomy is where the value is, and it is also where the risk is. The failure pattern is an agent population whose autonomy has climbed while oversight stayed flat, opening a gap where agents act faster than anyone can watch. Well-built agents help close that gap: guardrail agents that monitor other agents for policy and brand violations, running inside a defined control function, extend oversight at machine speed. Underneath them the operating model matters just as much. Business domains own day-to-day governance of their agent workflows, while a central data and AI team maintains the shared platforms and guardrails, a federated split that lets autonomy grow without losing accountability.
The practical answer is an execution environment where autonomy is bounded by construction. Every agent runs walled off, reaches only what it has been granted for the task in front of it, and writes every action to an audit trail you can hand to a regulator. Governance stops being the brake on agentic value and becomes the thing that lets you press the accelerator.
The dial shows the whole argument in one motion. Raise autonomy alone and a crimson gap opens. Raise oversight to match, with guardrail agents and logged, least-privilege actions, and the gap closes. Scaling agents safely is that balance, held deliberately.
Force five: own a small model on your governed data
The last move is where the foundation pays a dividend most leaders underestimate. Once you have a governed data foundation, you no longer have to rent all of your intelligence by the token. Organisations with well-structured internal data can fine-tune smaller, domain-specific models on that data, and those models are not only cheaper to run than a general endpoint, they are more resilient and more compliant (McKinsey, 2026). The reason is straightforward. A small model tuned on your own governed data learns your actual work, stays inside your perimeter, and removes the recurring cost and exposure of shipping sensitive data to someone else's endpoint.
The economics turn on data quality. At low data quality a general endpoint looks cheaper, because fine-tuning a small model on messy data buys little. As the governed data improves, the small model gets sharply more effective, and its total cost of ownership falls below the endpoint you were renting. Past that breakeven the decision inverts: the model you own is cheaper, and it keeps your data where your obligations live.
For a data leader this is the argument that ties the whole foundation together. The governed data you built for the first four forces is the same asset that makes a sovereign, auditable model affordable. The foundation is not a cost centre that agents draw down. It is the thing that lets you own your intelligence rather than lease it.
The crossover is the moment the foundation pays for itself. Build the governed data and a model you own becomes the cheaper, safer choice, not the idealistic one.
Where to start
The five forces are a sequence, not a menu, and the sequence is the same one that separates the few who scale from the many who stall: build the foundation, then let the agents stand on it.
Assess first. Rank your workflows by value and by data-readiness, and pick the few that are worth starting now. Name the data gaps behind the rest rather than pretending they are not there. A value-first assessment gets you a defensible plan in 4 to 6 weeks.
Transform the layers that matter. Modernise the weak layers of the data architecture for agents, evolving what you run rather than rebuilding it, and hold data quality continuously so both agent shapes act on one version of the truth. This is where the human operating model shifts too, from doing the work to supervising and orchestrating agents that do it, which is a capability a data organisation has to build deliberately rather than assume.
Govern from day one. Give every agent a walled-off, audited runtime, keep autonomy and oversight moving together, and split governance between the domains that own the workflows and the central team that owns the platform. Then, with the core in place, own the models that reason over your most sensitive work instead of renting them.
The organisations that win the agentic era will not be the ones with the most agents. They will be the ones whose foundation lets agents act safely, cheaply and accountably, over and over. That foundation is what a data leader is uniquely placed to build, and it is the difference between an agent demo and an agentic operation.
- 2/3
- Enterprises that have piloted agents (McKinsey, 2026)
- <10%
- That have scaled them to value (McKinsey, 2026)
- 8 in 10
- Blocked by data, not models (McKinsey, 2026)
- 4-6 wks
- To a value-first agentic roadmap (RealAI)
The organisations that win the agentic era will not have the most agents. They will have the foundation that lets a few agents run safely, cheaply and accountably, again and again.
This is the second in a three-part RealAI series on the foundations of scaled AI, written for Chief Data Officers and the leaders who own the estate with them. First was the data-readiness dividend. Next: reimagining the infrastructure that runs agents, without the cost curve that usually comes with it. The series responds to McKinsey's 2026 technology research on AI data readiness, agentic foundations, and infrastructure.
“A house is only as strong as its foundation. Agentic AI is only as strong as the data it stands on, and in 2026 that foundation, not the model, is what separates the few who scale from the many who stall.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.