Key Takeaways
- Agentic AI removes the human safety net from data pipelines, raising the bar from “good enough to interpret” to “good enough to act on autonomously, without review.”
- Closing the Agentic AI Data Integrity Gap requires unified, context-rich, and governed data that agents can reliably reason with and act on.
- Organizations should start with a single, high-value use case and treat it as the foundational build from which future Agentic workflows will scale.
I’ve spent a lot of time recently in conversations about Agentic AI — with customers, at conferences, and in a TDWI panel discussion I recently joined alongside Fern Halper, Ph.D. of TDWI and Priya Singh of Reltio. One theme that keeps surfacing is that organizations aren’t short on enthusiasm for Agentic AI, but where many fall short is data that’s actually ready for it.
TDWI research on Agentic AI readiness is clear on this point. While a majority of organizations are actively exploring agentic systems, fewer than 10% report having multi-agent systems in production. That gap is a data problem — and specifically, what we at Precisely call the Agentic AI Data Integrity Gap: the disconnect between AI ambition and the quality of the data powering autonomous systems.
What makes this moment different from every previous wave of AI investment is who, or what, is consuming the data. That shift changes everything. We covered this in detail during our discussion with TDWI, and you can find the highlights here.
Why Agentic AI Exposes Data Gaps Traditional Programs Missed
Data programs have historically been built for humans. Analysts, dashboards, reports — the underlying assumption was always that a person would be on the receiving end of that data, with the ability to pause when something looked off, ask a question, or check another source.
Agentic AI removes that safety net entirely. An agent doesn’t pause, but acts immediately across multiple systems simultaneously, triggering downstream decisions in real time. The standard shifts from “good enough for a human to interpret” to “good enough for an autonomous system to act on without review.”
That’s a materially higher bar, and it’s exposing underinvestment in three key areas that traditional programs were able to work around.
- Context. Agents need to understand not just what data says, but what it means in relation to other data — and to the business. A customer record is just a set of fields. That same record enriched with location, behavioral history, relationships to other customers, and business rules? That’s something an agent can reason with.
- Freshness. Batch-refreshed data may be fine when a human has time to interpret it. But for an agent making real-time decisions, stale data is likely wrong.
- Traceability. Agentic AI introduces a new requirement: traceability against the decision itself. When an agent acts, you need to reconstruct what it saw, what it reasoned from, and why it did what it did. That’s different from anything a conventional data catalog or lineage diagram was built to handle.
The gap organizations face isn’t necessarily because they’ve failed to invest in data, but because they’ve been building for humans — and now they need to build for a fundamentally different consumer: the agents themselves.
ANALYST REPORT
TDWI Benchmark Report: Agentic AI Readiness
Benchmark your organization against TDWI’s Agentic AI Readiness Assessment, a framework designed to evaluate an organization’s ability to move from experimentation to enterprise scale deployment using five dimensions.
What Does It Take to Build an Agentic-Ready Data Foundation?
Our discussion also touched on what data needs to be considered Agentic-Ready, so that an agent can act on it with confidence. Agentic-Ready Data requires several interconnected layers, each of which has to hold up under the demands of autonomous systems.
Here are a few of the requirements that Priya and I covered:
Unification without centralization. Agents don’t inherently know which systems to query. They need a connected foundation across structured and unstructured sources — cloud platforms, on-premises systems, and even legacy environments. There’s a tendency in the industry to overlook what still runs on IBM, SAP, or mainframe infrastructure. Those systems handle high-intensity workloads and contain data that’s often material to business decisions. Unlocking it — making it discoverable, trusted, and usable — remains one of the harder problems in this space.
Real-time access. If an agent is making decisions based on data refreshed two weeks ago, that’s not true autonomy. Data must be continuously refreshed at a cadence agents can trust.
Enrichment beyond internal data. Internal data never tells the complete story. Authoritative external data — location attributes, third-party signals, environmental context — gives agents the depth to make decisions that hold up to scrutiny.
The mistake I see most often is treating these as separate workstreams. Separate ownership is fine, but they have to operate in concert. Any agentic implementation depends on all of these working as a single foundation. If one is missing, the whole system is exposed.
How Should Organizations Think About Context and Semantics?
One of the richest conversations during our panel was about context — specifically, the machinery needed to make it actionable for AI agents. Ontologies and knowledge graphs are seeing a real resurgence, and for good reason.
Ontologies give agents a shared vocabulary. Knowledge graphs let them traverse the relationships between entities when data is being called upon. Without those two things working together, context is still just data about data. You activate it through these tools.
Priya made a point during our discussion that’s worth repeating. She observed that most of this isn’t information organizations need to create from scratch — it’s tribal knowledge that already exists in the business. As she put it: “It’s just your business blueprint. If you look at your business, you would be able to organize it into different domains, different processes, and you should have a common glossary across it.” She emphasized that by creating a semantic graph that your data can connect to, you gain a 360-degree view of your operations.
That resonates with what I’m hearing from customers. The organizations making the most progress aren’t building a semantic layer from zero. They’re taking metadata enriched through catalog initiatives, definitions maintained informally, and domain structures they already understand — and formalizing those into something agents can consume.
The key principle is to stop treating context as something you bolt on at the application layer. It needs to live alongside the data from the start. Otherwise, agents reinvent it inconsistently, fragmentation builds, and the downstream challenges compound.
What Does Responsible Data Governance Look Like When Agents Are Making the Decisions?
Another theme that ran throughout our discussion: governance for Agentic AI is a fundamentally different problem than the governance programs most enterprises have built.
Traditional governance asks who can access data and whether it’s accurate. Agentic governance adds a different set of questions:
- What did the agent do with it?
- Why did it make the decision it made?
- Can we reconstruct that later?
That’s closer to an audit log of agent behavior than anything you’d find in a conventional data catalog — and it has to run at the speed of the agents themselves. Policies encoded in PDFs, or approval workflows measured in hours, are irrelevant to systems making decisions in milliseconds.
A specific challenge worth calling out: the governance of unstructured data. Many of the organizations I speak with assume it’s ungovernable. It isn’t, but the approach has to be different.
- With structured data, governance lives in schemas and access controls defined in advance.
- With unstructured data — contracts, emails, call transcripts — governance must happen at the point of extraction and use. What did the agent pull from this document? What did it infer? Can you trace that inference back to the source?
As Priya noted, “It’s not only what the agent did with your data, but what exactly it did, at what time, and what systems it affected — upstream and downstream.” Auditability and traceability are non-negotiable for any organization thinking seriously about autonomous decision-making at scale.
There’s also a cultural dimension that doesn’t get enough attention: someone has to own the outcomes of autonomous decisions. TDWI research shows only about 32% of organizations have clear ownership and accountability assigned for agent-based systems. That number needs to move before the agents do.
Where Should Organizations Start Building Toward Agentic AI Readiness?
The answer isn’t to tackle everything at once — and it isn’t to start so small that the effort never requires solving foundational problems. The right answer is one use case, chosen intentionally.
Pick something narrow enough to deliver business value within a quarter or two, but real enough that it forces the foundational work: unifying relevant data, enriching it with context, governing it properly, and improving quality.
Treat it not as a one-off experiment, but as the foundation from which future agentic workflows will scale. Priya and I landed in exactly the same place on this point during our discussion, and it’s where I’d leave any organization starting this journey.
The sequence generally looks like this: assess data quality in the relevant domain, close the gaps, build the semantic layer alongside it, and put governance in place before the agents go live — not after.
For a more in-depth look at everything covered here — including TDWI’s Agentic AI Readiness Benchmark report — watch the full session: How to Make Enterprise Data Fit for Autonomous Decision-Making.
