Picture this: your team spends six months standing up an agentic AI system to handle customer support. The demo looks incredible. Then you push it to production, and within weeks it's confidently citing outdated pricing, pulling records from the wrong customer accounts, and hallucinating policy details that haven't been true since the last product refresh. The business case collapses. The pilot gets shelved. Everyone looks at each other.
This isn't a hypothetical. It's the story playing out across thousands of enterprises right now. A new survey of 600 chief data officers (CDOs) makes the cause painfully clear: the data foundations needed to run agentic AI at scale weren't built with agents in mind.
What 600 Data Leaders Are Actually Seeing
Informatica, Wakefield Research, and Deloitte surveyed 600 data leaders at companies with $500M+ in revenue across the U.S., UK/EU, and APAC. The report, CDO Insights 2026, was published in January and carries some numbers worth sitting with.
Agentic AI adoption has reached 47% of enterprises, an impressive figure given where we were just two years ago. Another 31% plan to adopt within the next year. But that growth curve is running into a wall most executive teams haven't prepared for.
The survey data tells a specific story:
- 50% of companies planning or deploying agentic AI cite data quality and retrieval as major challenges to agent adoption
- 76% of data leaders say their company's governance frameworks have not kept pace with how employees are actually using AI
- 57% of leaders view data reliability as a key barrier to moving AI initiatives from pilot to production, essentially unchanged from 56% the year before
- 90% of data leaders are concerned that new AI pilots are launching without resolving data reliability issues uncovered by earlier pilots
- 86% of CDOs plan to increase investments in data management in 2026-2027
That last number is the tell. When 86% of the people responsible for enterprise data strategy say they're about to spend significantly more on data management, they're not doing it because things are going well. They've watched enough agent deployments stumble to understand where the real bottleneck is.
Agents Don't Forgive Bad Data the Way Dashboards Do
Traditional analytics and even most generative AI applications are fairly forgiving of messy data. A dashboard with a few stale records still tells a useful story. A chatbot drawing on slightly outdated documentation might give a 90% accurate answer.
Agents are a different animal. They don't just read data; they act on it. When an agent decides to cancel an order or update a customer record, it does so based entirely on what the underlying data tells it. A wrong answer from a chatbot is an annoyance. A wrong action from an agent, taken at scale, is a different category of problem entirely.
Nearly 60% of organizations consider data management critical for AI success, yet less than 20% report high maturity in any aspect of data readiness, according to Capgemini research cited by Informatica.
That gap — between perceiving data as important and actually having data infrastructure that's ready — is exactly where expensive pilots stall. And the retrieval problem compounds the quality problem. If an agent can't reliably reach the right data at the right time (because it's siloed in a legacy system, locked behind inconsistent APIs, or buried in an unstructured format the agent can't parse), it fails regardless of how clean that data is.
Why Employee Trust in AI Data Is a Warning Sign, Not a Comfort
One of the more counterintuitive findings in CDO Insights 2026 is what Informatica calls the "trust paradox." Despite all the data quality concerns, 65% of data leaders say that most or nearly all of their employees trust the data being used for AI.
That number should unsettle anyone who's been treating employee confidence as a proxy for data quality.
When employees trust AI outputs built on data foundations the CDOs themselves know are shaky, you have a compounding problem. The people acting on agent outputs may not recognize when those outputs are wrong. The survey makes this explicit: 75% of CDOs say their workforce needs upskilling in data literacy, and 74% say the same about AI literacy. Employees who can't critically evaluate AI output won't catch what the agent gets wrong, and in agentic contexts, that means bad actions propagate without review.
Governance is lagging even further behind. Three-quarters of organizations say their AI governance hasn't kept pace with actual AI use. Most companies (48%) are trying to adapt existing data governance tools to cover AI rather than building AI-native governance. The scale of agentic workflows will break that approach.
The Company Size Gap Is Really a Data Maturity Gap
The Informatica survey found that agentic AI adoption sits at 54% among large enterprises (5,000+ employees) versus 44% among smaller companies. Budget is the obvious explanation, but not the complete one.
The more significant factor is data maturity. Larger enterprises have, often out of sheer necessity, invested in data infrastructure, governance programs, and data management platforms over longer periods. They have data catalogs. They have MDM systems. They have people whose entire job is ensuring data quality across domains. That foundation is what lets them move faster when agents arrive.
Smaller and mid-market companies can close this gap, but only if they stop treating data infrastructure as maintenance and start treating it as the prerequisite the CDO data already shows it to be. The companies saying they'll increase data management investment in 2026-2027 aren't behind; they've connected the dots.
Three Questions to Ask Before You Greenlight Any Agent Pilot
Based on where the CDO survey shows failure modes clustering, three questions are worth working through before approving an agent initiative.
1. Do we know where our data lives, and can an agent actually reach it?
This is more basic than it sounds. Many organizations have data spread across dozens of systems with inconsistent schemas, limited API coverage, and integration layers built for human users, not autonomous agents. Before deploying an agent that needs to query customer records, product inventory, pricing data, and transaction history in real time, map the actual data topology. Where does the relevant data live? How fresh is it? Can the agent reach it reliably at query time?
If the honest answer involves a lot of "it depends" or "we'd need to check on that," fix the plumbing first.
2. Can we measure data quality in the specific domains the agent will touch?
Data quality is a spectrum, and different agent use cases have different tolerances. The relevant question isn't "is our data good?" It's "is our data good enough for this specific agent to take autonomous action without someone reviewing every output?"
For customer-facing agents handling returns or recommendations, the data needs to be accurate, complete, and current. For internal analytics agents summarizing reports, the bar is lower. Before launch, define the quality metrics that matter for this particular use case: completeness rate, accuracy against a known ground truth, freshness window. Then measure where you actually stand. Measurement is the prerequisite. Without it, you're deploying on faith.
3. Who owns agent governance, and what does it actually cover?
An agent that can write to a database, trigger emails, update records, or initiate transactions needs governance coverage that specifies exactly what it can and cannot do, under what circumstances, and when it escalates to a human.
Given that 76% of CDOs say governance hasn't kept pace with AI use, the odds are good that your existing policies were written before agentic AI existed. They need updating: specifically to cover agent permissions, data access scope, audit logging, and escalation paths when the agent hits something it shouldn't handle autonomously.
Which Data Domains to Fix First
Not all data cleanup delivers equal value. If you're resource-constrained (and who isn't?), prioritize investment in domains that map to the highest-ROI agent use cases.
Customer data deserves the most attention. Customer support and customer experience automation rank as the top-reported benefit of agentic AI, cited by 29% of CDOs in the Informatica survey. That's not surprising: CX use cases are highly visible, directly tied to revenue, and relatively easy to measure against baseline performance. But CX agents are only as good as the customer records they pull from. Incomplete customer histories, duplicate records, and stale contact information are immediate blockers.
The practical investments here are CRM data quality, customer identity resolution, and real-time profile synchronization. What makes CX a particularly smart first domain is the feedback loop it creates. When an agent routes a customer to the wrong tier or quotes a discontinued price, you know about it quickly. That fast signal is what lets you identify actual data quality problems rather than theorizing about them.
Finance and transaction data comes next. Finance operations, including accounts payable, reconciliation, and compliance flagging, are among the highest-ROI agentic use cases because they involve high-volume, rules-based decisions that are expensive to staff manually. The stakes here are also higher: agents making payment decisions on dirty ledger data don't just create operational noise. They create regulatory and audit exposure. The investment in clean financial data infrastructure pays back quickly precisely because the use case value is so high and the cost of errors is so legible.
Operational and IT data is the practical third priority for organizations that want early wins without heavy regulatory exposure. IT operations, including incident triage, root cause analysis, and system monitoring, is emerging as one of the most scale-ready functions for agentic AI. The data challenge here is usually integration breadth. The investment is in a unified layer that lets agents draw from monitoring tools, ticketing systems, and infrastructure logs through a consistent interface.
Data Readiness Is an Executive Problem, Not an IT Problem
When 90% of data leaders say new AI pilots are launching without resolving data issues from earlier pilots, that's an organizational failure, not a technical one. When pressure to show AI progress overrides the discipline to get the foundation right, only executives can break the pattern.
The CDOs increasing data management investment in 2026-2027 aren't doing it grudgingly as cleanup. They're doing it because the cost of skipping it is now legible in their production deployments. The right framing isn't about deployment timelines. It's about what data infrastructure makes agents trustworthy enough to actually run the business.
Agentic AI is a data problem first, with extraordinary software capabilities on top. The difference between a successful deployment and an expensive pilot that never scales depends entirely on whether someone at the executive level decides the foundation matters more than the timeline.