Article contributed by Garrett Blythe, Senior Data Architect, WWT & Andrew Sellers, Vice President, Technology Strategy & Enablement, Confluent, an IBM company 

An airline built an AI agent to help customers with seat assignments, upgrades, and flight changes. In the pilot, it performed flawlessly. It ran on static data on a local machine. Then it reached production. The airline's systems of record were spread out, and their updates were not arriving in real time. No single system could give the agent a current view of a flight fast enough. So, the agent tried to seat travelers in seats that had been filled seconds earlier. The lag between a seat changing hands and the agent learning about it was the whole problem. The model had not changed. The airline fixed the problem by moving much of that data onto real-time infrastructure. 

The airline is not an outlier. IDC research conducted with Lenovo found that 88% of the AI proofs of concept (POCs) it observed never reached widescale deployment. For every 33 that a company launched, only four made it to production. IDC's authors read that low conversion as a sign of low organizational readiness in data, processes, and IT infrastructure. 

We have watched versions of this story play out across industries, and the lesson catches many IT leaders off guard: The bottleneck is no longer the model. It is access to trusted, real-time, domain-specific enterprise data. 

This is not a new problem. AI resurfaced an old one. Most enterprise data platforms were built years ago to solve narrow problems for a single team, and each is owned by that team. Agents do not just need data; they need context, and that context often sits on an island owned by someone else. The AI boom has made that old problem more urgent and, in some cases, a question of business survival. 

Something has also changed in the market. In the first wave of generative AI enthusiasm, most organizations were stuck in the discovery phase, with impressive proofs of concept and almost nothing in production. Over the last six to eight months, we have seen a threshold change: many Confluent customers are now in production with AI-enabled use cases and getting a real ROI. This success is not due to frontier models getting marginally better in that time. The meaningful factor was their customers' ability to manage their data, give models the right context, and monitor, observe, and test the results. The organizations that succeed do not have access to different models than everyone else. They have mastered their own data. WWT has seen the same shift from the services side. Within roughly the last year, clients have moved from a small team exploring AI tools to an expectation that AI enablement spans the whole organization. 

Six assumptions keep IT leaders from closing the gap between AI ambition and data readiness. Here is each one, and what we have found to be true instead. 

The myth What is true 
1. A better model gets us better AI outcomes. Today's models are good enough for most use cases when given relevant, real-time, domain-specific data. 
2. If the pilot works, production will too. Pilots run on curated data. Production data is siloed, incomplete, and late. 
3. Batch data is fast enough. Batch suited human analysts. Automation acts at machine speed and needs data that moves just as fast. 
4. Painful data silos are just the cost of doing business. Silos persist because legacy infrastructure was not built for reuse. A better alternative now exists. 
5. Governance can come later. If the data is not governed, the AI is not governed. 
6. Agentic AI isn't ready for production yet. Dozens of Confluent customers already run AI agents in production with significant ROI. 

 

Myth 1: A better model gets us better AI outcomes. 

The truth. A better model may improve results at the margin, but it is not where to invest. Today's frontier models are good enough for most enterprise use cases when they are contextualized by relevant, real-time, domain-specific data. Even models from about nine months ago generally qualify. 

One of us has spent 20 years in AI, first as a machine learning researcher and then commercializing AI-enabled technology. For most of that career, the model was the hard part: weeks of feature engineering, model selection, bias characterization, and training. That is no longer where outcomes are won or lost. Switching models is the glamorous place to spend money, but the data foundation is the biggest lever, whatever model sits on top, whether it is high-end or low-end. Executives are often well informed about where AI is heading and eager to chase it. What trips them up is not awareness. It is that their enterprise is not ready to adopt. 

The costly corollary: We should wait for better models. 

Waiting does not reduce risk. It cedes ground to competitors who are already in production. 

Invest here instead: If the model is not the bottleneck, the time and money going into choosing and tuning models is better spent on the data those models work from. Four principles follow: 

  • Build reusable data products. Take your most critical data and make it contextualized, discoverable, and ready for any team or AI agent to use, so each new application starts with data that is already prepared.
  • Pair that data with more commoditized models. Widely available models do many jobs well when they are fed good data. Let the frontier model builders compete on model quality.
  • Stay loosely coupled to reduce dependency on any one model. Don't build so tightly around one model that you can't change it. A data infrastructure that is ready for whatever comes lets you absorb model gains as they arrive.
  • Put the foundation first and the tooling second. An AI budget is not a reason to debate models and token strategies before the foundation exists. Once you understand your data footprint, the tooling decisions get more accurate.

Myth 2: If the pilot works, production will too. 

The truth. Pilots typically run on static, curated data. Production data is siloed, held in separate systems that each record only one slice of the business, such as customer details, orders, or tax policies. It is also less trustworthy, incomplete, inconsistent, sometimes wrong, and slower. An agent fed that data acts on an outdated picture of the business. If you are not testing effectively, you will not see the effect until the agent is live. 

Gartner's research validates this finding. In a July 2024 survey of 1,203 data management leaders, 63% said their organizations either lack the right data management practices for AI or are unsure whether they have them. In February 2025, Gartner also predicted that through 2026, organizations would abandon 60% of AI projects that are not supported by AI-ready data. 

The myth is understandable. A demo at a hackathon can look like magic, and it is tempting to assume it needs only a week to go live. It does not work that way. 

Figure 1. The data a pilot sees versus the data production sees.
Figure 1. The data a pilot sees versus the data production sees.
Figure 1. The data a pilot sees versus the data production sees. 

A utilities example. We have seen this more than once with large utilities providers. They saw a demo of an AI workload, tested it in a small sandbox, and secured funding. When it came time to scale, the other systems they needed context from were not open enough to share it across team boundaries. The initiative took a step back, and its focus shifted from deploying agents to improving the organization's data posture. For the executive approving the budget, the lesson is that a demo shows the model working. It does not show whether the surrounding data will support it. 

A telecom example. A large telecom used AI for customer service and for monitoring its network infrastructure. For a long period, the workloads were not producing positive results. The analysis centered on the model: was it the wrong one, or was it returning inaccurate answers? The cause was the data. It was not real-time, so it arrived late and out of date. 

What surprises leaders. The gap in resources between a compelling demo and a production application that creates real value is far larger than expected. Models that feel like magic can then fail to perform reliably, because the data in a typical batch-oriented enterprise degrades automation that worked in testing. 

Before you approve the next step, ask your team: 

  • Is the pilot running on the same data production will see, or on a cleaned-up copy?
  • How old is the data by the time the agent acts on it?
  • Which other teams' systems does the agent need context from, and are they open to sharing it?

Myth 3: Batch data is fast enough. 

The truth. Batch was fast enough in the era of business intelligence when human analysts made the decisions. Automation acts at machine speed, so the data it relies on has to move at machine speed. 

Automation is a response to events: a network alert, a customer order, a fault on a factory floor. It works best when it responds to those events as they happen rather than waiting to process a pile of them later. Producing and consuming events in real time is the core design pattern Confluent addresses, and the value of automation tends to fall as the delay between the event and the action grows. 

 

Figure 2. The value of an automated action falls as the delay grows.
Figure 2. The value of an automated action falls as the delay grows.
Figure 2. The value of an automated action falls as the delay grows. 

Think of it this way: you would not cross a street using the positions of cars from yesterday, an hour ago, or even 30 seconds ago. An agent making decisions in your network should not rely on information from an hour or a day ago either. 

The cost of delay. Fraud has to be caught at the point of sale. If it is missed there, the opportunity is gone. In utilities and manufacturing, we have seen slow insights cost millions of dollars a day while a factory or pipeline is down. In these settings, true enterprise observability has to go beyond the older model of simple monitoring. Without a real-time streaming element, time to resolve stretches, operations teams get swamped, and the cost mounts. 

The scale argument. Batch might get one agent to market. It will not support the dozens or hundreds of agents that competitive enterprise organizations will run within five years. Batch always seems fast enough until it is not. An organization that builds on batch today may find itself behind its competitors, and behind the needs of its own business, within three years. 

Myth 4: Painful data silos are just the cost of doing business. 

The truth. Silos persist because legacy data infrastructure was never designed for data to be reused in more than one application. There is now a better alternative. 

Most organizations know about the problem. Data silos have been a drag on the world of data since long before the AI boom. What has been missing is a major organizational push to fix them, and a credible alternative to the status quo. 

How we got here 

Thirty years ago, storage was expensive, and developers needed protection from the errors that creep in when the same fact is stored in several places. So, databases stored each piece of information exactly once, in compact, tightly organized forms that carry little context on their own. That works when a human developer who understands the data model builds the application. An AI agent does not have that understanding. Storage is cheap now, and we have many options we did not have then, but data infrastructure in many places has not caught up. Many stacks were also tuned for one storage approach or one query pattern, and the bigger the data got, the more specialized that tuning became. 

The compounding cost 

In a batch-oriented enterprise, every new automated process or insight tends to require reworking how the data is organized, start to finish. A new copy of the data is moved from where it was generated to where it can be analyzed, with some transformation along the way. Those bespoke, brittle, batch-based pipelines deliver complete data hours after the events it describes. That delay was tolerable when human analysts set the speed of the business. As applications and analytics multiply, the complexity compounds. 

Figure 3. Each new use of data in a batch world adds a copy and a pipeline. In a streaming world, it adds a subscription.
Figure 3. Each new use of data in a batch world adds a copy and a pipeline. In a streaming world, it adds a subscription.
Figure 3. Each new use of data in a batch world potentially add a copy of each source and a pipeline. In a streaming world, it adds a single subscription. 

The alternative: data that is consumable by design 

Data streaming is built so that data can be consumed. A new application subscribes to an existing stream and gets contextualized, discoverable, trustworthy, secure data products, without reworking how the data is organized each time. That shortens the time to market for the next application or insight. It also changes how to think about the data itself: treat it as a product, not as a means to an end. 

Readiness is enterprise-wide 

Getting your own team's data in order is no longer enough. Agents need context that lives on other teams' islands, so organizations need governed, consumable data products across every corner of the business. No more operating in a vacuum. An AI or data initiative is not the responsibility of one team or one vertical. It is an enterprise-wide initiative, and everyone has to be on board. 

Share with guardrails 

How do you make data available across the business while keeping guardrails and usage standards in place? The pattern we see work best is centralized governance with decentralized execution. A governance framework with rules and standards of practice is set in one place and widely adopted. Producers own their data, and consumers serve themselves. This pattern is what Confluent has observed distinguishes those who maximize RoI from their data infrastructure across its roughly 6,500 customers over the last decade. Socio-technical practices such as data mesh help. Technology alone is not a silver bullet; organizations also have to make process and organizational changes. 

Myth 5: Governance can come later. 

The truth. You cannot govern AI unless you first govern the data behind it. An AI agent is only as good as the information it draws on. Data governance means having clear, enforced rules about what data you hold, who may use it, and whether it can be trusted. If the data is not governed, the AI is not governed. There is no other way to put it. AI governance as a whole is not a solved problem, but nobody can solve it without governing the data first. 

The stakes are high. An agent that is right 99 times and wrong once can erase the return an organization has already earned, and one bad automated decision can damage the brand. For most companies, a gain in worker productivity is not worth that risk. 

An airline learned what this looks like in 2024. A customer asked its website chatbot about bereavement fares after a death in the family. The chatbot told him he could claim the discount after booking, within 90 days of buying the ticket. That was wrong. When he asked for the refund, Air Canada argued, among other things, that the chatbot was a separate entity responsible for its own statements. A British Columbia tribunal rejected that argument and held the airline responsible for what its chatbot said. The amount at stake was small, a little over $800. The lesson was not: when your AI speaks or acts, your company owns the result. That is why governance cannot wait until after launch. 

Leaders tend to underrate four capabilities that effective automation depends on, and that data streaming helps provide. The first, governance, sets the rules for who may use which data. The other three are how you keep control of automation once it is running. They answer the questions any executive will ask: what is it doing right now, why did it do that, and can we recreate what it knew and try again? 

These four underrated capabilities are:  

  • Governance. Clear, enforced rules for who can use which data, and how.
  • Operations. Visibility into what the automation is doing, and the ability to act on it. Data streams help with visibility and with working out afterward what happened.
  • Observability. It is not enough to know that something happened; you need to know every input that informed it. That means tying the activity data a system produces (its telemetry) to specific moments and to the specific decisions or actions the agent took. It is much more feasbile to find a root cause when you have the whole ordered series of events rather than a couple of events with no order or timing between them.
  • Replayability. The ability to reconstruct what an agent knew when it made a decision, so you can fix the problem and try again. For 50 years, software engineering has built deterministic applications, meaning the same input produces the same result. AI agents are stochastic: they are probabilistic by nature, and they sometimes give different answers. You cannot recreate an agent's answer exactly, but you can replay the inputs it was given. That requires saving them. The Model Context Protocol (MCP), widely used to connect agents to tools and data, is fundamentally stateless. If those calls are not saved in a data stream, such as Kafka, it is hard to say afterward why an agent did something unexpected.

Here is what replayability looks like in practice: Suppose an agent denies an insurance claim it should have paid. Teams need to reconstruct the context the agent had when it decided, including prompts, system prompts, and tool calls. Then they can update the agent, run it again, and test whether a future version behaves in line with the goal. In much of today's tooling, everything is ephemeral. If data is lost or an incorrect action is taken, it disappears into the ether. 

How do you know an agent's answers are right? Best practices for monitoring the accuracy of AI answers about enterprise data are still emerging, they depend on the application, and we do not know of one that is completely perfect yet. What holds up is the foundation: well-governed, trustworthy, secure data, plus the ability to observe and replay what an agent did, so you can put guardrails around it. Keep a person in the loop as well, someone who checks the agent's work. Without a solid governance foundation, organizations tend to face one of two problems: distrusting output that is good or, in worse cases, fully trusting output that eventually proves wrong. 

Myth 6: Agentic AI isn't ready for production yet. 

The truth. Confluent has dozens of customers running AI agents in production with significant ROI, and agent-driven automation is already a competitive differentiator. Many people also meet agentic workloads every day as customers, without realizing it. Across the wider market, production is still the minority outcome. In Confluent's own 2026 Data Streaming Report, a survey of 4,625 IT leaders, 32% said they have agentic AI in production. Two-thirds (66%) cited data infrastructure and data quality as barriers to adopting it. 

What ROI looks like. It shows up in three main ways: 

  • More productive knowledge workers. A machine takes over routine, repetitive work it does well, such as summarization, and a person is freed up to do more. One banking client pairs anti-money-laundering agents with human investigators and believes the combination makes the team 70% more productive, which means it can process about 70% more investigations in a given period. That is the bank's own estimate.
  • New revenue. Hyper-personalization can raise attach rates, meaning the extra items customers add to a purchase, and enable things the business could not do before.
  • More service capacity. Automation lets a service center handle more calls, which in practice can mean a smaller service center with fewer workers handling a larger volume of calls.

Where we see agentic AI. The need is broad, and we do not differentiate between public and private organizations. The benefit categories we see most are healthcare, with personalized treatment and better patient outcomes; retail, with personalization; utilities, with operational efficiency; manufacturing, with faster speed to market; and financial services, with fraud detection and anti-money-laundering. E-commerce is another heavy user because hyper-personalization depends on real-time data. 

Patterns in the market. Behind those industries are specific customer case study examples. Each one depends on real-time data:  

  • Real-time driver dispatch, where optimization is not possible without live data.
  • Recommendations based on clickstream, search history, and how long a shopper lingers on an item.
  • Job matching, where bad data broke downstream systems and any attempt at personalization.

Medical-device telemetry, where brittle batch pipelines limited how fast a company could respond to machine monitoring and predictive maintenance. 

The foundation: how Confluent, IBM watsonx.data, and WWT work together 

If the myths describe what goes wrong, this is how the pieces fit when it goes right. 

Who does what? Confluent is a portfolio of technologies that stream, connect, process, and govern data in real time. IBM watsonx.data, an open hybrid lakehouse, is a set of tools that makes data governed, contextualized, and ready for AI. WWT, a Confluent Premier Partner, brings consulting and implementation services from the first assessment through full implementation. 

The handoff. Apache Flink cleans and enriches events upstream. Confluent's Tableflow materializes Kafka topics into Apache Iceberg tables, so watsonx.data can query live event data alongside historical data without a hand-built pipeline between them. 

Live context for agents. Confluent's Real-Time Context Engine serves agents the latest state of the business over MCP, with role-based access control and audit logging. IBM's developer tutorial shows the pattern with a retail example: Flink works out current product availability from a stream of inventory transactions, and watsonx Orchestrate agents answer a store associate's stock question from that live data instead of a periodic batch update. 

Reach across hybrid estates. IBM completed its acquisition of Confluent in March 2026. Confluent has more than 6,500 customers, including 40% of the Fortune 500. With the purchase, IBM can now extend streaming to mainframe, on-premises, and multi-cloud data, wherever an enterprise client keeps it. For example, IBM Data Gate for Confluent sends each change made in a mainframe database (IBM's Db2 for z/OS) into Confluent as it happens, instead of waiting for a batch copy. 

Figure 4. From event to agent, with a replay and observability loop.
Figure 4. From event to agent, with a replay and observability loop.
Figure 4. From event to agent, with a replay and observability loop. 

A  practical path forward 

Four steps help leaders move from a promising pilot to dependable production automation. 

  1. Assess data readiness before choosing tools. Map your data footprint, then pick the most critical data to turn into products. Decisions about models and tooling go better, with more nuance and accuracy once you understand your footprint.
  2. Secure enterprise-wide buy-in. Data and AI readiness is an organization-wide initiative, not one team's project.
  3. Sequence adoption. Teams already using data streaming or watsonx move fastest. Both streaming and AI are hard the first time, so AI is a poor gateway into data streaming, and organizations new to both should not take on both at once.
  4. Build in governance, observability, and replay from day one. Keep humans in the loop. Measure ROI in productivity, cost avoided, and revenue.

Where WWT can help. You do not have to take these steps alone. WWT, serves as an implementation partner for organizations on their data and AI journey. WWT helps customers assess their data readiness, modernize their data foundation, and translate AI ambition into practical business outcomes. Its teams can manage and design projects and provide full implementation teams, taking an organization from the first assessment through to a working deployment. The first step is the same as step one above: a clear understanding of your data footprint before any decision about models or tooling. 

A few questions can open the conversation with your technology leaders, and with WWT. How do we determine whether our data is AI-ready? How do we make data available across the organization while keeping guardrails and usage standards in place? What guardrails do we need around AI-driven automation? How will we measure the return on our automation investments? 

Ready to move AI from pilot to production? Contact your WWT account team. Start Today 

Technologies