What makes an AI agent trustworthy
Prajakta Damle
Senior Director, Product Management
To prevent AI agents from making errors, businesses must focus on building trusted business context.
Picture an AI agent that pulls last quarter's sales, builds a forecast, and sends it to your regional teams before lunch. It moves fast, and it sounds certain. It also counted returns as revenue, with no way to tell the difference. By the time someone catches the error, the regional teams have already started planning around it.
Context is fast becoming one of the most valuable assets a company owns. Much of the conversation surrounding AI today still centers on the model. While a powerful model is essential, it is only part of the story. The real differentiator for your business is the ability to make sound decisions quickly and accurately, which depends on a trustworthy picture of your operations. That means ensuring your AI agents know what your data represents, how it fits together, and the rules that govern it.
The evidence points the same way. Only 7% of enterprises say their data is completely ready for AI, according to a 2026 study from Cloudera and Harvard Business Review Analytic Services, and 73% say data quality deserves more priority than it gets. Gartner® predicts that through 2026, organizations will abandon 60% of the AI projects that aren't supported by AI-ready data.1 The data foundation, not the model, is where AI efforts stall.
From answering to acting
The first wave of enterprise AI mostly answered simple questions. Someone asked a chatbot for a summary, read the reply, and decided what to do next. A wrong answer was usually easy to catch.
Agents work differently. They can act, carrying out multi-step work across your systems without a person reading every line. That's where the real value lies, and the real risk, too. When an agent acts without a solid grasp of your business, the failures tend to fall into three patterns:
-
Hallucinations: Without grounding in business truth, an agent guesses. It fills gaps with invented facts that sound authoritative even when they're wrong, and even capable models are prone to it.
-
Operational lag: The agent acts autonomously based on static data, missing real-time operational updates and executing decisions that are directionally correct but factually outdated. This delay can create critical discrepancies across downstream workflows before human teams can intervene or catch the error.
-
Unauthorized access: The agent reaches data it was never cleared to see, because the rules about who can access what didn't travel with the data. In IBM's 2025 analysis of data breaches, 97% of organizations that experienced an AI-related breach lacked proper AI access controls.
A list of coordinates isn't a map
The data conversation has long been about logistics: where data lives, who owns it, how to keep it clean. Agents need more than that. They need the meaning around the data, what it represents and how it connects, in a form they can act on.
A list of GPS coordinates is just numbers. Accurate, but they won't get anyone anywhere. A map takes those same points and adds what makes them useful: the roads between them, what sits near what, which routes are open. Data is the coordinates. Context is the map. Most companies have handed their AI agents a list of coordinates and asked them to find their way around the business.
So the work for leaders is shifting from managing data to building context: the meaning, the relationships, and the rules that let an agent reason about your business the way an experienced employee would. Treat it as something to capture, own, and improve, and you hand your agents a real advantage.
What it takes to get this right
Building context an agent can trust comes down to three things.
One trusted source of meaning. Context has to come from everywhere the company operates, not just the data warehouse. That means gathering the meaning behind core systems like SAP, Salesforce, and Workday, alongside the data your teams touch every day, so an agent reasons from one governed picture instead of a dozen partial ones.
Context that keeps itself current. A business generates new data every day, and context maintained by hand goes stale about as fast as people can write it down. The solution is to have a self-learning context layer that identifies gaps in context based on agent interactions and automatically captures meaning as information is created. This enables context to be derived as new information arrives across sources such as files, records, emails, meeting notes, chat threads, ticketing systems, and code repositories. Today's AI can read new content as it lands and add the business context people rarely have time to record by hand.
When organizations automate their context and metadata, the payoff is measurable, and so is the cost of skipping it. Gartner has put numbers on both. It predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance and access failures.2 And in a separate Gartner survey of 360 IT leaders in the second quarter of 2025, only 23% said they were very confident in their organization’s ability to manage security and governance when deploying generative AI tools.3
Strong context speaks directly to those gaps. Gartner found that organizations with the most mature AI-ready data capabilities are achieving up to 65% greater business outcomes, including revenue growth and cost optimization. It calls context, semantics, and metadata mission-critical for data and analytics and even describes them as the brain of a working AI system.3
Retrieval you can trust and verify. In an agent-driven company, search becomes how agents find what they need to act. Two things make it dependable. It has to honor permissions, so an agent only acts on what it's cleared for. And you need a way to confirm that the context it pulls is the right context, which turns "I think it's working" into something you can test and improve. Optimized search is critical here, not only to ensure the agent has the most relevant context but also for cost optimization; better retrieval leads to lower token usage and lower costs overall.
Bringing these capabilities together is what Google Cloud's Knowledge Catalog is built to do, operating on three foundational pillars: aggregation, which unifies the entire enterprise data estate by combining native Google sources, multi-cloud platforms, and partner catalogs into a single governed source of truth; enrichment, which drives continuous automated learning by using Gemini to actively mine structured schemas, query logs, and unstructured data relationships; and search, which delivers secure, low-latency, Google-grade retrieval at scale while strictly adhering to metadata access permissions.
Putting the idea to work
Bloomberg Media offers a useful example. Like many large organizations, it sat on a deep, sprawling data lake. The stakeholders who most needed answers had business questions, but those answers were buried in technical data, so getting them meant waiting on a specialist.
So the team built an internal agent, its Data Access AI Agent, that lets stakeholders explore the data with plain-language questions and trust the answers that come back. Those answers hold up because of the context underneath them. Using Google Cloud's Knowledge Catalog, the agent draws on the company's own institutional knowledge, so its answers reflect how the data actually fits together.
William Anderson, Bloomberg Media's chief technology officer, put the value plainly: "By grounding our AI in a trusted institutional context, we ensure confidence in the accuracy and quality of every insight generated."
Bloomberg Media launched its Data Access AI Agent by unifying enterprise metadata and business context using Knowledge Catalog that generated 102 usage notes across 20 tables and improved the agent's SQL accuracy by 63% during initial development.
By blending automation with human intelligence, Virgin Media O2 built a scalable, governed, and intelligent metadata foundation. Over 20,000 data assets are automatically documented with Knowledge Catalog to enhance data discovery, strengthen data quality, and empower teams across departments to make confident, data-driven decisions — turning metadata into a strategic enabler of innovation, trust, and enterprise-wide value.
Where to start
You don't have to solve all of this at once. A sensible first step is to look honestly at where your agents get their context today: what they're answering, what they draw on to do it, and who owns that. The answers usually show both the risk you're carrying and where to begin.
The companies that pull ahead with AI agents will be the ones that treat context as what it's become: a strategic asset worth building and protecting. It's what turns a capable agent into one the business can actually rely on. To go deeper, start with Google Cloud's Knowledge Catalog and map it to where your agents need context most.
1. Gartner Q&A, Lack of AI-Ready Data Puts AI Projects at Risk, February 2025. Gartner is a trademark of Gartner, Inc. and/or its affiliates.
2. Gartner Press Release, Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, May 26, 2026
3. Gartner Press Release, Gartner Says Organizations with Successful AI Initiatives Invest Up to Four Times More in Data and Analytics Foundations, April 16, 2026
