Why agentics projects fail and how to fix them

It’s amazing what deltas exist between AI implementations.

Wayfair built agents to support its suppliers and now automates 41,000 support tickets a month. C.H. Robinson built agents that read shipping-request emails, connect information across messages and attachments, fetch additional context, and automatically create more than 5,500 shipping orders per day, saving 600 person-hours per day.

But then there's Klarna, which one analyst called the "poster child for bad AI deployments." A customer service agent meant to do the work of over 850 employees contributed to quality issues and declining customer satisfaction.

Same technology, wildly different outcomes. And the deciding factor is rarely the model. As we’ll discuss below, what separates the wins from the cautionary tales is the underlying data, and whether users can trust it.

Why agentic AI projects fail without trusted data

Agentic AI is undergoing the most aggressive technology adoption curve in a generation.

According to the 2026 Gartner CIO and Technology Executive Survey, more than 60% of organizations plan to deploy AI agents in the next two years, and only 17% have done so today. McKinsey estimates that gen AI could add $2.6 trillion to $4.4 trillion in value annually across enterprise use cases.

Yet Gartner also projects that more than 40% of agentic AI projects will be canceled by the end of 2027. A 2026 Fivetran report found that only 15% of organizations are fully ready, even as the vast majority have already invested millions.

That gap between ambition and readiness has a cause, and it's a specific one. The key limiting factor for successful agentic AI implementation is generally not model quality, but poor data quality and governance.

Frontier AI labs have released powerful foundational models. But if the data that feeds them is inaccurate, incomplete, or inconsistent, you get poor results.

Agentic AI extends the reasoning ability of generative AI to decisions and actions performed through software, not only producing information but also acting in the world. That shift is exactly what raises the stakes.

Without adequate data and context, agentic AI can misanalyze a situation, choose the wrong response, perform the wrong action at scale, and cause cascading workflow errors. With poor security, governance, and accountability, including at the level of data assets, it can become difficult or impossible to trace the origin of a wrong decision and remediate it.

The public examples are instructive. A chatbot deployed by Air Canada gave a customer inaccurate information about bereavement pricing, failing to refer to and correctly cite the company's internal policies. Replit's software development copilot deleted a production database during a code freeze and created false data in the process, the kind of incident that hard production controls should prevent, whether the actor is human or AI.

We find it useful to organize the risks posed by agentic AI into three categories:

1. Operational correctness risk: the agent acts on stale, missing, or misunderstood data.
2. Control-plane risk: the agent has the wrong permissions, weak auditability, or poor security boundaries.
3. Human-system risk: humans overtrust, under-review, or cannot effectively supervise the agent.

The through-line in all three is context. For agents, stale data is not merely an analytics problem; it can become an operational action taken on the wrong version of reality. A dashboard built on last week's numbers is a bad report. An agent acting on last week's numbers is a bad decision, executed automatically, at machine speed.

Agents demand more from your data than dashboards ever did

Agentic AI imposes far greater demands on an organization's data infrastructure than human-centric analytics workflows, especially reports and other forms of decision support.

A human analyst consumes data intermittently; an agent consumes it continuously. A human can absorb tacit, tribal knowledge through experience and can intuit, remember, or investigate where a data asset came from. An agent needs explicit access to context, explicit governance, and access to data lineage.

The key challenge is the lack of a governed, consistent context across the enterprise data estate. AI agents need more than raw data. They need reliable data movement, shared business logic, semantic context, lineage, and access controls across every source and consumer.

Meeting that requirement rests on two pillars: automation and centralization. Centralizing data and ensuring that it's inventoried and defined once is essential for scaling access to trusted data, controlling infrastructure costs, and managing compliance risks. Once data is centralized, your team has to systematically transform, that is, model, it into a usable context layer for AI. That's what makes lineage, semantic layers, and governance critical:

• Lineage shows users, including agents, where data came from and whether it can be trusted.
• Semantic layers apply shared business meaning to tables, granting humans and agents alike a shared, consistent understanding of how data maps to real-world business concepts.
• Governance controls what users and agents can access, decide, and change.

There's one more requirement that's easy to underrate: interoperability. As AI tooling continues to evolve, the best model, compute engine, orchestration layer, or activation channel for one workflow may not be the best for another.

Interoperability gives teams the freedom to connect systems without duplicating data, rebuilding pipelines, or locking agent workflows into a single vendor. If each AI use case requires copying data into a proprietary silo, organizations lose governance, portability, and control.

Where agentic AI actually works best

Deciding where to point an agent matters as much as the infrastructure underneath it. AI has a very jagged ability profile due to its design, excelling at some tasks while deficient at others.

It's strong at pattern recognition and completion for text and code, including drafting, editing, summarizing, translation, and coding. It's strong at ideation and brainstorming, especially when breadth is required, and the cost of a bad suggestion is low. It’s skilled at reasoning through problems with clear, specified premises and constraints.

On the other hand, it's weak at discerning truth from plausibility when facts are obscure or highly specific and at knowing when not to answer. It struggles with open-ended problems that require causal reasoning from limited evidence and long-horizon planning, and with adversarial interactions.

That profile points to a clear set of characteristics for good agentic use cases:

• High volume
• Repeatable structure
• Text/code-heavy inputs
• Clear success criteria
• Low-cost human review
• Reversible or low-risk actions
• Available authoritative data

Poor use cases have the inverse: ambiguous accountability, high legal or safety stakes, sparse data, adversarial users, long-horizon planning, and irreversible actions.

It helps to remember that these systems tend to augment work rather than replace it. In 2016, Geoffrey Hinton, who would later win the 2024 Nobel Prize in Physics for his work on artificial neural networks, predicted that radiologists would be extinct as a profession by 2021 due to AI image recognition. By 2025, radiologists' pay, employment, and workloads had never been higher. AI is far likelier to augment complex workflows than eliminate roles.

We've put these principles to work ourselves. Fivetran's Chief Product Officer uses agentic AI to perform conversational analytics on Jira data. Directly querying Jira's MCP server was untenable at scale, so the team moved Jira data into BigQuery via Fivetran, used a Claude Skill to query it, and produced product-ops insights in hours rather than multiple analyst sprints.

Separately, our support team embedded a custom AI app in Zendesk to answer questions, draft responses, summarize handovers, and find similar tickets. Built using Fivetran and dbt, it centralizes knowledge from Zendesk, Slab, Jira, GitHub, Google Drive, Gong, Salesforce, and docs.

The common thread: each agent is narrow, high-volume, and grounded in authoritative data it can‌ reach.

Build an agent you can trust

Once you've picked a workflow, the build itself is more approachable than most teams assume. Building agentic AI models from scratch is a complex undertaking that can cost many millions of dollars and months of development time.

A more practical and less risky option is to augment a foundation model with your organization's unique, proprietary data using a RAG architecture. You can create specialized agents that perform specific tasks by interacting with your operations through the dbt MCP server and similar controlled interfaces.

The safest way to roll that out is in tiers of progressively growing autonomy:

• Start with read-only agents that retrieve and summarize information.
• Then, build drafting agents that prepare outputs for human review and final implementation.
• Next, build bounded write-back agents that act within strict, narrow limits. Potentially risky actions should require approval, while sensitive, irreversible, regulated, or safety-critical actions should remain prohibited from autonomous execution.

Getting each of those tiers right depends on several details, such as your reference architecture, the discipline of context engineering, and a concrete readiness checklist.

We’ve laid out this blueprint in The data leader's primer for agentic AI. It walks through the reference architecture step by step, the strengths-and-weaknesses map for choosing use cases, and the checklist we use to take an agent from idea to production so your project avoids ending up part of the 40% cancellation statistic.

Download the full guide for the roadmap.