It was 8 pm on a Thursday when the personalization team went into crisis mode.

“​​It wasn’t an outage. It was a slow-motion failure.”

Engagement metrics had tanked just days after launching a new AI-powered recommendation engine. Click-through rates (CTRs) nosedived. User feedback poured in: “Why am I seeing this?” “I’ve never bought anything like this.”

Sales sounded alarms over missed revenue. Customer experience teams were overwhelmed with complaints. Product marketing scrambled to craft a response. The product team insisted the model had tested just fine.

At first, the team blamed seasonality. Then, the UI.

But the real issue started upstream. The data team updated the taxonomy, introducing new categories, renaming segments, and reorganizing hierarchies. These changes shifted the data feeding the model, but no one retrained it. No alerts fired — no contracts required coordination — no team took ownership of the drift.

Sure, this is a fictional e-commerce company. But the story? All too real.

“The algorithm didn’t fail. The trust chain did.”

These types of failures don’t come with error messages. They show up as subtle misalignments between what the system believes and what the world has already changed.

And they’re becoming alarmingly common.

Most companies still struggle to trust their data. Research from Precisely shows that only about a third feel confident in their own data, and McKinsey found that fewer than one in five data projects meet their financial goals. These studies reveal a deeper issue: the need to address trust issues with data is quintessential to getting the most out of data initiatives.

This article introduces the Trust Erosion Curve, a framework for identifying early warning signs that your data strategy is quietly unraveling and offers a practical playbook for rebuilding trust before AI fails loudly and publicly.

The Trust Erosion Curve is a conceptual model that describes how organizational confidence in data systems erodes gradually, often imperceptibly, before collapsing rapidly due to compounding failures. This model draws from reliability engineering and behavioral science to illustrate how trust usually degrades in nonlinear ways.

Initially, trust declines almost imperceptibly as minor issues in data discovery, quality, or governance go unnoticed. But over time, these issues interact and compound, eventually reaching a tipping point where trust collapses rapidly.

McKinsey’s “The Data Dividend: Fueling Generative AI” (September 2023) reinforces this progression through several key observations.

The article emphasizes that foundational cracks in data systems, such as fragmented architectures, low observability, and inconsistent preprocessing, become “big problems” when organizations try to scale generative AI (GenAI) initiatives. These issues are often tolerable during pilot phases but emerge forcefully during operationalization, leading to performance degradation, mounting technical debt, and business mistrust.

The result: a slow buildup of data friction that eventually undermines user trust and model outcomes.

This progression follows the arc of the Trust Erosion Curve:

• First, a subtle decline as quality inconsistencies, latency, or explainability issues accumulate quietly.
• Then, a steep drop in trust occurs when these issues interact, causing visible model failures, broken service-level agreements (SLAs), or audit failures.

McKinsey’s recommendation to “track rigorously and intervene quickly” aligns with the curve’s purpose: recognizing the early signals of trust breakdown is crucial for CDOs and data leaders who want to scale innovation without risking catastrophic failure.

The model serves as both a warning and a diagnostic tool, helping data leaders identify and mitigate risks before they cascade across AI and analytics ecosystems.

Here are four critical signals that your data strategy may be quietly unraveling, and how to act before they impact your AI ambitions.

Signal 1: Data discovery is broken

When finding data feels like an exercise in frustration rather than empowerment, your discovery infrastructure fails. If analysts, data scientists, or business users must ask around, rely on tribal knowledge, or search multiple disconnected systems to find basic datasets, discovery gaps are widening.

As the effort to locate data increases, velocity slows, and so does trust. Teams begin creating duplicate datasets, relying on stale sources, or abandoning self-service altogether.

Before you build AI, fix how people find data

How to fix it: Make metadata work for the business

• Invest in enterprise-wide metadata management and domain-driven architecture that enables the proper context around the data.
• Implement intuitive data catalogs integrated into everyday workflows.
• Capture lineage automatically to show where data lives and how it flows.

Take Kroger Co., one of the largest retailers in the US. It operates a wide range of business units, brands, and subsidiaries, resulting in the organization’s data being scattered across different systems that didn’t communicate with each other. Todd James, (former) Chief Data & Technology Officer at 84.51° (Kroger’s retail data science division), explains that their strategy focused on building context-aware systems.

By establishing a domain-driven data architecture, Kroger can not only stitch together disconnected systems but also power features like semantic search, product recommendations, and helpful reminders such as “Did you forget something?” prompts. This approach enables a seamless and personalized shopping experience throughout the customer journey.

Signal 2: Quality conversations are always reactive and not proactive

While data quality covers core dimensions such as accuracy, consistency, validity, and freshness, it becomes effective only when teams actively detect and resolve issues before they escalate.

If quality concerns only surface during post-mortems or when something breaks, it indicates that governance practices are lagging and unable to support scalable data reliability.

If you’re fixing quality after it fails, it’s already too late

When an executive questions a KPI or a machine learning model returns flawed predictions, the damage to trust has already occurred.

A resilient data strategy surfaces anomalies, freshness issues, and consistency gaps before they impact decision-making or AI models.

How to fix it: Invest in data observability

• Deploy automated data observability frameworks across ingestion and transformation pipelines.
• Monitor critical data assets continuously for freshness, drift, and schema changes.
• Establish proactive data health dashboards that business stakeholders can also monitor.

Signal 3: Data contracts exist in people, not systems

In many data organizations, data sharing begins with informal handshake agreements(“I’ll send you this file daily”). While expedient at first, these undocumented expectations create fragile dependencies. As systems evolve, producers may change schemas, delay updates, or retire fields without notice, breaking downstream dashboards or models. The result is a slow erosion of platform trust and rising technical debt.

Make expectations enforceable, not implied

How to fix it: Enforce contracts like code

• Make contracts explicit and versioned: A robust data contract is a formal agreement between producers and consumers. It should define:
  • Schema structure (field names, types, nullability)
  • Data freshness SLAs (e.g., “available by 6 a.m. daily”)
  •  Quality thresholds (e.g., “<1% null values allowed”)
  • Backward compatibility rules (e.g., additive-only changes unless versioned)
  •  Change notification policies (e.g., “notify consumers two sprints in advance”)
  • Like APIs, these contracts should be version-controlled to allow for safe iteration without breaking existing dependencies.

• Make contracts visible and traceable: Embed contracts into your data catalog, pipeline orchestration tools, and monitoring layers. This visibility enables traceability, impact analysis, and automated enforcement.
• Designate enforcers and equip them: Enforcing data contracts isn’t solely IT’s job. The accountability typically falls to:
  • Data product owners or platform teams who establish enforcement policies
  • Data producers who commit to contract terms
  • Data SREs or governance stewards, who implement observability rules
• Enforcement mechanisms can include: Automated schema checks in CI/CD pipelines:
  • Monitoring dashboards that trigger alerts when contracts are breached
  • Governance policies that block deployments if schema validation fails

A visible, enforced data contract transforms ad hoc exchanges into resilient, governed data products, enabling teams to scale trust across the ecosystem.

Signal 4: AI models age faster than you think

A sudden decline in model performance is rarely just an AI problem: It’s a data trust problem manifesting downstream.

Upstream drift is the model’s silent saboteur.

When AI models drift or degrade after deployment, it’s often because the input data they rely on has changed, degraded, or lost relevance. These changes in the structure, semantics, or statistical properties of input data that AI models depend on are referred to as upstream data shifts.

Modifications in source systems, data pipelines, or user behavior often cause these shifts. When these upstream data shifts are unchecked, model drift feels like a persistent challenge, suggesting a warning sign that trust across your data supply chain is breaking down.

How to fix it: Catch data drift before the model does

• Integrate model monitoring into your data observability strategy.
• Tie model input and output quality metrics directly to data product owners.
• Implement retraining pipelines that trigger based on both model performance and upstream data drift signals.

Closing the loop on trust failures: The trust recovery playbook

Trust in data and AI systems often fails in silence until the dashboard is wrong, the model is off, or the executive asks, “Why didn’t we catch this sooner?” But trust decay isn’t random. It follows patterns. And with the right feedback loops, it can be caught early and reversed.

Each of the four early warning signs – discovery breakdowns, reactive quality, invisible data contracts, and upstream model drifts – is a signal that trust is slipping.

“Trust isn’t a nice-to-have; it’s infrastructure for AI at scale.”

The good news is that each signal already suggests its fix. By instrumenting for these signals and operationalizing the responses, embedding them into catalogs, contracts, monitoring pipelines, and ownership models, organizations can transform fragile trust into a self-reinforcing loop.

When response mechanisms become system design, not just incident reaction, trust becomes durable, measurable, and scalable.

For data leaders, this isn’t just about avoiding failure; it’s about safeguarding organizations by eliminating the sharp drop-off possible on the Trust Erosion Curve. The ability to scale AI and analytics with confidence does not come from perfect data but from resilient systems that catch imperfections early and respond by design.

In that ability lies your competitive edge.

About the Author:

Anusha Dwivedula is the Director of Product at Morningstar, where she leads data and analytics initiatives that power smarter decision-making across the financial ecosystem. Over the past 12 years, she has helped build and scale platforms that combine cloud infrastructure, data quality frameworks, and observability tools, always with a focus on making data more trusted and accessible.

Dwivedula has played a key role in Morningstar’s cloud transformation, leading the design of resilient, self-serve data systems that support analysts, researchers, and AI initiatives. She often shares her experiences at industry conferences, discussing practical lessons in data strategy, governance, and building trust in AI.

She believes that great data platforms don’t just move information, but also build confidence. Anusha holds an MBA from the University of Chicago Booth School of Business, with a concentration in strategy, entrepreneurship, and finance.