Sagar Gupta, computer scientist and tech executive with 25+ years of experience in enterprise systems and digital transformation using AI
Artificial intelligence (AI) dominates business conversations. Companies invest heavily in models that improve customer experience, forecast revenue, detect fraud or automate operations. Yet one foundational issue often receives less attention: whether the data feeding these systems is reliable.
Sophisticated models can still produce misleading results when the underlying data is incomplete, outdated or incorrectly mapped. A more powerful model may simply generate wrong recommendations faster. Technology leaders should ask not only whether a model can perform a task, but whether the data is trustworthy enough to support the decisions expected from it.
Measure data quality against business needs.
Data quality is frequently treated as a technical checklist—removing duplicates, filling gaps and standardizing formats. These steps matter, but they do not guarantee reliability for every business use.
Consider revenue forecasting. An AI system may draw on sales opportunities, customer history, contract dates, inventory and project schedules. The data may look complete, yet if opportunity stages are outdated, installation timelines are stale or revenue classifications shifted during an ERP migration, the forecast can diverge sharply from reality. The system may show no technical errors while the information remains unsuitable for the decision.
Reliability must be evaluated against the specific business requirement. Daily updates may suffice for monthly planning but fall short for real-time inventory allocation or fraud detection. Before deploying data in an AI application, leaders should ask: Is it complete and accurate across systems? How recent is it? Has its meaning changed? What happens if the model decides on incorrect information?
The goal is not perfect data—an often impossible standard—but data sufficiently reliable for the decision it must support.
Automate carefully.
Automation can detect duplicates, fill missing values and flag outliers. Assuming every anomaly is an error, however, carries risk. An unusually large customer order may be flagged and adjusted, distorting the revenue forecast. Missing values replaced by averages may conceal process failures or genuine exceptions.
Clear guidelines are needed: Automatic correction is appropriate for verifiable, low-impact issues such as date formats. When changes affect financial reporting, customer classification, risk or planning, business owners should intervene. When accuracy cannot be confirmed, excluding the record may be safer than inventing plausible data. Automation works best inside defined business controls. Sometimes admitting insufficient reliable information is better than manufacturing the appearance of accuracy.
Know where data comes from.
Enterprise data often travels through multiple systems—from CRM to ERP to integration platform to warehouse—being transformed and reclassified along the way. By the time it reaches an AI model, it may differ significantly from the original transaction.
When a forecast surprises management, teams must be able to trace the information to its source. Without lineage—origin, update frequency, transformations, mapping changes and processing history—it becomes difficult to distinguish model problems from data or process issues. For critical applications, organizations should be able to recreate the exact dataset used in an analysis. These practices prove invaluable when results go wrong. A reliable AI system produces answers the organization can also explain.
Reliability changes even when models do not.
Data is not static. New products, process changes, evolving data entry habits and system upgrades continually alter it. A minor shift in CRM opportunity classification or ERP field mapping can degrade model performance while the application continues running without errors.
Validation cannot be a one-time implementation task. Continuous monitoring for pattern shifts, missing data, schema changes and integration failures is essential, as is tracking whether the model still delivers consistent business outcomes. A sudden drop in forecast accuracy after an ERP upgrade may point first to changed data definitions rather than the need to retrain the model on the same flawed inputs. AI governance must evolve with the business systems that support it.
Start small.
Years of accumulated quality issues across ERP, CRM and finance systems can make a comprehensive cleanup an overwhelming project. A practical path begins with one high-value use case, such as revenue forecasting. Identify the critical data elements—opportunity values, close probabilities, contract status, invoicing history, delivery schedules—then assess completeness, accuracy, consistency and freshness. Agree with stakeholders on acceptable thresholds and exception handling. After controls are in place, measure results against prior performance: improved accuracy, fewer surprises, less manual verification, greater decision confidence.
Business outcomes matter more than cleaned-data percentages. Success with one focused case creates a framework that can expand to other initiatives.
Confidence in AI begins with confidence in data.
More powerful models and automation will continue to appear. Advanced technology alone does not guarantee better decisions. Organizations that understand their data, maintain visibility into its processing and continually assess its reliability will extract greater value from AI.
For technology leaders planning the next investment, start with practical questions: Which decisions will the system influence? Which data elements matter most? Who owns the information, what reliability level is acceptable and how will exceptions be handled? Ensure the organization can trace data to its source and detect when system changes affect reliability. Measure success by improved business decisions, not solely by model metrics or processing efficiency.
Even the smartest AI model is only as dependable as the data and business understanding behind it.
Disclosure: The author is the applicant and inventor of a pending patent application related to adaptive data processing, automated data-quality detection, feature reliability scoring and model-ready dataset generation.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?