The Hidden Costs of Poor Data Quality in AI, From Bias to Context

Watch the full session · 1 hr

In a Data Science Connect webinar, experts from Revefi, Snowplow, Informatica and Fivetran explained how poor data quality undermines AI. They covered silent data failures, unrepresentative data, the shift to data and context quality, and how to win executive buy-in for data governance.

Poor data quality quietly undermines AI accuracy, scalability and return on investment, and the problem now reaches beyond the data itself into the context models receive. That was the message of a Data Science Connect expert panel webinar moderated by Brian Mink, co-founder and president of Data Science Connect, and sponsored by Fivetran, Informatica, Revefi, SAP and Snowplow.

Panelists were Girish Bhat, SVP at Revefi; John Mallaloy, senior technical account manager at Snowplow; Nate Rain, data governance, quality and privacy specialist at Informatica; and Kelly Kiful, senior global director of partner sales engineering at Fivetran.

Why is data quality so important to AI success?

Rain said data scientists often spend 60% to 80% of their time preparing data, which is not what they were hired to do. Mallaloy described the two bad choices poor data leaves: spend many hours cleaning it, or trust pipeline outputs that contain hidden biases and make business decisions before anyone notices.

Bhat said AI raises the stakes compared with business intelligence. A wrong number on a dashboard does limited damage, but a model used by thousands of people every second "acts as an amplifier of bad data," he said.

Perfect data everywhere is unrealistic, Kiful argued. Getting every one of the thousands of data sets in a large enterprise perfect "will never happen." A focused AI application only needs the right data and context, and higher-stakes uses such as financial or risk decisions need a higher quality bar than others.

What data quality issues hurt AI performance most?

For Kiful, the biggest problem now is context. Building AI applications with MCP servers that span multiple data tools has meant struggling to load enough data and tools without exhausting the model's context window before it can answer. "This data quality thing has kind of expanded into data and context quality," Kiful said.

Mallaloy's biggest issue is missing data. "What data do I not have? And data that I might not have that I think I do," he said. Users who opt out of data collection, for example, can leave a data set unrepresentative without anyone realizing.

Bhat pointed to silent data failures, in which a pipeline runs successfully and dashboards show green while the data is wrong. He also cited data timeliness, especially in regulated industries with legacy systems, and data bias.

Rain said the classic problems remain: missing and outdated values, inconsistent formats across sources, duplicate records and subtle biases. Fixing data where it is used, instead of at the source with the people who enter it, means the same problem returns tomorrow.

How do you get executive buy-in for data governance?

Rain said a focused use case is what worked for him. As a head of BI, he profiled one problem domain's data after its business owner said she was confident it was good. The results, including errors entered the day before, changed the conversation. "Data quality is a hook for data governance," he said. The key is finding the organization's pain and getting C-suite agreement that fixing it adds value.

Mallaloy said setting a governance strategy before collecting data takes time and can delay a first proof of concept, but skipping it means discovering problems after costs are already sunk. Bhat recommended the NIST risk management framework, with its govern, map, measure and manage functions, as a vendor-neutral starting point.

How do biases in data distort AI results?

Mallaloy described a customer whose AI analysis showed extremely engaged users, but new content based on those results performed poorly. The training data came only from users who had chosen to share their data, a small and unrepresentative group. Nothing in the warehouse table revealed that, he said.

Kiful said bias can enter through system instructions, tools, semantic views, memory and chat history as well as data, so all of them need alignment.

Can AI check data quality on its own?

Automation is necessary for scale, Rain said, but human judgment remains essential. "I want my AI model to know enough to doubt itself," he said, adding that AI won't do that on its own.

Mallaloy said AI can find nulls and hidden bad characters, but organizations must agree on definitions, such as what eight product views actually mean. Rain recalled a hospital system where finance and clinical leaders reported different average lengths of stay because their denominators differed, while a regulator already defined the term. Checking for an existing regulatory definition is a good standard to start from, he said.

To enforce standards across teams, Kiful recommended templatizing an approach for data and context quality on one use case, then repeating it. Bhat stressed collaboration, automation and modern platforms. Mallaloy said dedicated data governance roles pay off. "You don't need a lot of people," he said, noting that a two- or three-person team can drive significant value.

Catch the next expert panel live

Data Science Connect has 24 free live webinars on the 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Sessions like this one bring enterprise practitioners and vendors together to take audience questions live. See the schedule at webinars.

Watch: The Hidden Costs of Poor Data Quality in AI on YouTube
Event: Webinars · More coverage: Webinars insights