GenAI makes data quality harder because teams must manage unstructured inputs, prompts, and model outputs.
2
Moving a GenAI proof of concept into production requires checking document recency, privacy, access, costs, and human review.
3
LLMs can help with pipeline code, testing, synthetic data, cataloguing, and data classification, while engineers still need guard rails and judgment.
Summary
Anu Arora and Anass Bensrhir explain why data engineering remains difficult as companies adopt GenAI. Traditional data systems already had methods for measuring quality, building pipelines, and tracking lineage. GenAI adds documents, emails, videos, vector databases, retrieval systems, prompts, and model outputs that also need to be managed. A proof of concept can work while using an old contract or an outdated policy, then produce harmful answers in production. The guests recommend treating quality as an input, prompt, and output problem, with checks for recency, privacy, and access. They also discuss LLM support for pipeline development, unit tests, synthetic data, cataloguing, and data classification. For deployment, Anass recommends ranking use cases by value, feasibility, and risk, then starting with limited users and narrow intents. Both guests argue that anonymisation, controlled hosting, authorisation, and human review remain necessary.
GenAI makes unstructured data management a production problem
Anass says data engineering has always connected data to models, but GenAI expands the work beyond structured tables. Teams now use documents, email servers, and video storage, then need to retrieve, transform, store, and assess that material. Existing data lakes and data-management investments remain useful. The additions include vector databases for retrieval-augmented generation, integration with language models, and ways to measure quality and recency in unstructured sources. A PDF raises a different quality question from a website event because the system must determine whether its contents were parsed correctly and whether the document is still current.
Anu explains that large documents need preprocessing before a model can use them. A PDF is broken into chunks, otherwise the cost of sending it to an LLM can become very high. Quality checks can look for missing values during preprocessing, validate interpreted content with prompts, and check for sensitive or personally identifiable information. Anass adds that quality needs to be assessed across the input, the prompt and context, and the output. A chatbot can produce an answer even when the underlying information is absent or wrong, so the model response also needs checking.
Production systems must verify that retrieved documents are current
Anass gives an insurance example involving long business contracts. A proof of concept used an LLM to compare a claim with a policy and help an insurance agent decide whether the claim was covered. The prototype worked when tested, but the production question was whether the scanned contract in document storage was the latest version. Using a two-year-old policy could lead the system to miss a newer cover and create serious consequences. The team therefore had to establish that the document was current, correct, and tied to the right customer before relying on its answer.
LLMs can speed up data work without replacing data engineers
Anu lists several uses for GenAI in data engineering. It can help write data pipelines, generate unit tests, create synthetic data for integration testing, classify personally identifiable information, and support data cataloguing. Anass describes using natural language to find data in a catalogue, identify relevant tables, suggest joins, and produce queries for users who do not know SQL. LLMs can also inspect tables for possible data-quality problems and help trace lineage. The guests still require human oversight because generated code can raise intellectual-property concerns, and models can hallucinate answers.
Privacy starts with deciding what data the model actually needs
Anu advises teams to ask whether an application really needs personal data before sending anything to a model. In most cases, customer IDs, customer numbers, or bank-card details can be anonymised, tokenised, or hashed. If sensitive data must be sent, teams need to know where the LLM is hosted, where the data will travel, and which users and applications can access it. Anass describes alternatives for banks, including keeping company data in an internally hosted vector database while using a GenAI system for interaction and retrieval. Offline and hybrid models are also options when the data must remain on company premises.
Teams should limit early releases by risk and intent
Anass recommends ranking possible use cases by value, feasibility, and risk. A company might begin with an internal HR chatbot that answers a narrow set of questions, then expand its scope under control. Another approach is to release a product to a limited user group, learn from failures, and build from there. The team needs to validate the business value, understand the technology and its guard rails, and measure the cost. Anu summarises the controls as privacy, quality, and access. Anass also expects human review to remain part of the process while organisations are still learning how to operate these systems.
GenAI cost savings need to be proven for each use case
Anass says the expected benefit differs by application. A team might be trying to serve many customers, retrieve information faster, or optimise another specific outcome. Each use case needs a business value and a cost assessment before significant engineering work begins. He warns that companies can spend heavily building a capability that data-management vendors later offer off the shelf. His suggested sequence is to rank use cases, start with high-value and feasible work, test uptake, measure the cost, and then move toward more complex or higher-risk applications.
"I think data quality is way more important because we are basically touching and using unstructured data."Anass Bensrhir08:06
Who should watch
You are taking a GenAI prototype that searches documents or answers questions toward production and need to deal with stale sources, parsing errors, and output checks.
Your data platform team is deciding where LLMs can help with pipelines, testing, cataloguing, or synthetic data without handing over engineering judgment.
You work with regulated or sensitive data and need practical options for anonymisation, internal hosting, access control, and limited user rollouts.
