About This Episode
In episode 17 of Data & AI Heads, host Ian Allison sits down with data quality specialist Susan Walsh to discuss the realities of cleaning messy enterprise data, the genuine promise and current shortcomings of generative AI in data quality workflows, and the practical techniques that separate expert data cleaners from well-intentioned amateurs. Susan draws on nearly eight years of hands-on work, primarily with procurement and spend analytics teams, to share concrete stories of supplier name deduplication, address normalization gone wrong, and the downstream business value that rigorous data quality unlocks.
The conversation covers the upcoming second edition of Susan's book "Between the Spreadsheets," which adds new chapters on generative AI with real ChatGPT examples showing both successes and failures, including hallucinated addresses and improperly normalized supplier names. Susan also introduces Samification, a tool she has built that uses private GPT models trained on verified, human-cleaned data to automate supplier name classification, and previews her forthcoming sales and marketing data cleaning book.
For data and AI leaders, this episode is a grounded, practitioner-level reality check: AI initiatives require investment in data quality upfront, data cleaning is a specialist skill that is undervalued and underrepresented in formal education, and the human element in data quality work is not going away anytime soon.
Key Takeaways
- Generative AI tools like ChatGPT currently struggle with core data quality tasks such as address normalization and supplier name standardization, producing hallucinated addresses and inconsistent formats when used with everyday prompts.
- Investing in data quality at the start of an AI initiative reduces costly rework later, since data scientists already spend 30 to 60 percent of their time on data engineering and cleaning rather than model development.
- Never overwrite original data columns: always create a separate clean column alongside the source data so you can benchmark progress, demonstrate value, and recover from errors.
- Standardization rules must be agreed upon before a data cleaning project begins, covering decisions such as abbreviation conventions, casing, date formats, phone number formatting, and which fields to collect, because retrofitting standards midway through is far more expensive.
- Always maintain a backup of the original dataset in a separate location before any cleaning work starts, because accidental overwrites or format conversions, especially with dates across regional Excel settings, can corrupt days of work irreversibly.
- Deduplication of supplier or customer records must come after data cleaning, not before, because only clean, normalized data reveals that multiple differently spelled records represent the same entity.
- Data quality work creates measurable business value, including enabling AI tooling, justifying new headcount, revealing off-contract spend, surfacing supplier risk, and preventing duplicate outreach in CRM systems, even when clients rarely report those outcomes back to the consultant.
- The data quality specialism is not being automated away: Susan built Samification, a tool using private GPT models trained on her own verified cleaned data, to improve automation while preserving the accuracy that only expert-curated training data provides.
What We Cover
Frequently Asked Questions
Is ChatGPT reliable for cleaning and normalizing data quality tasks like addresses and supplier names?
According to Susan Walsh, ChatGPT is not yet reliable for core data quality tasks when used with everyday prompts. In her testing, it hallucinated addresses when the correct address was available for comparison and failed to normalize supplier names properly. She emphasizes that results improve with detailed prompt engineering, but most users are not working at that level of sophistication in day-to-day operations.
What percentage of a data scientist's time is typically spent on data cleaning and engineering rather than actual modeling?
Susan Walsh notes that data scientists commonly spend 30 to 60 percent of their time on data engineering alone, with additional time dedicated to data quality work. This leaves a relatively small proportion of their time for the modeling and analysis they are primarily hired to perform, making upfront investment in dedicated data quality expertise a significant efficiency lever.
What are the top practical tips for cleaning CRM address data?
Susan Walsh recommends three foundational practices. First, never overwrite original data columns: place cleaned data in a new adjacent column so you can benchmark against the source and demonstrate value. Second, agree on all standardization rules before starting, including abbreviation conventions, casing, date and phone number formats, and required fields. Third, always keep a backup of the original dataset in a completely separate location before any cleaning begins, as accidental overwrites or regional format conversions can cause irreversible data loss.
Why should deduplication happen after data cleaning rather than before?
Susan Walsh explains that deduplication depends on records being normalized first. Until supplier names, addresses, and other identifiers are cleaned and standardized, the same entity can appear under multiple different spellings and formats, making it impossible to reliably identify duplicates. Only after cleaning can you confidently consolidate records and recognize that, for example, IBM and International Business Machines refer to the same supplier.
What is Samification and how does it use AI for data quality?
Samification is a tool developed by Susan Walsh that uses private GPT models trained on data she and her team have already cleaned and verified. Because the training data comes from expert-curated, high-quality sources rather than general internet data, the models can classify and normalize supplier names with greater accuracy. The tool started as an internal resource and is now publicly available at Samification.com, where users can test up to 250 supplier company names for free.
How does poor data quality affect AI initiatives and what should organizations budget for?
Susan Walsh and Ian Allison agree that poor data quality creates compounding problems in AI projects: it inflates the time data scientists spend on remediation rather than modeling, introduces errors that are far more expensive to fix mid-project than at the start, and undermines the reliability of model outputs. They recommend that a meaningful portion of any AI initiative budget be explicitly allocated to data quality work upfront, treating it as a prerequisite rather than an afterthought.
Will AI eventually replace the need for human data quality specialists?
Susan Walsh does not believe AI will eliminate the need for experienced data quality practitioners. She founded her business eight years ago expecting it to become redundant and has found the opposite to be true. While AI will reduce the volume of manual cleaning work and help standardize data inputs, she argues there will always be edge cases, context-dependent judgment calls, and training data curation tasks that require human expertise. She compares the specialism to skilled crafts that persist alongside industrial automation.
About the Guest
Susan Walsh
Susan Walsh is a data quality specialist and consultant with nearly eight years of experience working primarily with procurement and spend analytics teams. She focuses on the most complex data cleaning challenges, including supplier name normalization, spend classification using taxonomies such as UNSPSC, address deduplication, and contract-to-supplier mapping. She is the author of "Between the Spreadsheets," a practical data quality book now in its second edition with new content on generative AI, and is completing a forthcoming book on sales and marketing data cleaning. Susan is also the founder of Samification, a supplier name classification tool that uses private GPT models trained on expert-verified data. She is a recognized speaker who appears at data and procurement industry events and brings a practitioner-first perspective to data quality, emphasizing methodology, standards, and the irreplaceable value of specialist human judgment.
Connect on LinkedIn →More Episodes