AI in Finance #005 — Data in AI #001: Data Is the Hurdle
Originally published in Weekend Reads on LinkedIn · Subscribe
As companies are moving from deploying pilot projects to full-scale enterprise AI adoption, it is natural for such organizations to face certain hurdles.
One such hurdle is " DATA ". Data has been a difficult thing to manage, specifically in the BFSI domain. For instance, millions of trades across geographies, and the numerous platforms executing these trades, generate a plethora of data. Maintaining client data in wealth management firms across various types of accounts, and catering to the different needs of each client, is another example of how systems can become complex and generate multidimensional data.
I have worked on various projects, from client-facing trading platforms to the back office, where all trades are captured and maintained. I have also worked for clients in asset management and wealth management. Creating data, managing it, and reconciling it whenever new functionality is added to a system has always been a difficult task. This problem becomes magnified when dealing with legacy systems or applications that have been in place for several years.
Another challenge comes from the mandatory changes introduced by regulatory bodies. Sometimes these changes are forward-looking, but at times they have to be applied retrospectively.
These challenges are already there. Now add another layer over them — AI.
Many organizations adopting AI are going to face, or are already facing, challenges related to data. Gartner estimates that poor data quality costs organizations an average of $12.9 million a year — a figure drawn from its 2020 research across 154 enterprises already mature enough to be buying data-quality software. More pointed for anyone funding an AI program: Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data. Not because the models were weak — because the data wasn't ready to feed them.
LLMs are data-hungry. Data is their food. On the surface, AI adoption seems very simple: choose the best model (this is where most organizations focus), give it the problem, and it will magically fix everything and produce intelligent outputs.
In reality, an LLM trained on incomplete, stale, biased, or mislabeled data will learn incomplete, stale, biased, or mislabeled patterns — faithfully, at scale, and in the confident tone that makes them dangerous.
In the finance domain, this is not just a problem — it is what corrupts the image of the company and opens it to lawsuits, because in finance the data is the business. A credit decision, a suitability check, a surveillance alert, a margin call — each is a data artifact before it is anything else. When the inputs drift, the intelligence built on top of them drifts with them, silently.
I came across the example below, where bad data caused losses of millions of dollars:
The clearest illustration isn't an exotic neural network — it's a credit-scoring pipeline, the most established algorithmic decision system in finance. In spring 2022, a coding change in a legacy Equifax server being migrated to the cloud corrupted score calculations over a roughly three-week window. The scoring logic wasn't the problem. The data flowing through it was. The consequences were entirely downstream and entirely real. By Equifax's own account, fewer than 300,000 consumers saw their score shift by 25 points or more — enough to change an interest rate or flip an approval to a denial. Incorrect scores went to major lenders, and reporting identified institutions including JPMorgan Chase and Wells Fargo among the recipients, affecting applications for mortgages, auto loans, and credit cards. Equifax now faces a proposed $100 million settlement of the resulting class action — granted preliminary court approval in August 2026, with a final fairness hearing set for January 2027 and Equifax denying wrongdoing — described by plaintiffs' counsel as the largest in the history of the Fair Credit Reporting Act.
The lesson generalizes directly to AI: a mature, well-understood decision engine produced wrong real-world outcomes at scale, not because the algorithm failed, but because the data feeding it was wrong and no one caught it in time. Every AI system you deploy inherits this exposure — and amplifies it, because AI acts on more data, faster, with less human inspection per decision.
In the same Gartner research, 63% of organizations said they either lack, or aren't sure they have, the right data-management practices for AI. That is not a technology gap so much as an ownership gap — a question of who is accountable for the data an AI system consumes, answered before the first line of model code is written.
For a regulated firm, the stakes are sharper than lost efficiency. Wrong data doesn't just degrade a model's accuracy; it manufactures decisions you are legally answerable for. A frequently cited 2016 estimate, published in Harvard Business Review, put the annual cost of bad data to the US economy at $3.1 trillion — a number now dated and questioned on methodology, but useful as an order-of-magnitude reminder that this problem is structural, not incidental.
The industry keeps hunting for a better model when the constraint is almost always a better dataset. Reliable AI in finance is built the way any sound structure is: foundation first, verified before you build on it. That is what behaving by design rather than by hope actually requires — not a smarter model hoped into good behavior, but a data foundation engineered to earn trust.
Subscribe and follow this space as we explore why data in AI matters, how it should be governed, and everything else that goes into making AI solutions work efficiently and within regulatory boundaries — by respecting the data underneath them.
References
Gartner, Data Quality: Why It Matters (drawing on the 2020 Magic Quadrant for Data Quality Solutions, 154 reference customers) — gartner.com/en/data-analytics/topics/data-quality
Gartner press release, Lack of AI-Ready Data Puts AI Projects at Risk, 26 Feb 2025 — gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
Equifax Newsroom, Equifax Statement on Recent Coding Issue / Credit Scoring and the Equifax Coding Issue, Aug 2022 — equifax.com/newsroom
Reporting on the proposed $100M Equifax settlement, Aug 2026 (WSB-TV / court filings) - https://www.wsbtv.com/news/local/atlanta/equifax-settlement-consumers-impacted-by-2022-credit-score-error-could-receive-payments/IFCDD7H6OJEABG7X45NPRZUQUM/
T. Redman, Harvard Business Review (2016), $3.1T bad-data estimate — cited with methodology caveat - https://hbr.org/2016/09/bad-data-costs-the-u-s-3-trillion-per-year
Data in AI
Have thoughts on this?
Let's continue the conversation on LinkedIn.
Get new posts by email
Occasional notes on the BA → AI PM journey. No spam.