How ‘Watson Era’ Thinking Is Holding Back Banks’ AI Benefits
By Alex Jiménez , managing director, consulting, at Finalytics.AI
Simple Subscribe
Subscribe Now!
In 2011, IBM’s Watson beat two Jeopardy champions on national television. The victory was the product of years of investment and some of the most advanced natural language technology of its time.
Banks paid attention. Executives who had never talked about artificial intelligence in a board meeting were suddenly asking their teams what Watson could do for underwriting, for fraud detection, for customer service.
Then bankers stopped talking about it.
By the time most people started talking about AI again, ChatGPT drove the conversation, not Watson. An entire decade of banking’s first serious AI moment came and went, and most people in the industry today never knew why.
Key point: That reason matters again now, with a new wave of AI excitement underway.
Need to Know:
- IBM’s Watson required massive massaging of databases in order to be useful. Bank data often defied the process.
- GenAI, by contrast to the original Watson, can handle data without such extensive prep and cleanup.
- Banks that view today’s GenAI through the same lens as Watson will lose the potential benefits of adopting the new tools.
What Watson Needed that Doomed Its Influence on Banking
Calling Watson a failure gives the wrong impression. It won on national television against two of the best Jeopardy players alive, and IBM built a real business around it afterward, selling Watson into healthcare, retail and other industries for years.
Banks hit the problem when they tried to move that same approach into their own operations.
Key insight: Watson needed specific, hand-written rules, built by people who understood both the technology and the problem it was solving. It needed the data cleaned and organized before it could answer a question well.
Think of it as a librarian who could answer any question you asked — but only after someone had already sorted every book in the library, labeled every shelf, and cross-referenced every entry.
Banks don’t have libraries like that. They have core systems built decades apart, data warehouses that don’t agree with data lakes, and product names that mean different things in different departments.
Pymnts Intelligence’s “Core Strength” report, done with Galileo, found that 75% of banks struggle to roll out new digital solutions because of their legacy infrastructure.
Sorting out all of that by hand wasn’t something most banks could pull off, and that’s why bankers moved on. Banks needed an approach built for the data they already had, not a decade of manual cleanup first.
Key insight: Some bankers, and even some bank technologists, still resist newer AI approaches for the same reason banks backed off Watson: They assume it requires the same setup, hand-built rules, and scrubbed data before anything useful can happen.
That belief made sense a decade ago. It doesn’t anymore.
Read more: AI Is the Answer for the Banking Industry. But It’s Also the Problem
What’s Different about AI Today
Generative AI doesn’t need the shelves sorted first. It can work with the mess directly, because it already understands the structure of language and how ideas connect to each other before it ever sees your specific data. That’s the real distinction, though vendor vocabulary buries it fast.
Late last year, before joining Finalytics, I built a proposed structure for moving a large asset-based loan application and approval process from paper to a fully agent-driven workflow.
The existing process ran on paper files, scanned documents, inconsistent formats, and years of accumulated exceptions. None of it was clean. None of it was standardized.
The system I designed didn’t need any of that. It could read the paper trail as it existed, route pieces of the application to the right specialized function, and flag what needed a person to review.
Different parts handled document extraction, checked figures against source records, and flagged anything that didn’t fit the expected pattern. No one had to spend a year cleaning the archive first.
Key insight: Watson needed the mess solved before it could help. Generative AI can help while the mess is still there.
Read more: How Financial Institutions Must Update Data Strategies for the Next Wave of AI
Where the Real Work is Now
Watson’s limit was technical. The bank’s limit in this story wasn’t.
The proposed loan system was designed to work. The bank got scared anyway.
A pilot could have proven the system out. Run it against real applications and expand once the results held up.
Take one step as an example. One agent could read a financial statement, paper or digital, and enter the figures directly, what underwriters call “spreading it.” That took milliseconds.
Built into that step was a human check that sampled the agent’s output on a regular basis, watching for drift or hallucination rather than reviewing every record. That sampling took a few minutes here and there, nowhere near the hours the manual process took, and it never needed to touch every loan to do its job.
The bank’s actual process had someone enter the data by hand, then upload it again into a second system, a job that took several hours with no equivalent check at all. That was one step out of many.
Multiplied across the full application, the process could have gone from weeks to days, with real savings behind it. Instead, the bank applied older technology to a couple of steps in the same process. The timeline dropped by a couple of days. The savings were close to nothing.
Key insight: The bank hesitated because of trust, not technology, and it’s not alone.
A 2026 Grant Thornton survey found that half of banking executives say governance and compliance concerns are already limiting how well their AI performs, and only 18% feel confident their AI controls could pass an independent audit.
That gap is about readiness, not capability, the same distinction the bank in this story never worked through.
Read more: How to Succeed with Agentic AI: Give It Major Tasks, But Don’t Hand It Entire Jobs
Understanding the Appropriate Role of AI Governance
A study by Columbia University and FS-ISAC show what that readiness work looks like. Researchers interviewed U.S. bank executives in 2025 and found the disciplined ones operate on what one executive called a “trust, but verify” philosophy: pilots stay narrow, oversight stays heavy, and governance gets built in before a system ever reaches a client.
Key insight: That’s the harder path the bank in this story skipped entirely, in favor of no pilot and no proof at all.
A system that can read messy, inconsistent, decades-old paper records and make sense of them fast is impressive right up until someone asks how it decided, and whether that call can be defended to an examiner.
Banks can’t treat governance as an afterthought once a system is making real recommendations. Every recommendation needs a trail back to why the system made it, and every handoff to a human needs enough context to act on, not just a transcript to sort through.
None of that must slow the system down. Banks just need to build the record from the start, not bolt it on after something goes wrong.
Read more: Retail Bankers Are Adopting AI for All the Wrong Reasons
What Banks Must Do Differently
The lesson is simple. Watson needed the mess solved before it could help, and most banks couldn’t solve it fast enough to justify the investment. GenAI doesn’t share that problem. It reads the mess directly.
The bank in this story avoided the work real caution requires, and called it caution anyway. A real pilot, with sampled human review built in, would have given the bank both proof and speed. Instead, the bank picked older technology, kept the same lack of proof it claimed to worry about, and gave up the savings too.
That’s cost, not risk management.
Banks often confuse the two. Avoidance feels like governance, but it isn’t governance unless someone can point to the audit plan behind it. A board that blocks an AI pilot without building a way to verify it is paying for comfort, not protection.
Read next: Who Will Protect Banking Consumers’ Rights in the Age of AI?
