Narmi founder: ‘Don’t wait for AI to be perfect’

Michiel Willems
August 11, 2026

Banks are moving artificial intelligence into customer service, account opening and financial decision-making, promising faster responses and fewer manual checks.

But as adoption accelerates, QA teams face a more difficult question: how can they prove these systems remove friction without introducing unreliable guidance, biased outcomes or an additional layer of work?

Testing AI in banking requires more than measuring whether a model produces an accurate answer. Financial institutions must be able to explain how recommendations were reached, identify when models begin to drift and demonstrate to regulators that every important decision can be reconstructed.

They also need fallback processes that allow essential banking services to continue when an AI tool fails or becomes unavailable.

New York City-based Nikhil Lakhanpal, co-founder of digital banking technology provider Narmi, believes governance, auditability and human oversight must therefore be built into testing from the outset.

“QA teams should first focus on governance as a key pillar. Are audit trails recorded? Can an auditor get a full report of what decisions were made and by what piece of software? This will be critical during governance exams,” he told QA Financial in an exclusive interview.

Testing whether AI creates more work

Lakhanpal warned that banks should not assume the presence of AI automatically results in greater efficiency. Testing must establish whether the technology genuinely reduces the workload for employees and customers.

“Next, financial institution teams should avoid the term ‘AI Slop’. For example, if a customer messages their bank and the AI drafts a reply, does staff have to rewrite the whole thing? If so, that’s actually creating more work than automating it.”

That assessment requires QA teams to look beyond headline accuracy rates and examine how employees use, correct and override AI-generated material in real workflows.

Narmi applies a similar test to explainability when developing its own AI functionality, according to Lakhanpal.

“Honestly, one way we test AI functionality is if a regulator walks into a client’s office tomorrow and asks ‘why did you approve this?’, Narmi’s AI needs to show, in plain English, a full audit trail.”

Biased outcomes

Human intervention is also an important source of testing data. When employees repeatedly reject or amend recommendations, those overrides may reveal defects, poor calibration or patterns of unfair treatment.

“Building off AI DecisionAssist, we track how often staff override the AI’s recommendation and why. If a certain type of application keeps getting flagged unfairly, that pattern will show up, and it’s on us and the institution to catch it and fix it.”

“QA teams should first focus on governance as a key pillar.”
– Nikhil Lakhanpal

Lakhanpal stressed that AI should support financial decisions without becoming their final authority.

“And no matter how good the model gets, a human makes the final call on anything that touches a customer’s finances.”

The range of users involved in testing matters too. AI services can perform differently for customers depending on age, disability, digital confidence and familiarity with banking technology.

“We encourage financial institutions to test with outside consumers and businesses, not just the internal team who already knows how the product works,” Lakhanpal said.

“The good news is that user-facing AI can be restricted to certain segments based on the financial institution’s setup.”

Models cannot be left unmonitored

Pre-deployment testing alone cannot account for changes to models, prompts, data or customer behaviour. Lakhanpal argued that financial institutions must continue evaluating AI after it enters production.

“We also keep testing after launch because a model that’s calibrated correctly today can quickly drift if you’re not watching it. We also make it a priority to actively speak to our customers who have deployed AI tools in production to get feedback.”

That continuous-testing principle applies whenever Narmi updates its technology.

“We don’t believe in ‘ship it and walk away’ with any part of our product, ever. Every update gets tested before it goes live, and we watch how staff actually respond to the AI’s recommendations after it’s live.”

“Given how new this technology is, it’s important to stay close with those clients using AI in production.”

Banks need an AI failure plan

AI must also form part of banks’ digital-resilience planning. Institutions need defined fallback processes for situations in which an AI service fails, becomes unavailable or begins producing output that cannot be trusted.

“Even with fully autonomous agents, control should never shift away from the financial institution,” Lakhanpal said.

“If the AI goes down entirely, your staff is alerted immediately, and serves as the final safeguard before anything reaches a user, while the underlying manual processes continue to operate.”

Audit trails then become essential not only for regulatory reviews but also for incident investigation and remediation.

“And when something does go wrong, you should be able to pull up exactly what happened, what the AI recommended, and what your team did about it.”

Despite those risks, Lakhanpal cautioned banks against delaying adoption until the technology is flawless. He recommended beginning with tightly defined operational problems and introducing safeguards proportionate to the potential impact.

“Don’t wait for AI to be perfect before you let it help your team. Pick one or two real problems your staff deals with every single day, and go solve that one thing well, with the right guardrails in place. That’s how you build trust in this technology instead of fear of it.”

“We’re already seeing our clients’ manual reviews faster and more consistently because of AI DecisionAssist. My personal hope is that a Narmi client can review applications just as fast and just as thoroughly as one 50 times its size.”

“That’s the whole point of practical AI, and we’re building toward it with our clients every day,” Lakhanpal concluded.