5 things you should know about AI in banking for 2026: insights from a recent panel of experts
AI this. AI that. We see it and hear it all over the place.
The noise is deafening. But while the industry buzzes with hype, most banks are still watching from the shore, because their technology simply won't let them move.
At ENGAGE London, Backbase CMO Tim Rutten sat down with three leaders actively doing the real-life work: Marco Martinez (Microsoft), Kanishka Bhattacharya (Bain & Company), and Dave Murphy (Publicis Sapient).
They talked about the hard, messy, real work of deploying AI in banking right now. The code that nobody knows how to read anymore. The cost of talent versus technology. The shift from helpful chatbots to agents that can do the work.
The panel agreed: The time for "wait and see" is over.
Here are the five realities you need to take into 2026 if you want to scale AI in your bank.
1. Legacy code is no longer an excuse
Generative AI can now translate and analyze legacy banking code that's been undocumented for decades. Large Language Models ingest ancient mainframe code and output human-readable functional specifications, finally making legacy modernization economically viable.
We all know the story. You run on mainframes written 30 or 40 years ago. The documentation - if it ever existed - is gone.
Dave Murphy highlighted this massive unlock. Generative AI has changed the economics of modernization.
Here's what AI-powered legacy modernization delivers:
- Code translation: Convert undocumented mainframes into readable specifications
- Dependency mapping: Understand system connections without guessing
- Risk reduction: Know exactly what breaks before you touch it
- Cost efficiency: Replace "patch and pray" with strategic modernization
This is a literal breakthrough.
You can finally understand what your core is actually doing. You can document the undocumented. Once you understand the code, you can dismantle it and move from a "keep the lights on" strategy to a true modernization roadmap.
The shackles of technical debt are heavy. But for the first time in history, we have the key to unlock them and wave goodbye to the legacy constraints.
2. Unstructured data is your biggest untapped asset
Banks are sitting on mountains of data they can't use.
You have millions of hours of call center recordings. You have servers full of PDFs, contracts, and scanned documents. In the old world, this was "unstructured data" - messy, dark, requiring armies of humans to manually review it file by file.
Martinez pointed out that this is where the immediate ROI lives.
Take compliance as a prime example. Traditional manual review checks maybe 3% of call center recordings and hopes the other 97% are compliant.
AI changes everything:
- Call recordings: 100% compliance checking instead of 3% sampling
- Trade finance: Instant jurisdiction compliance instead of hours of manual review
- Document analysis: Every PDF, contract, and email gets interrogated in seconds
You can feed every single call recording, every text interaction, and every email into the model. You can ask, "Show me every instance where an agent failed to read the disclosure statement," and get an answer in seconds.
The data is the same, but your ability to use it has changed completely.
3. Stop building chatbots. Start building agents.
Most banking chatbots are glorified FAQ pages with a search function. You ask a question, and they give you a generic answer with a link to a policy document. They deflect problems instead of solving them.
Murphy warned against falling into the "optimization trap." Using AI to build a "more sophisticated assistant" might get you a 10% or 15% efficiency bump, but it's missing the point.
The real shift - the one that changes the industry - is the move to Agentic AI.
We're moving from software that retrieves information to software that executes tasks.
The difference is execution:
- Chatbot: Tells you how to apply for a mortgage
- Agent: Applies for the mortgage for you
- Chatbot: Shows you account balance
- Agent: Moves money between accounts
- Chatbot: Explains loan terms
- Agent: Processes the loan application
Martinez shared a telling story about his five-year-old daughter. She tried to talk to a CD player to change the music. When it didn't answer, she was confused.
That is no different from your future customer.
They'll expect to tell your bank, "Move $500 to savings and pay my electric bill," and they'll expect the app or agent to just do it.
4. The real cost is talent
Is AI expensive? That depends on how you look at the P&L.
Bhattacharya dropped a necessary reality check on the financials. The cost of the technology itself is plummeting. The cost of tokens and hosting models on Azure or AWS is a fraction of the value they deliver.
The unit economics are undeniable:
- Human interaction: $24 fully loaded cost
- AI agent resolution: 20-40 cents per interaction
- ROI potential: 60-120x cost reduction per transaction
But the real cost is talent.
To build these systems, you need engineers who have built systems at scale. You need "architects and builders" who understand how to integrate new AI models into old banking architectures.
Those people are rare. They're expensive. And you're competing with every other bank (and Big Tech) to hire them.
The bottleneck is your ability to attract the leadership and engineering talent required to execute.
5. Safety comes from experience
There's a lot of fear in the room.
We worry about "deceptive alignment" - where a model pretends to follow your rules. We worry about hallucinations. We worry about the societal impact of automation on jobs.
These are valid concerns. But avoiding the technology just makes you obsolete.
Kanishka compared our current moment to the early days of aviation. Flying became safe because engineers built rigorous checks, redundancies, and protocols. We're on that same curve with AI.
You get safe by building the right guardrails.
For now, that means keeping a "Human in the loop."
AI safety guardrails for banks:
- Design time: AI writes code and summaries, humans review before deployment
- Run time: AI drafts customer responses, humans approve before sending
- Monitoring: Continuous human oversight until confidence builds
- Escalation: Clear protocols for when AI reaches its limits
As confidence grows, the human steps back. But trust is the currency of banking. You can't afford to lose it.
The banks that win will be the ones that move fast with strong controls.
Your next step
The panel was asked a simple question: "What is the one thing banks should do right now?"
Their answers were direct.
Marco: Start with yourself. Treat AI like an "exoskeleton for your brain." Learn to prompt. Ask it to find flaws in your strategic logic. If you understand the reasoning power of these models personally, you can lead an organization to use them.
Kanishka: Find a "no-brainer" use case in your bank - something small but meaningful - and execute it end-to-end with extreme urgency.
Dave: Reinvent. Look for the processes you can rebuild entirely with agents, and then do so.
The AI banking transformation is happening now:
- Legacy excuse eliminated: AI can decode your oldest systems
- Data goldmine ready: Your unstructured data becomes instantly valuable
- Agent revolution: Move beyond chatbots to task execution
- Talent bottleneck: Real but solvable with the right strategy
- Safety through action: Controls come from experience, not avoidance
The banks that scale AI will be the ones with the courage to start, the discipline to execute, and the willingness to reinvent from the ground up.
So what's stopping you?
FAQ: AI Implementation in Banking
Q: How long does AI-powered legacy code analysis take?
A: Large Language Models can analyze and document legacy code in days or weeks, not the months or years traditional approaches require.
Q: What's the biggest risk when implementing AI agents?
A: Moving too fast without proper human oversight and control mechanisms in the early stages.
Q: Should smaller banks attempt AI transformation?
A: Start with narrow use cases like compliance checking or document analysis where the ROI is immediate and measurable.
