Sponsored deep dive essays are opportunities for me to work with specific companies on topics and issues important to them, me and bank readers. The sponsor provides input and information, but each piece is researched and written by me, and are my honest thoughts.
Thanks to Taktile for partnering with me on this piece and helping me to better understand how to make the most of the tech moment we’re in.
I have some good news and some bad news when it comes to AI at banks.
Good news: Banker enthusiasm for artificial intelligence, specifically agentic AI, has led to more experimentation and faster adoption. Boards are asking executives where they can prudently leverage AI to cut inefficiencies and operating costs. Some banks are implementing AI use mandates or token quotas.
Bad news: Most of that experimentation isn’t showing up on the income statement. While more than 80% of organizations have explored or piloted AI tools and nearly 40% report deployment, “these tools primarily enhance individual productivity, not P&L performance,” wrote researchers in a 2025 MIT study.
That gap makes sense once you look at where bankers are applying AI. Summarizing a document or drafting an email saves time, but it doesn’t touch the highly manual processes where AI could likely deliver the biggest return, like commercial underwriting and Know Your Business (KYB).
The core business of banking is making good loans to good customers. Who does the bank let in the door, and how much does it lend them once they’re inside? It turns out those decisions are also where AI offers the largest impact for banks — if they’ve got the right infrastructure to leverage it.
Why Are Many Decisions Still So Manual?
Every new customer relationship and every loan starts with two questions. Is it safe to do business with this customer? And how much can we lend them? To answer either of them, bank employees need to gather all relevant information and weigh it against what the bank knows and what it’s allowed to do before reaching a conclusion.
This work involves many more touchpoints to service commercial customers. A bank needs to verify the business’s ownership, weed out bad actors and screen for adverse media. It collects and analyzes years of financial statements, applies its own credit policy, ensures compliance with regulations and finally puts the decision in front of an expert for the final judgment.
Even though banks use deterministic rules to automate certain steps in the decision process, much of the work remains highly manual and fragmented. Large teams of analysts review financial statements and research company ownership. Rules engines are still great technology for what they were built to do: black-and-white, if-then logic. Does this applicant meet the minimum credit score? Is this business based in a country the bank can’t serve?
Complex decisions, though, require processing messy, unstructured data that rules engines can’t handle. Financial spreading in commercial lending, or KYB verification for a company with a layered ownership structure. The rise of AI only emphasizes the cracks in banks’ decision-making infrastructure. These systems aren’t designed to seamlessly connect data, AI agents, rules and human experts in an end-to-end process. How would a traditional rules engine coordinate human review of an AI agent’s work? Where does the central audit trail of all AI-driven decisions live?
These systems were built for a world without AI, and they’re struggling to keep up with a world where money and information move faster.
So Can AI Fix These Inefficiencies?
Yes — with the right approach.
Agentic AI adds a powerful new ingredient to this process. It can handle unstructured data, like customer documents and website information. It can pull those threads together into a fuller picture and assemble the essential information an employee needs to rule on an ambiguous case faster.
And the models are getting better all the time. Taktile Labs, the research arm of this deep dive’s sponsor, found that agents could spread financial statements with 96.5% accuracy, compared with an 89% human baseline. In KYB, agents produced higher-quality adverse media evidence than human analysts, and an agent-first hybrid approach reduced analyst workload by 93%.
Looking at what AI is now capable of, it might be tempting for a bank to spin up dozens of bespoke, internally built agents and point them at the problem. But adding a potent new ingredient to a process that already struggles to unify data, rules and human expertise doesn’t fix the process. It makes the cracks more obvious. Hundreds of disconnected agents won’t transform a bank’s workflows or lower its costs, and they’re likely to make an already fragile setup more brittle.
The technology for faster, more intelligent decisions exists. What limits banks from getting real value from this technology is the fragmented, antiquated infrastructure around these complex decisions. To capture AI’s biggest upside, banks will need a tech layer that orchestrates the infrastructure, rules, models, people and data that go into making a decision.
That might sound like a tall order, but the technology already exists. It’s called a decision layer.
Decision Layer 101
What, you might ask, is a decision layer? Taktile, the sponsor of this deep dive, has created a decision layer that it has implemented at banks around the world. The way Taktile sees it, a decision layer is a technology-agnostic, horizontal platform where teams across the bank work in a single stack.
Importantly, it should be modifiable: banks should have the ability to gradually increase AI-driven automation as their confidence builds. And it should bring together four components that keep the system governed, auditable and explainable:
- Context. Data is often disconnected across a bank, with no central source of truth. The decision layer should give a bank a unified, real-time view of the customer that includes structured and unstructured data. Bankers should feel confident they have a complete picture based on the available information.
- Decision logic. Not every decision needs a model or an agent behind it. Why task an agent with determining whether the bank can do business with a customer based in another country, when simpler automated tech can answer that? Clear-cut points in the decision should run on deterministic rules, which are fast, auditable and cheap to operate.
- Agents and AI. For ambiguous cases, such as unstructured documents, incomplete data or unusual patterns, an AI agent gathers information and prepares the case. That could mean spreading financials, mapping an ownership structure or searching for adverse media.
- Humans. When a case warrants enhanced due diligence, a person makes the final call, backed by everything the system has already assembled. The layer should make that handoff easy: when an agent flags a case for expert review, the case and its context are delivered to a reviewer.
Orchestration ties these components together. It lets a bank establish how a case moves between rules, agents and people, and adjust as a use case evolves, without ripping out the underlying infrastructure. This could look like: decision logic automating simple cases, agents preparing the tricky ones and humans applying judgement when it’s needed. A decision layer should be able to combine all three and plug into the systems that banks already run.
Done well, a decision layer changes who gets to shape a bank’s decisions. It enables collaboration: business and technical teams share a workspace and can see the complete customer journey. It gives business owners control over the agents and key performance indicators they’re accountable for, so that the head of lending with slipping approval rates can see what’s happening and make changes directly. And it supports continuous improvement, since the teams monitoring outcomes can keep iterating as markets, regulations and models change.
And because the decision layer is designed and maintained by tech experts, bankers don’t have to build, monitor, maintain and evolve something in-house.
The “Build versus Buy” Debate Is the Wrong Debate
Now, artificial intelligence could be a truly revolutionary tool for banks. But I’ve seen banks go through a couple of tech cycles, and I worry they might repeat the same pattern of adoption with AI as they have with other technology that came before it.
I know bankers are experimenting with agentic AI and are spinning up internal AI tools or workflows. But can that vibe-coded tool scale? Is it auditable? How much time will bankers spend training their agents, or figuring out how to add appropriate human oversight? Teams within a bank can now build agents for their own tasks — but what happens when they need those agents to connect to an end-to-end decision process?
There’s also the question of who keeps the lights on. Many banks think about building the internal tools — not who will maintain them and keep them governed, secure and aligned with changing regulations. What happens when the people who built an agent leave, and a critical process gets lost in the handover?
The DIY approach has some real downsides that bankers need to think about long-term. It doesn’t scale. It’s expensive. It creates tech debt. It falls behind the latest models. All of that adds up to an underwhelming, fragmented, brittle technology experience. And it’s not really the core work of a bank, which is to make good loans to good customers.
It’s the “build versus buy” debate all over again. And a DIY approach risks rebuilding silos within a bank instead of tearing them down.
This doesn’t mean bankers are being careless! The pressure to adopt AI or be left behind is pushing teams to make complex technology decisions quickly. Building an AI agent can be fast and cheap. But building the infrastructure that connects, guides and governs agents could become a multiyear undertaking.
That’s why banks have decided to buy rather than build in so many areas of the business. The data supports that choice: strategic partnerships are twice as likely to succeed as internal builds, according to the same MIT study. Banks can keep experimenting and still turn to a tech partner for a cohesive, transparent platform that combines AI and other modern tech to enhance the decisions that define their P&L. That’s the job of a decision layer.
Today’s Strategic Choice for Tomorrow’s Tech Reality
AI could be truly transformative for banks, and the enthusiasm for agentic AI inside banks is heartening and refreshing. But uncoordinated experimentation has downsides. It reopens the build versus buy debate, saddles teams with bespoke agents to maintain and rebuilds the silos between them.
Whether to use AI, and where, is the easier question for banks to answer. The harder question is whether a bank’s infrastructure can use automation and AI efficiently. Does it unify disparate systems, workflows and data sets so teams share context? Does it provide visibility and auditability throughout? Does it keep the people accountable for outcomes in control of them, even as technology’s change of pace accelerates?
Banks that make this infrastructure choice today will likely be better positioned for whatever comes next, while managing costs and improving their workflows. The result is better decisions, faster.
