Don't force AI into your business. Build it around how your customers already work.
Stop Building AI Features Nobody Asked For
By Daniel Lambo, Blaze Group
In July 2025, the Commonwealth Bank of Australia told 45 customer-service workers that an AI voice bot would replace them. Within weeks, the Finance Sector Union reported surging call volumes, overtime offers, and team leaders answering phones.
On August 21, the bank reversed the layoffs and apologized, admitting its review failed to weigh "all relevant business considerations.”
The bot didn’t fail. The arithmetic did. The bank assumed that automating tasks would eliminate the need for human labor. It learned that the two are not interchangeable.
While many evaluate AI through surface metrics like total conversations, automated summaries, or deflected tickets, its real impact lies in the effort saved for the user. True success is measured by how seamlessly a tool simplifies a customer's workflow and eliminates unnecessary steps.
Today, the real opportunity isn't just adopting AI. It's thoughtfully embedding it into the experiences customers already love to create genuine, lasting value.
At the start of 2023, roughly 10% of finance and insurance businesses using Ramp, an all-in-one finance automation platform, had adopted an AI product. By September 2026, that figure had climbed to 73.7%, marking a major shift in industry adoption.
Where AI sits matters more than how smart it is
In 1997, at Apple’s Worldwide Developers Conference, an audience member challenged Steve Jobs’s technology strategy. Jobs, then serving as Apple's interim chief executive, famously replied: “You’ve got to start with the customer experience and work backwards to the technology.” He acknowledged making that same mistake himself multiple times.
Nearly thirty years later, much of fintech AI is still built the other way around: model first, customer second. The first mistake is often placement. AI can sit in three places:
In front: The customer interacts with it directly, often through a chat window they must find, learn, and trust.
Beside: It helps an employee perform their job more efficiently.
Underneath: It matches, flags, checks, pre-fills, or routes work within an existing process.
Companies start in front because it demos well. Yet, it is also the hardest place to win.
Gartner found 49% of customers would have been willing to use a chatbot in a recent service interaction, but only 7% actually did. After a bad chatbot experience, just 27% would try one again. Done badly, the front-end chatbot becomes this decade's "press 1 for billing": a wall built to keep customers away, sold as a way to serve them.
AI underneath works like the best waitress you’ve ever had. Your sweet tea gets refilled before you notice it’s low, and you never have to ask. If she grabs the wrong pitcher, she catches the mistake at the table, before you take a sip.
Good AI underneath works the same way. It runs on data you already have, within steps your customer already takes, and knows when it’s uncertain. Say it matches a receipt to a card charge. When it’s confident, it files the match. When it isn’t, the match goes where a bookkeeper or operations analyst can confirm it with one click. The customer only sees finished books. Mistakes can still happen, but someone on your team catches them before they reach the customer.
In front can still work, but it has more to prove. It has to beat the traditional route on speed. It has to know enough that the customer never has to repeat themselves. It has to finish the task, not just explain how. And it has to hand off to a person the moment it becomes uncertain.
So before asking which model to use, ask where the output lands, what happens when it's wrong, and who has to do extra work because of the mistake. That sounds simple on a whiteboard. Two companies illustrate what it looks like in practice: one placed its AI underneath, the other in front.
Two case studies: One AI screened transactions, the other screened the customer
Stripe demonstrates what effective integration looks like. In 2025, it extended Radar, its fraud detection system, to ACH and SEPA bank payments. Radar evaluates more than 1,000 transaction signals in under 100 milliseconds. Stripe reports that users experience 42% less SEPA fraud and 20% less ACH fraud on average. No customer has to understand how Radar works. They simply make the payment they were already going to make.
In contrast, Virgin Money placed its chatbot directly in front. In January 2025, fintech commentator David Birch, a Virgin Money customer, opened the bank’s chat with a routine banking question: “I have two ISAs with Virgin Money, how do I merge them?” ISAs are UK tax-free savings accounts, and consolidating them is a common banking request.
The bot replied, “Please don’t use words like that,” and warned that it would end the chat if he continued.
The offending word, it appeared, was “Virgin.” The chatbot had been programmed to block crude language, and it couldn’t distinguish vulgarity from the name on its own front door.
Birch posted a screenshot on LinkedIn and asked which words were now banned. Virgin Money apologized in the comments, confirmed it could merge the accounts, and pulled the bot. It turned out to be an older system using basic language processing, not the bank’s newer assistant.
Consider what the exchange cost. A customer with a one-line question ended up screenshotting the conversation, posting it publicly, and getting his answer from the bank’s social media team. The bot didn’t eliminate a step. It added several. And the failure wasn’t intelligence. A blunt filter stood between a customer and a simple answer.
Virgin Money later introduced a more capable front end. Its newer assistant, Redi, was developed with frontline contact-center staff and escalates complex cases to human agents. Microsoft reported a 97% completion rate for outbound journeys initiated by Redi.
Company | What They Built | Habit Used | Result | Decision That Mattered |
 | Radar fraud detection extended to ACH & SEPA payments | Existing payment flow; no new customer behavior required | 42% less SEPA fraud; 20% less ACH fraud | Embedded AI underneath existing processes |
 | Customer-facing chatbot | Required customers to interact through a chat interface | Bot flagged a routine query as inappropriate; withdrawn after an apology | Placed AI in front without sufficient contextual understanding or effective escalation |
Sources: Stripe Newsroom (April 30, 2025); David Birch LinkedIn post (January 2025); Fortune.
The lesson isn't "never put AI in front." The closer AI gets to the customer, the greater the cost of a mistake, and the more important a seamless path to resolution has to be.
At Blaze, that's the standard: if your customer has to work harder because of your AI, what exactly have you automated?
The case against this: customers will learn new habits
The strongest counterargument is that people learn new interfaces all the time. ChatGPT went from launch to more than 900 million weekly users in about three years. In fintech, Cleo built a chat-first money assistant for Gen Z and ended 2025 with 1.3 million paying subscribers and $254 million in revenue, according to Sacra.
If people are willing to talk to an app about their money, perhaps designing around existing customer behavior is the more timid approach.
But look closer. ChatGPT asked people to do what Google had trained them to do for twenty years: type a question into a box. And Sacra estimates 62% of Cleo's 2025 revenue came from cash-advance fees, addressing one of the oldest needs in consumer finance: access to money before payday. The chat earned the download. A need as old as payday paid the bills. Old needs come with old rules, too: Cleo paid $17 million in 2025 to settle FTC allegations concerning how it marketed those advances.
Harvard Business School professor John Gourville calls this disconnect the "9x effect." Customers overvalue what they already use by roughly three times, and builders overvalue what they've created by about the same margin. The result is a ninefold gap in perceived value. New habits that stick often bridge that gap by building on familiar ones.
You can ask customers to adopt a new habit. Just pay for it with an old one.
Bridging that gap means shifting focus from changing user behavior to improving the workflows customers already know.
What to do on Monday: start with the job, not the model
Paul Graham tells founders to "do things that don't scale": do the work by hand before you automate it. For AI in fintech, that's the starting point.
Watch five customers do the job. Not a survey. Sit beside them or share a screen while they dispute a charge, onboard a business, pay a bill, or close their books. Capture every tool they touch, including the spreadsheet and the text to their bookkeeper.
Count the steps and circle the one they hate. Document minutes, screens, fields, repeat contacts, and handoffs before anyone writes a prompt. That one step, not the entire journey, is your starting point for AI.
Place the AI in this order: underneath, beside, then in front. Put AI in front only when conversation is already how customers complete the task, and only if the first two approaches can't eliminate the friction.
Decide who catches mistakes before you ship. In 2024, a British Columbia tribunal held Air Canada responsible for incorrect fare advice its chatbot gave a grieving customer. Your bot is not a liability shield. Keep a person one tap away.
Report the step you removed. "Four fewer fields" and "30% fewer repeat contacts" measure outcomes. "12,000 AI conversations" measures activity.
We turned these five steps into the Workflow Fit Scorecard below.
What customers say, and what they mean
Fintech friction rarely shows up labeled as an AI problem. It shows up in something a customer says to a representative, a form they abandon, or a call they have to make twice. Here's how we interpret those sentences.
What customers say | What's actually wrong | Don't build | Build instead | Where it sits | Measure |
"I already sent you that." | Their information doesn't follow them between channels | A bot that asks for it again | AI that puts their history on the representative's screen before the call begins | Beside | Repeat contacts within 7 days |
"I'll finish this later." | Paperwork friction | An AI onboarding guide | AI that reads uploaded documents and pre-fills the form | Underneath | Completion rate; minutes to approval |
"I'll sort my books out at tax time." | Time constraints | An "AI CFO" chatbot | Automatically categorize transactions from bank feeds; flag only exceptions | Underneath | % auto-categorized and unchanged at month-end |
"I get paid Friday. The bill's due Wednesday." | Cash flow timing, not discipline | An AI budgeting coach | Allow customers to align due dates with payday; warn only when funds are insufficient | Underneath | On-time payment rate |
"Can I just talk to someone?" | Lack of trust | An app-only assistant | AI that supports tellers and phone representatives | Beside | Time to resolution |
"Why was I turned down?" | Unclear decision criteria | AI that makes decisions alone | AI that drafts a case summary and plain-language explanation, with human review and approval | Beside | Days to decision; reversal rate |
Notice how few of these solutions require a conversation.
The final row has data behind it. In TD's 2026 survey of 2,504 U.S. adults, 67% were comfortable with AI handling fraud detection, but only 18% would trust it to make financial recommendations without human involvement.
The "talk to someone" row is especially relevant to the communities we serve. The FDIC finds that teller use is highest among older, lower-income, and rural households. In Mississippi, 9.4% of households were unbanked in 2023. These customers aren't behind. They're showing us which channels they trust. Help the people in those channels serve them better.
The Workflow Fit Scorecard from Blaze Group®
Hit a dealbreaker, or stuck below the bar? That's the conversation we have most often with founders. Here's how we usually start.
Commonly asked questions
Should we build our own model?
Probably not at first. Prove the workflow and the outcome before making model ownership part of your strategy. Your advantage lies in the context only you have: your customers' data, your integrations, your controls, and knowing which steps frustrate your customers most.
Is a chatbot ever the right first move?
Yes, if three things are true: customers already ask those questions in writing, the bot can complete a clearly defined task, and a person can pick up the conversation without making the customer start over. Keep it away from making consequential decisions about people's money without human oversight.
Our investors want to hear "AI-first." How do we pitch this?
Pitch the metric, not the label. "We removed four steps from onboarding and cut approval time in half" holds up under diligence. In 2024, the SEC fined two investment advisers a combined $400,000 for misleading claims about their use of AI.
The best AI roadmap is already on your payroll
In August 2025, team leaders at Australia's biggest bank were back on the phones, handling calls a voice bot was supposed to absorb. CBA's mistake wasn't using AI. It was deciding which work had disappeared before understanding all the work around it.
Every fintech has people who know where that work really happens: the representative who hears why customers call twice, the operations analyst with the ugly spreadsheet, the teller who knows which form makes people give up. They will never appear in an AI pitch deck. But they're the best roadmap you'll get.
Talk to them first. Then build something that makes their day, and their customers’ days, one step shorter.
Casey Ariel Diké spent 13 years in global finance, including roles at Bank of America Merrill Lynch and as Chief Credit Officer at Hello Tractor, before founding Blaze Group. Alongside Daniel Lambo, a Blaze Apprentice specializing in computer science and emerging technologies, she offers limited AI Workflow Fit office hours each month to help fintech founders determine where AI belongs in their products. Book a session.
SOURCES & REFERENCES
Commonwealth Bank of Australia (CBA): Finance Sector Union reports on AI voice bot rollout and reversal (July–August 2025).
Ramp AI Index: Data on AI adoption in finance and insurance businesses via FRED (September 2026).
Apple WWDC (1997): Steve Jobs on starting with customer experience and working backwards to technology.
Gartner: Research on customer willingness and usage rates for service chatbots.
Stripe Newsroom: Stripe Radar expansion to ACH and SEPA payments (April 30, 2025).
Virgin Money & David Birch: Coverage of chatbot query misclassification and deployment of Redi assistant (January 2025; Fortune, Microsoft reports).
Cleo & Sacra: Cleo financial profile, subscriber count, revenue breakdowns, and FTC settlement details (Sacra, 2025).
John Gourville (Harvard Business School): "Eager Sellers and Stony Buyers: Understanding the Psychology of New-Product Adoption" (The 9x Effect).
Civil Resolution Tribunal (British Columbia): Air Canada chatbot liability ruling (2024).
TD Bank Group: Consumer survey on trust in AI for fraud detection vs. financial recommendations (2026).
FDIC: Survey of Unbanked and Underbanked Households (2023).
U.S. Securities and Exchange Commission (SEC): Enforcement actions against investment advisers for AI washing claims (2024).
Comments