5 Key Takeaways: Why Millions in Tech Spend Still Leaves Manual Work Behind

TL;DR

In our latest fireside chat, Uptiq's Kyle Kneubuhl sat down with John Mignone, Institutional Client Group at U.S. Bank and a lecturer at the University of Texas at Dallas, talk about why so much manual work survives underneath systems that cost millions. Here are five takeaways from the conversation.

1. Manual work is the hidden tax nobody puts on a balance sheet

Across audits at banks, credit unions, and enterprise tools generally, most institutions are only using somewhere between 15% and 70% of what they're actually paying for in any given system. "You're paying for the full enterprise price, majority of the time you're getting things that are in there that you probably don't use," John noted, pointing to Microsoft Office as the example everyone recognizes but nobody audits. That underused capability is exactly what gets filled in with manual workarounds, staff copy-pasting between systems, dropping into an Excel worksheet to do what the system won't, and bringing it back in. As Kyle put it, "Manual work is often like the hidden tax of disconnected technology." That workaround time never shows up as a line item, but it's real cost, and it compounds every time a new system gets bolted on without retiring the old one.

2. AI doesn't mean ripping out your systems, it means filling the gaps between them

The instinct when technology isn't working is to assume the fix is a bigger purchase, a new LOS, a full core replacement. Both panelists pushed back on that. "A lot of people think, well, if I want to go into AI, I've got to rip out systems, or it's like buying a whole new LOS, and it's actually not," Kyle said. "It's something that can be pieced together to fill manual gaps." John summed up the same idea from the institution's side: "You're not replacing the processes and procedures around what you're doing from an AI agent... You're leveraging what you currently are doing." His advice for legacy cores specifically was blunt: let core be a ledger, nothing more, and don't build your innovation roadmap on its timeline.

3. Most of a credit decision is math, protect the sliver that isn't

John broke down what actually makes up a lending decision: "If you look at credit, and I take the simplistic approach, 90–95% is going to be the same. It's mathematics. It's spreading, it's client documentation intake... 5% or less is effectively a finger in the air." That 5% is the judgment call, the "secret sauce" that varies institution to institution. The goal isn't to automate that part away, it's to protect it, by letting agents absorb the repeatable 95% so your team's expertise gets spent entirely on the decisions that actually need a human weighing them.

4. A trustworthy AI partner has to be provable, grounded, and guided, and stick around after the sale

Kyle's bar for any AI-driven process: it has to be grounded in the institution's actual policies and regulatory obligations, provable in terms of what it did and why, and guided, with hard stops that route back to a human when it doesn't know what to do. "At the end of the day, the institution has to be able to do three things," he said. "It has to show that these agents or processes are grounded... They have to be provable... And then they have to be guided." John added the relationship test: "You're buying the seller... the people that you're interacting with are actually able to [help] when there's a problem to get it solved." If a vendor disappears after the contract is signed, "you have the opportunity to move. You should, and you should be looking elsewhere."

5. Start small, let it snowball, and don't be afraid to break it first

Neither panelist recommended a big-bang rollout. "Entering into AI currently is actually a lot easier than the legacy purchasing process," John said. "We really believe that people entering AI should do it at small to medium size. It's not a gigantic lift." The safety net is what makes it low-risk: pick one manual process, e.g., spreading or document intake, and if the agent underperforms, you simply revert to the manual process you were already running. "If for some reason it doesn't work... you just go back to the manual. There's no production letdown," Kyle said. Fix seven, ten, fifteen of those individually, and the orchestration between them eventually becomes the workflow itself.

If you want the full conversation, watch the full fireside chat recording here.

Or, if you're curious how this could look inside your own workflows, book a discovery call with Uptiq.

‍