Utah’s AI-Native banking conference: from AI demos to accountable workflows

Tap a dotted-underlined term for a definition. Use Aa in the navigation for reading preferences.

Analysis

The strongest signal from the conference’s published program and related announcements is a move toward AI embedded in operating workflows. For consumer lenders, the opportunity includes faster document handling, broader review of customer interactions and more consistent servicing. The corresponding test is whether institutions can measure errors, constrain actions and reconstruct decisions. The implications below are The Credit Current’s analysis; they are not attributed conclusions from unreviewed panel recordings.

The event—and what this report can establish

The AI-Native Banking & Fintech Conference took place September 29, 2026, at the Ken Garff University Club at Rice-Eccles Stadium, University of Utah. Spring Labs hosted it with the Utah Bankers Association as co-host; the event site also identifies the American Fintech Council, the university’s Fintech Center and Utah’s Governor’s Office of Economic Opportunity as partners. Training and networking preceded the main program on September 28. [1]

An event-day public LinkedIn post by Peter Renton reported that the conference was underway and identified Phil Goldfeder’s opening interview with Michael Hsu. FinWise Bank’s own event listing separately identifies the date and venue. These corroborate the event; they do not establish attendance totals or the contents of every session. [2] [3]

This is a reported analysis of the organizer’s program, public event evidence and company materials available the following morning. It is not an on-site account or a transcript-based recap. Session descriptions below identify the published program; no unverified remarks are assigned to speakers.

Who was on the program

The opening conversation paired former acting Comptroller Michael Hsu with the American Fintech Council’s Phil Goldfeder under the title “AI and the Appearance of Oversight.” The operating-use-case panel listed Figure, Zions Bank, Visa and Avant. [4]

Other sessions paired Chime’s Jeff Currier with Fintech Takes founder Alex Johnson; examined bank implementation with Square Financial Services’ Richard Rosenthal and the University of Utah’s Ryan Christiansen; and addressed small-business lending with LoanPro, Newity, Fundbox and ByzFunder. [4]

The program also included a security panel with CALM xAI, FairPlay, Guardrail Technologies and Uptiq; customer-experience coverage involving Self Financial, Avant, Healthcare Finance Direct and Imprint; and workshops on complaints, financial crime and cybersecurity. That breadth connects origination, servicing and risk management rather than treating AI as a single technology purchase. [4]

The central distinction: an answer versus an action

An assistant that summarizes a call creates an output for someone to inspect. An agent that changes a payment arrangement, initiates a refund or sends a dispute response changes the customer’s position. In our analysis, that is the useful dividing line for assessing the conference’s agentic-AI theme: what authority crosses from a person or controlled workflow into software?

A convincing deployment should make that authority explicit. Readers evaluating a system should be able to identify what data it may retrieve, which actions it may propose, which it may execute, and which require approval. A permissions boundary enforced by the surrounding application is more testable than a prompt that merely tells the model to be careful.

Human review needs equally concrete treatment. A reviewer needs the underlying evidence, adequate time, and the ability to stop or reverse an action. If a queue is too large to examine or a screen hides the relevant source record, a human approval button can create confidence without meaningful challenge. This is our operational interpretation of the oversight issue, not a quotation from Hsu.

The clearest product development: LoanPro and Spring Labs

A September 28 company release announced a LoanPro–Spring Labs partnership to embed complaint, dispute and quality-management capabilities within LoanPro. The release describes analysis of supported calls, emails and chats while LoanPro remains the system of record. [5]

The rollout is phased: the companies say quality management and complaint workflows will begin reaching LoanPro customers in Q4 2026; credit, transaction and fraud disputes are planned for 2027. Those dates matter: the announcement does not establish that every capability was already available at the September 29 conference. [5]

For a lender, the integration’s potential value is reducing the movement of information between servicing records, customer contacts and compliance review. Our analysis: the same integration makes permissions, record reconciliation and escalation design more consequential. A mistaken classification that stays in a draft is different from one that drives an account action.

The release does not provide independent outcome evidence for the announced integration. Procurement teams should ask which channels and products are covered, how omissions are detected, what the human workload becomes, and whether audit records preserve the original contact alongside the AI’s interpretation.

Lending: faster preparation is only part of the credit decision

Uptiq’s event page advertised demonstrations spanning document classification, extraction, validation and commercial-lending work, with source provenance, exception paths and approval controls. These are the vendor’s descriptions of its offering, not independently observed conference test results. [6]

The economically useful starting point is often a bounded task: extracting a field, finding a missing document, assembling a credit file or comparing an application with an approved policy. These tasks can shorten turnaround while leaving a clear checkpoint before an approval, limit or price reaches a customer.

That checkpoint matters because different errors have different costs. A document agent can read an income figure correctly yet use the wrong period, overlook an obligation or attach it to the wrong applicant. Field accuracy, decision accuracy and eventual loan performance therefore need separate measurement. Faster processing does not by itself demonstrate better credit selection.

For a credit team, a useful pilot compares the AI-assisted process with the existing process on the same kinds of files. Track exceptions, rework, missed material facts and the disposition of overrides. If the tool contributes to credit decisions, also examine whether its rationale reflects the actual approved decision process. These are proposed evaluation questions, not results reported by conference participants.

Servicing, disputes and merchant oversight deserve equal attention

The conference’s servicing and complaint themes are particularly relevant to installment lending and merchant-originated credit. Origination is only the beginning of the customer relationship; problems can emerge later through payment allocation, refunds, product delivery, a merchant representation or a customer’s difficulty obtaining help.

A hypothetical example illustrates the opportunity. Several customers at one merchant say a financed service was never delivered, but their contacts are coded under different servicing categories. An AI system could help surface that pattern and assemble the underlying contacts. A reviewer would still need to distinguish duplicate reports, merchant explanations, transaction evidence and confirmed outcomes before drawing a conclusion.

Useful measures would include the share of eligible contacts actually examined, missed complaint types, repeated contacts after closure, reopening rates and the time from intake to escalation. A rising complaint count may initially reflect better detection. A falling count may reflect genuine improvement, missed intake or premature closure. Interpretation requires both a denominator and a review of case quality.

The practical takeaway is to treat AI-generated themes as investigative leads with traceable evidence. A fluent summary alone should not determine a merchant restriction, a customer remedy or the conclusion of a dispute.

Customer experience: useful personalization requires clear authority

The program’s combination of primary-account relationships and customer-experience sessions points to a broader commercial ambition: software that helps people complete financial tasks, rather than simply answers product questions. The relevant business question is whether a service reduces friction and improves completed outcomes without creating avoidable mistakes.

Consider a customer asking for help with a payment. Explaining available options, recommending an option and executing a change are separate steps. A well-defined journey makes the selected action, amount, timing and consequences clear before it is carried out. The institution also needs a reliable record of what the customer authorized.

Assess the experience through resolution quality, repeat contacts, complaints, inappropriate recommendations and unsuccessful transfers to a person. Containment rate—the share of contacts that never reach an employee—can be useful, but it can also reward a system that makes human help harder to obtain. Our view is that revenue and efficiency claims need to be evaluated alongside customer outcomes.

Governance context: do not misread the 2026 model-risk guidance

The April 17 interagency model-risk guidance, issued by the Federal Reserve as SR 26-2, replaced SR 11-7. Its scope note explicitly excludes generative and agentic AI models while stating that an institution’s risk-management and governance practices should guide controls for systems outside the document. It is , not a set of enforceable standards, and is expected to be most relevant to banks above $30 billion in assets, with qualifications for some smaller institutions. This is background context, not a conference announcement. [7]

That scope distinction prevents two mistakes: presenting SR 26-2 as a complete agentic-AI rulebook, or treating exclusion from that particular document as permission to operate without governance. A workflow may also combine a traditional quantitative model with a generative interface and action-taking software. Each component’s purpose and authority should be understood.

For management, a practical inventory would connect each use case to its business owner, data sources, vendor dependencies, allowed actions, review process and escalation route. Our analytical recommendation is to test the complete workflow—including integrations and human handoffs—rather than evaluate model responses in isolation.

A practical scorecard for evaluating the ideas

The following questions translate the conference’s themes into an internal evaluation framework. They are The Credit Current’s proposed scorecard, not reported conference findings or a statement of legal requirements.

Area	Evidence to request	Failure to look for
Authority	A list of permitted actions, approval gates and tested stop controls	The agent can execute more than the business intended
Source quality	Original documents or contacts linked to each material conclusion	A plausible conclusion cannot be traced to evidence
Credit quality	Comparable file reviews, override analysis and eventual outcome monitoring	Faster decisions conceal missed obligations or inconsistent treatment
Servicing quality	Missed-case testing, reopens, repeat contacts and escalation results	High closure rates hide unresolved customer problems
Security	Tests of malicious input, access boundaries and cross-customer isolation	Untrusted text changes instructions or reveals another customer’s data
Economics	Total operating cost including review, exceptions, integration and remediation	Headline savings exclude the work moved elsewhere
Change management	Recorded versions, regression checks and a tested fallback process	A vendor or policy change silently changes customer outcomes

What would turn the conference’s promise into evidence

Watch for named production deployments, clearly defined coverage, comparable before-and-after measurements and disclosures about exceptions. For the LoanPro partnership, the first milestone is the announced Q4 introduction of complaint and quality workflows; later dispute functionality needs its own verification. A release schedule is useful evidence of intent, not evidence of delivered performance. [5]

Ask vendors to report both the gains and the remaining work. An accuracy statistic is hard to interpret without the task, sample, denominator and treatment of ambiguous cases. Savings claims are more useful when they include implementation costs, human review and remediation rather than only the time spent on one automated step.

Utah’s conference brings together a consequential set of operating questions. The opportunity for banks and lenders is to make everyday work faster and more consistent while keeping responsibility understandable. The test after the event is whether an institution can show what its agents did, why they were permitted to do it, what went wrong and how the institution responded.

What remains uncertain

No complete September 29 session recordings or transcripts were reviewed, so this article does not attribute unverified recommendations or quotations to speakers. The published agenda was still marked subject to change; agenda participation is distinguished from independently confirmed remarks. No verified 2026 attendance count is reported. Product functionality, rollout dates and performance descriptions remain company claims until independently demonstrated. The public LinkedIn evidence was an indexed event-day excerpt, not access to a private feed.