2025: The Year AI Stopped Being a Pilot Project

Published: December 22, 2025

A Message from Anna

In January 2025, every financial institution had an "AI strategy."

By December 2025, the gap between strategy and execution had widened into a chasm. Some institutions deployed AI at scale. Others were still running pilots. The distance between them may take years to close.

This is the story of the year AI stopped being a pilot project—and what that means for everyone still experimenting.

The Scale Shift: From Experiments to Infrastructure

JPMorgan: 200,000 Employees on LLM Suite

The clearest signal that AI had moved from pilot to production came from the largest U.S. bank.

JPMorgan Chase deployed its LLM Suite to 200,000 employees and ran over 450 proof-of-concept projects through 2025.

What LLM Suite does:

Research and analysis assistance
Document summarization
Code generation
Communication drafting
Knowledge retrieval across enterprise data

Why it matters:

When the world's largest bank gives 200,000 employees access to AI tools, it's not an experiment. It's a statement: AI is infrastructure now.

The 450 proof-of-concept projects signal something equally important—JPMorgan isn't just deploying what works today. They're systematically discovering what will work tomorrow.

Cash App: Moneybot Goes Live

Square's Cash App introduced Moneybot, described as an "advanced financial chatbot."

What Moneybot does:

Analyzes user spending patterns
Recommends financial decisions
Creates personalized savings plans
Executes trades in stocks and cryptocurrencies
All within a conversational interface

The shift:

Previous financial chatbots waited for questions. Moneybot actively monitors, identifies opportunities, and takes action within user-defined parameters.

This is agentic AI in consumer finance—not just answering questions, but making things happen.

Accountants: 46% Daily AI Usage

The Intuit QuickBooks 2025 Accountant Technology Survey revealed that AI adoption among accountants outpaces nearly every other profession:

Metric	Accountants	Small Businesses
Daily AI usage	46%	28%
AI boosts productivity	81%	—
AI reduces mental load	86%	—
Use AI for advisory	93%	—

What changed:

AI moved from "optional enhancement" to "competitive requirement." Firms not using AI daily are falling behind on speed, accuracy, and capacity.

2025: The Year AI Stopped Being a Pilot

The Agentic AI Emergence

2024 was Generative AI experimentation. 2025 was the Year of the Agent.

What's Different About Agentic AI

Traditional AI (including chatbots):

Waits for input
Processes single requests
Returns output
Stops

Agentic AI:

Monitors continuously
Identifies issues proactively
Reasons about multi-step solutions
Executes within guardrails
Learns from outcomes

As Mitch Rutledge, CEO of Vertice AI, put it:

"Agents, agents and more agents—all the leading fintechs are embracing agentic AI functions to empower consumers and staff members."

Where Agentic AI Deployed in 2025

Autonomous member service:

Red Rocks Credit Union launched "Roxie," an AI assistant handling account inquiries, processing requests, and routing complex issues—without waiting for human intervention
PenFed reported resolving 20% of member cases on first contact through AI agents

Self-driving back office:

Transaction monitoring
Automatic reconciliation
Regulatory report drafting
Compliance alert generation

Proactive financial guidance:

Spending pattern analysis
Automatic savings optimization
Real-time tax impact calculations
Cash flow predictions with intervention suggestions

The Guardrail Requirement

The institutions that deployed successfully weren't necessarily the most aggressive. They were the ones with proper controls:

"Many think they're ready for AI, but they're not. Guardrails, monitoring, oversight—most credit unions haven't even started." — Bill Illis, MDT

What successful deployment required:

Clear boundaries on AI authority
Human escalation triggers
Audit trails for all AI decisions
Continuous monitoring for drift
Regular model validation

The Widening Gap

Leaders vs. Laggards

The gap between AI leaders and laggards widened significantly in 2025:

AI Leaders:

Deployed multiple AI systems in production
Built internal AI expertise
Established governance frameworks
Accumulated months of operational data
Identified next-generation use cases

AI Laggards:

Still running pilot projects
Relying entirely on vendor AI
No formal governance
Limited data on AI performance
Uncertain about next steps

Why the gap is hard to close:

AI systems improve with use. The institutions that deployed in 2025 now have:

Training data from real interactions
Learned understanding of edge cases
Calibrated confidence thresholds
Staff experienced in AI oversight
Proven ROI to justify further investment

Those still piloting have none of this. Every month they delay, the gap widens.

The 93% Failure Rate Context

CCG Catalyst reported that 93% of bank digital transformation initiatives fell short of expectations.

But here's the nuance: AI initiatives were not the primary failure mode.

What failed:

Legacy system modernization projects
Large-scale core transformations
Integration-dependent initiatives
Projects requiring organizational change

What succeeded:

Targeted AI deployments with clear scope
Point solutions with measurable ROI
Incremental automation of specific workflows
AI tools deployed to prepared teams

The lesson: AI works when deployed with focus. It fails when treated as a magic solution for systemic problems.

What We Learned About AI That Works

1. Context Beats Capability

The most powerful AI models in the world still fail without context.

Example:

Generic AI: "Transaction of $14,457. Category: Miscellaneous."

Context-aware AI: "Conference expenses at Jack Henry Connect, September 13-15. Categories: Travel ($4,200), Lodging ($3,800), Meals ($2,100), Booth ($4,357). Total deductible: $14,457."

Same data. Radically different value.

2. Guardrails Enable Speed

Institutions with strong governance deployed faster, not slower.

Why:

Clear boundaries mean less fear. When everyone knows what AI can and can't do:

Approval processes are faster
Scope creep is prevented
Failures are contained
Learning is systematic

3. Small Wins Beat Big Visions

The successful AI deployments of 2025 were often modest:

Automate this one report
Handle these routine inquiries
Categorize these transactions
Draft these responses

Institutions that tried to "transform everything with AI" mostly ended up with PowerPoints.

4. Human + AI > Either Alone

The most effective implementations kept humans in the loop:

AI handles routine cases; humans handle exceptions
AI drafts; humans approve
AI analyzes; humans decide
AI suggests; humans verify

Pure automation without oversight created risks. Pure manual processes missed the efficiency gains. The sweet spot was hybrid.

What This Means for 2026

For Financial Institutions

If you deployed in 2025:

Expand scope based on what worked
Build on your data advantage
Develop next-generation use cases
Share learnings across the organization

If you're still piloting:

Pick one high-value, well-scoped use case
Deploy in production (with guardrails)
Accumulate real operational experience
Accept that you're playing catch-up

For Small Businesses

What to expect:

Your bank/credit union may offer AI-powered services
Your accountant likely uses AI (even if they don't say so)
AI-native financial tools are increasingly available
Real-time, context-aware services are becoming the norm

What to look for:

Systems that understand your business context
Proactive insights, not just reactive reporting
Integration across your financial ecosystem
Clear explanations of AI-driven recommendations

For the Industry

2025 established that AI is infrastructure, not innovation.

The conversation in 2026 won't be "should we use AI?" It will be "how do we use AI better than our competitors?"

This shift has implications:

Vendor selection will prioritize AI capabilities
Talent strategies must include AI expertise
Governance becomes a competitive advantage
Data quality determines AI effectiveness

The Uncomfortable Truth

Here's what nobody wants to say publicly:

Many institutions that claim to be "AI-first" are actually "AI-adjacent." They have AI features in vendor products. They've run demos. They've attended conferences.

But they haven't deployed AI in a way that changes their operations.

The institutions that actually deployed—that accumulated real data, learned real lessons, and proved real ROI—have a head start that will be very hard to overcome.

The pilot phase is over. The deployment phase has begun.

Those still experimenting need to decide: join the leaders, or accept the growing gap.

I'm Anna Khalzova, Co-founder of Jupid. We built Jupid to be AI-native from day one—not AI-assisted, not AI-enhanced, but AI at the core. Because we believe that's what small businesses deserve, and that's where finance is headed. If you're interested in what AI-native accounting looks like, let's connect.