This is a summary of an episode of Main Street AI, an educational podcast on AI led by our founder. Join 3,700+ business leaders and AI enthusiasts and be the first to know when new episodes go live. Subscribe to our newsletter here.

TL;DR:

• Agentic AI at credit unions splits into two tracks: personal productivity and partner-delivered back/middle office work
• Leadership has to visibly use the tools, not just approve them
• Uneven adoption is the real organizational risk, not slow adoption
• Member experience and lending are the live use cases; leadership funds revenue growth faster than cost efficiency
• Get one of the big three assistants (Microsoft Copilot, Anthropic's Claude, or OpenAI's ChatGPT) to every employee
• Agentic pricing is moving to usage and tokens — ask vendors for simulated usage numbers before signing
• A workflow should be live in two to six weeks; the spend to learn is a tuition expense

Before we dive into the key takeaways from this episode, be sure to catch the full episode here:

The Two Categories of Agentic AI Every Credit Union Is Now Running

Patel splits what's happening inside credit unions into two tracks, budgeted and measured differently. The first is personal productivity: employees connect Slack or Teams, email, call notes, and a storage drive to pull context and draft material. The second arrives through existing software partners.

"That is more behind the scenes agentic AI delivered through software partners that you have to make work easier, faster, better." — Ankur Patel

In 2025, agentic tooling was mostly an engineering phenomenon, concentrated in Cursor and Claude Code. "The beginning of 2026 was a big wake-up call for all business people, but especially at credit unions, around what agentic really means," Patel says.

Leadership Has to Use It, Not Just Approve It

Ankur's answer to how institutions build adoption isn't a memo.

"It can't be just them saying that people should use agentic AI. They need to show and teach and train as well."

Where leadership visibly spends the effort to learn, the organizations do better, he says. The second mechanism is a small internal group, stewards or shepherds, who bridge the conversation for everyone else, since most people won't pick this up on their own.

The Split That Uneven Adoption Creates

His real concern isn't slow adoption across the board. It's when some teams move and others don't.

"When there is uneven adoption, that's where we see a lot of organizational challenges, because it's almost like there are two parts of the organization, but they're all playing to a different beat."

Getting the whole institution to a more uniform pace before 2027 execution begins is the work left on the table this year.

Where the Real Use Cases Are: Member Experience and Lending

Member experience comes first because it's what credit unions protect most: personalized interaction, faster responses, staff empowered by agentic tools. "The things that make credit unions special and how they treat members, that still needs to come forth," Ankur says.

Lending is the second, and the most direct case. Getting back to a borrower fast with an appropriate quote wins business across indirect auto, mortgage, commercial, and HELOC.

"That's a slam dunk of a use case, because it's allowing your existing team to deliver better products, better services."

Revenue growth funds faster than cost efficiency at most institutions. "They get far more excited by potential revenue growth versus just cost efficiency," Patel adds.

Three Things to Fund for 2027

Ankur gives a sequence, not a shopping list. Get one of the big three assistants — Microsoft Copilot, Anthropic's Claude, or OpenAI's ChatGPT — to everyone first, with clarity on what data can connect and what queries are appropriate. Second, audit whether incumbent vendors' agentic features stay locked inside their own product. Third, partner with a few natively agentic startups, since they can move faster than incumbents carrying legacy product lines.

"I would really put the existing incumbent vendors through the wringer. Just because you have them today doesn't mean they're the right partner for the next five to ten years."

The Walled Garden Problem

Defensive incumbents are responding to the shift with fewer integrations, thin APIs, and features that only work inside their own loan origination software.

"The true power of agentic AI is for you to be able to traverse all of your software that you use. It allows you to operate without boundaries, without borders."

The wariness tracks with what credit unions report industry-wide: among top-tier credit unions, 76% said external partners were helping deliver digital onboarding and authentication capabilities, while 73% reported using partners to develop new payment user experiences, per PYMNTS. Vendors that block that kind of partnership are working against where the industry is already headed.

Questions to Ask AI Vendors About Token Pricing

Every vendor has access to similar underlying models, so the model isn't the differentiator.

"What matters most is what sort of harness. The harness, it used to be called wrapper, and I think harness is a better term."

His math: take the real volume of work in the operation you're automating and multiply by the vendor's token pricing to get total workload cost before you sign, not after. Get simulated usage numbers at your projected volume, and ask for the same projections at two and three times growth.

The New ROI Bar: Two to Six Weeks

The old tolerance was six months to see an efficiency gain. Patel's benchmark now is much tighter.

"There is a tuition expense to agentic AI, which is you need to try and test and spend some to learn. You should get a workflow live within two to six weeks."

The integration test that separates a real deployment from a demo: if staff still have to log into a separate application to load and review things, the workflow isn't actually live.

How This Works in Practice

Ankur ties his own team's pricing to the same standard. "Your pricing should be commensurate with the ROI that you deliver. If the workflows we deliver don't add business value, our customers won't use it, and if they don't use it, we also don't get paid much." Multimodal's AgentFlow platform is built around that test, pre-built, auditable workflows for credit union lending and member experience meant to prove value inside a two to six week window instead of a six month pilot.

Want more on financial services and AI? Check other episodes here.

Frequently Asked Questions

1. How should a credit union budget for AI in 2027?
Fund three tracks: a general assistant for every employee, an audit of incumbent vendors, and a few experiments with natively agentic startups. Treat the spend as tuition that buys organizational learning.

2. What questions should I ask an AI vendor before signing?
Ask for simulated usage numbers at your real workload volume, not just the entry price. Multiply that volume by the vendor's token pricing to get total cost, and ask what happens to that number at two and three times growth.

3. What is a walled garden in AI vendor software?
An agentic capability that only works inside a single vendor's own product. Warning signs include thin APIs, few integrations, and features locked to the vendor's own loan origination system.

4. How fast should a credit union see ROI from an AI workflow?
Two to six weeks to get a single workflow live. That replaces the older six-month expectation and forces faster learning about real cost and value.

5. What is an AI harness, and why does it matter more than the model?
The application layer around the model: routing, integrations, auditability, human-in-the-loop review. Every vendor uses similar underlying models, so the harness is what actually differs.

6. /Why does uneven AI adoption hurt a credit union?
It splits the institution into teams moving at different speeds. The fix is leadership that visibly uses the tools plus internal stewards who bridge the gap for slower-adopting teams.

‍

‍