Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter curates critical developments in agentic AI and explains why they matter for your credit union.

Goldman Sachs disclosed six months of agent development using Anthropic's Claude for accounting and compliance. 🌱 CUltivate, a new CUSO, launched a foundation model built on cooperative principles. Anthropic also released enterprise plug-ins aimed at replacing finance and HR software.

The U.S. Treasury published its first AI risk framework for financial services. Credit unions now have a purpose-built model, Wall Street blueprints, and federal guidance arriving in the same week.

🌱 CUltivate gets a full analysis at the bottom of this issue. ⬇️

Banking Dive · Feb 18, 2026

TD Bank's U.S. head of AI, Ted Paris, says successful AI agent projects require rethinking full business processes. TD implemented 75 AI use cases generating $170 million in value in 2025 and expects $200 million in 2026.

Why it matters: TD found that AI failures stem from poor process integration, a lesson credit union leaders should weigh before deploying agents.

CNBC

Goldman Sachs has spent six months co-developing AI agents with Anthropic's Claude model. The agents target trade accounting and client onboarding. The firm expects efficiency gains and faster processes, with more use cases planned.

Why it matters: If Claude handles rules-based accounting and compliance at Goldman, similar agents will soon reach FIs of every size.

Captain Compliance • Feb 25, 2026

The U.S. Treasury published two new AI resources: an AI Lexicon and a Financial Services AI Risk Management Framework. The documents aim to help FIs adopt AI more safely with practical, sector-specific guidance.

Why it matters: Credit union compliance teams should review the Treasury's new AI risk framework before expanding any AI initiatives.

Whether you're weighing credit-union-specific AI platforms, rethinking process design the way TD Bank has, or just trying to get your arms around the security and governance questions that come with agentic deployments, I'd love to hear what's shaping your thinking right now. Drop me a note at [email protected].

Know someone who should be reading this? Send them the subscribe link.

TL;DR:

Ask AI a compliance question at a credit union and the first few minutes typically go to context-setting. What's a credit union? How does NCUA oversight work? What's a field of membership? The model catches up, then it answers.

CUltivate, which launched February 24, is built to skip that. It's an AI model trained on Filene research, National Credit Union Administration (NCUA) guidance, and other credit union content, structured as a majority credit union-owned Credit Union Service Organization (CUSO) with Filene Research Institute, Vertice AI, and Vizo Financial as founding partners. An advisory committee oversees training data selection. I imagine a compliance question might land the way it would with a 20-year movement veteran.

Credit unions have a long history of building shared infrastructure when outside alternatives don't fit. This is that instinct applied to an LLM, and it could be a big deal.

The press release and website are light on technical details, which is normal for a launch week. Here's what I'm curious about...

What's it built on?

Why it matters: whether it's fine-tuned or built from scratch determines how it reasons, and how often it needs updating.

Is this a fine-tuned version of an existing model (Llama? Mistral? DeepSeek?), or was it pre-trained from scratch? The distinction matters more than it sounds.

Fine-tuning takes an existing model and specializes it. The underlying reasoning stays borrowed from whatever model is underneath. That works fine for a lot of tasks. But for something like fair lending rules under the Equal Credit Opportunity Act (ECOA), the model might be drawing on general legal patterns from its base training, not the regulation itself. There's no way to tell from the outside.

Training from scratch is the other path. More precise on what it trained on, but it needs full retraining to absorb regulatory updates, not just the cheaper re-fine-tuning option.

Given Vertice AI's background in analytics and personalization rather than LLM research, the website's own language about "leveraging existing AI models" is a tell. Fine-tuning on top of an existing base model is the more likely scenario. I'd want to know which base model it sits on, what it was trained on, and how often it gets updated.

What does the training data cover?

Why it matters: "credit union content" covers a lot of ground. The sources matter as much as the volume.

"Credit union content" covers a lot of ground. The composition of the training set (what sources, at what depth, with what recency) determines what the model is actually authoritative on. That shapes which use cases make sense to start with.

Recency is a separate question. NCUA issues guidance regularly, and a model with a training cutoff can be behind on current requirements in areas like overdraft programs, member business lending, and how CFPB rules apply to credit unions.Trickier though, is that the model won't flag which parts are stale. How often CUltivate retrains, and on what trigger, is something I'm genuinely curious about.

How does it handle regulatory edge cases?

Why it matters: on compliance questions, a confident wrong answer looks identical to a correct one.

The worst-case failure mode is a confident wrong answer. On ECOA, Bank Secrecy Act / Anti-Money Laundering (BSA/AML), or NCUA exam guidance, a fluent hallucination looks identical to a correct answer.

The key architectural question is whether CUltivate uses RAG (retrieval-augmented generation). A RAG system — which is basically connecting a knowledge base to a foundation model — looks up truth from source document before answering, instead of relying on what the model memorized during training. OpenEvidence does this for clinicians, citing sources inline so doctors can trace every claim to the underlying study. Whether CUltivate works the same way, or relies on weights alone, determines how much verification compliance teams should plan for.

Without retrieval against a live source, even a well-trained model can produce specific thresholds, effective dates, or agency guidance that sounds exactly right and is wrong. The fluency makes it worse: a confident wrong answer reads the same as a confident correct one.

Where's the right place to start?

Why it matters: domain-specific models have a real home turf. Knowing what it is helps pick the first use case.

This is a genuine use case. Summarizing exam findings, explaining established procedures, drafting from familiar templates. Standard member communications and policy drafting are solid starting points.

They get shaky on multi-regulation interactions, novel interpretive questions, and anything updated after their training cutoff. BSA scenarios and multi-regulation edge cases are where any model, domain-specific or general, needs the most human review.

How does it compare to general-purpose models?

Why it matters: CUltivate hasn’t shared public benchmarks yet. Here's what I’d love to see -

The performance of foundation models (like GPT, Claude, Gemini, etc) are compared using evidence-based, accepted benchmarks. A few relevant benchmarks for CUlitivate include:

I wouldn’t mind seeing how it performs in Vals.ai’s Finance Agent benchmarks. So far the highest performer is Claude Opus 4.6, with 67% accuracy. Leapfrogging the major models’ performance on financial benchmarks would be quite the coup.

Here's a case study worth knowing:

A domain-specific model's advantage over general-purpose LLM can shrink fast. The infrastructure built around it usually outlasts the model. Bloomberg found this out.

Bloomberg launched BloombergGPT in March 2023 with 50 billion parameters and 363 billion tokens of proprietary financial data. At launch, it outperformed GPT-3.5 on financial benchmarks by meaningful margins.

By 2024, GPT-4 had closed that gap. Bloomberg quietly stopped marketing it as a standalone product and absorbed it into the Terminal, where it reportedly still runs alongside newer general models.

The data pipelines, workflow integrations, and internal tooling persist. The model itself got replaced. The durable asset turns out to be the infrastructure, not the weights.

CUltivate has the advantage of knowing that history at launch. And it has something Bloomberg could never have built: a cooperative. Bloomberg's model was singular and centralized. A CUSO where credit unions contribute data, govern the model, and own the infrastructure collectively could keep the technology at the edge of performance vs overtaken.

The model will improve or get replaced. Models always do. The rails, the data governance, the value staying inside the system...these are what I'm the most excited about in this announcement. CUltivate's team has built a promising container for the industry to fill.

This is worth keeping an eye on. If your credit union is involved, I'd love to chat.

Some of this gets technical. It's easier to get started than it might seem. Let's talk: [email protected]

How this newsletter is made: Brent curates links, Claude (Opus 4.6) generates summaries and the intro, Brent edits, sometimes writes an editorial, and publishes on Beehiiv.

⚡ Alakazam ⚡.