Most credit unions are already using AI somewhere in the marketing function, just not for the thing that would move the needle most. Depending on size and tech maturity, it's usually meeting-note summarization, first drafts of marketing copy and email, or some basic member analytics. All useful. None of it touches strategy. This piece is about the next step: using AI to make better strategic decisions, not just faster tactical ones.
The data is already there and it doesn't talk to itself
Every credit union sits on data that's disconnected by default. Member survey data from vendors like Raddon. Competitive intelligence on the banks and credit unions you compete with. Performance benchmarking from Callahan. Your own product roadmap, goals, and plans. Individually, each of these tells you something. Cross-referenced against each other, they tell you something much more useful, and doing that cross-referencing by hand, as a human, is a genuinely time-consuming, error-prone exercise. That's the problem worth solving. A lot of strategic decisions at credit unions get made on a single data point sitting in its own silo, when several other data points that could change the outcome are sitting right next to it, unconnected. The need isn't more data. It's a way to incorporate the data you already have, all at once, without it taking days. I first built a version of this at a previous credit union using ChatGPT's Projects feature. At the Pacific Northwest credit union where I now run marketing, I've rebuilt and expanded the same system using Claude and Perplexity instead.
How I built it
Claude has a feature called Projects that lets you load a set of reference files the model treats as permanent, persistent context for every conversation inside that project. I used it to build out a working knowledge base of the organization:
- Recent member survey data from Raddon
- A voice-recorded walkthrough of how our teams are structured
- Our existing strategic planning documents, including our long-range plan and our ideal-member personas
- Every current and legacy product, along with our own notes on what actually differentiates them, including where we might be wrong about that
- Our purpose, mission, vision, and values
- A competitive analysis I ran using a tool that reads competitors' social media, content, and websites and summarizes how they position themselves
- That same competitor list run through Callahan, so I had both the front-end story (how competitors present themselves publicly) and the back-end story (how they're performing)
- Notable partnerships and events, plus context on our SEG (Select Employee Group) makeup, since we're a SEG-based credit union, including the incentives we offer SEGs and examples of specific groups
- Product brochures and collateral
One deliberate exclusion: member data. No names, no account information, nothing member-identifiable, only aggregated, anonymous survey responses. Before loading anything, I confirmed with our data governance team that what I was uploading, mostly public or already-aggregated information, was appropriate to expose, and I made sure the setting that opts our data out of being used for model training was turned on. Worth flagging clearly for anyone doing this: even with that setting on, there's still some residual risk of exposure, so treat anything sensitive as off-limits until you have a better answer than "I turned off a toggle." Before trusting it with anything strategic, I tested it. I'd ask basic questions, like how our checking product works, or what we know about a given SEG, specifically to catch anything missing, outdated, or wrong in what I'd loaded. That validation step matters. Don't skip it.
What it did for actual planning
Once I trusted what it knew, I used it for real planning at the start of the year. I asked it to draft a product marketing plan, and because it already knew our products, how they stacked up against competitors, and what members thought of them, the first draft came back genuinely informed instead of generic. A few rounds of back-and-forth (change this, don't touch that, we can't fix this particular pain point right now) got it to something actionable after three or four passes. The whole cycle took maybe 20 to 30 minutes, for work that would otherwise take hours if not days. Once the base files are loaded, which itself only takes about 10 minutes if you already have the documents on hand, each new plan or analysis costs about 20 to 30 minutes of iteration. Multiply that across a year of planning cycles and ad hoc requests, and the hours saved add up fast. It's not limited to formal planning either. My team covers business development, community engagement and events, and brand and communications, and other departments regularly bring us requests or ideas. I can validate those quickly against the model instead of pulling together a team discussion from scratch. Giving the tool even a simple internal shorthand, so people can reference it by name instead of describing it as "the AI thing," made it easier for leadership and my direct reports to engage with it as something real and specific rather than an abstract project. One more habit worth mentioning: I often talk through a request out loud and dictate it rather than typing, which makes it faster to explain what I need.
Expanding it in the moment
The standing knowledge base isn't the whole story. When we shifted card networks from Mastercard to Visa this year, I pulled PDFs of articles from Credit Union Times and BrandForum covering card-transition best practices and dropped them into that specific conversation. The base files gave the model institutional context; the added documents gave it situational context. That combination, a stable core plus topic-specific additions on the fly, is where this actually starts to pay off.
Why Claude and Perplexity together
I use both, for different reasons. Claude's Projects feature holds the core knowledge base and does the heavier strategic reasoning, and it has a real edge over ChatGPT here: rather than a hard cap on the number of files (ChatGPT limits Projects to 25 files on Plus and 40 on its business tiers), Claude allows effectively unlimited files, each up to 30MB, and once your total content exceeds its context window, it automatically shifts into a retrieval mode rather than blocking new uploads. Perplexity fills a different role. Its equivalent feature, Spaces, accepts links as a resource, so I've added our own website, which gets crawled periodically for product and brand updates, along with industry sites like TheFinancialBrand.com and Credit Union Times. I've also used Perplexity's Deep Research function, and its newer Labs feature, which turns a prompt into a finished document, slide deck, dashboard, or spreadsheet. Just as important, moving this work into a shared space let my team work alongside me instead of routing every request through me. Watching someone on the team get their own "wait, that was fast" or "oh, I hadn't thought of that" moment did more for internal AI adoption than any training session could have.
Guardrails
None of this works without discipline around a few things:
- No sensitive or member-identifiable data, full stop, regardless of what training-data settings you've toggled off.
- Keep a human in the loop. Don't take the first output at face value. Review it, push back on it, revise it.
- Change your language, not just your tools. We don't use AI here. We work with it. Treating it like Outlook or Excel, a tool you operate, limits what you get out of it. Treating it like a collaborator changes the outputs.
- Train your team on prompting. There's a lot of good material publicly available on how to prompt well, and you can even use AI itself to help draft better prompts.
Why not just Copilot?
Until recently the answer was easy: Copilot had no persistent workspace, so it couldn't hold a knowledge base at all. That changed this summer. Microsoft shipped Copilot Notebooks to general availability in mid-2026, which gives you a persistent container for references and chats, and Copilot Memory now carries context between sessions. So the gap is narrower than it was when I built this.
What I'd still say: Notebooks are a workspace, not a reasoning upgrade, and rollout is uneven by license and tenant, so check what your organization actually has before assuming it's there. If Copilot Notebooks is live in your tenant and your data governance answer for Claude is "no," start there instead of not starting. The setup in this piece is a pattern, not a product endorsement.
What we still can't do
This is still early. The next step is a properly walled-off environment where we can safely bring in member data itself rather than aggregated survey responses only, likely through a dedicated vendor rather than a general-purpose consumer AI product. We're also in active conversations as a leadership team about data governance and a formal AI policy, and a CRM or CDP is somewhere on our horizon, which would open up another level of this entirely. None of that is required to start, though.
You don't need a data lake to start
You need clean-enough data, a clear sense of what you're trying to decide, and a willingness to build the knowledge base once so you can reuse it constantly. If your organization isn't yet using AI for strategic decisions, find someone internally who's willing to dig in, or bring in someone from outside who works in this space. The time savings are real, but the bigger win is the quality of decisions you get when you're cross-referencing your data instead of guessing off one number in isolation.