What marketers in financial services are learning about AI

Financial services marketing has an AI problem that nobody talks about honestly.

It's not that the tools aren't good enough. It's not that leadership doesn't care. It's that most teams are stuck somewhere between "we should really be doing something with AI" and "compliance won't let us touch it." Meanwhile, the to-do list keeps growing, the headcount doesn't, and every second LinkedIn post makes you feel like you missed the boat.

Over the past several months, Optimizely ran dedicated Opal University cohorts for marketing and digital leaders in financial services — banks, insurers, credit unions, wealth managers, fintechs. Five days. One hour of live training per day. Three working AI agents built by the end of the week. Participants came from the US, UK, continental Europe, and beyond. Some were Optimizely customers. Many weren't. The only requirement: willingness to get your hands dirty.

This piece captures the patterns that kept showing up. Not theory. Not product marketing. The actual themes, pain points, and breakthroughs that surfaced when compliance-heavy teams stopped debating AI and started building with it.

Pattern 1: Everyone thinks they're behind. Almost nobody actually is.

The single most consistent feeling across every cohort was anxiety. Participants arrived convinced their competitors were miles ahead on AI adoption. That their teams were too slow, too conservative, too tangled in governance to catch up.

The reality, once people started talking openly, was the opposite. Almost every financial services team — regardless of size, region, or sophistication — is navigating the exact same uncertainty. Governance frameworks are still being written. AI policies are moving targets. Internal risk teams are asking questions nobody has clean answers for yet.

"They were literally building the plane in the meeting, talking about what model this is, whether data is shared, what is within our risk tolerance. Everybody was throwing out new questions and trying to figure out what their bare minimum was."— Senior operations lead, large US bank

One credit union marketer put it more bluntly:

"I'm not really seeing that sudden AI push happening here. We're still a little more conservative. Legal and compliance reviews are still all manual."

The pattern was the same from London to Anchorage to Madrid: marketing teams are interested, often excited, but stuck behind layers of internal approval before they can operationalise anything. Knowing that this is an industry-wide reality — not a personal failure — changed the energy in every room.

As Giulia Camargo, the programme's lead educator, framed it in every opening session: "The fear of irrelevance is the strongest shame trigger in our professional lives. If we're constantly stressed about what the future holds, it's really hard to get your learning hats on."

Pattern 2: Compliance isn't the blocker people think it is — but governance is a 9-month sport

The assumption walking in was that regulated industries simply can't use AI. The reality that emerged across cohorts was more nuanced: you absolutely can, but you have to be strategic about what you build first and how you frame it internally.

One participant — a digital marketing leader at a major bank — shared that her AI governance approval process had taken nine months from first proposal to production sign-off. Not because Opal or the underlying technology failed a security review, but because the internal AI governance body required evidence across five domains: data classification, human oversight, model design and output accuracy, API integration security, and third-party due diligence.

"Just the mere fact that I had my lawyer's first name and last name in the system — because they have access to CMP to review content — that was classified as confidential. So even though we're reviewing public-facing marketing material, that single data point flagged the entire use case."

Her advice to peers was direct: look at your data classification first. Understand what your AI governance team considers restricted, confidential, and public. Then design your first use case to sit entirely within the public or internal-only tier.

The winning approach she described: pitch your first agent as automation, not generation. "I told them: this is the same thing we used to call automation. We're using an AI engine to do it, but it's still automation. We are not generating content. We are not manipulating data. We are triaging work requests." That framing — automation within an existing tool, not a new generative AI experiment — is what got her through the governance gate.

"My goal is to start on the automation side and build confidence with the teams. If I can show Opal can do X, well then — let's test it with Y. It's not dissimilar to when nobody had websites, and then we started showing traffic."

For teams still early in the governance conversation, participants highlighted two critical accelerators: the product's data privacy architecture (data stays in your own instance; no LLM training on your content) and having vendor documentation ready before the risk review meeting, not during it.

Pattern 3: The 80% trap — AI gets you there fast, the last 15% is where careers are made

Every cohort hit the same wall on day two: AI generates content quickly, but the output sounds like… AI. Generic. Competent. Forgettable. Participants who had been using ChatGPT or Copilot for months recognised the problem immediately.

"A lot of times nowadays, people are sending just AI outputs to each other. It's not useful, it's not good, and it kind of annoys everyone — you didn't synthesise the information, you didn't read it."

The breakthrough came when participants realised that the quality gap isn't a technology limitation — it's a briefing problem. Most people tell AI what to do but never show it what good looks like. They don't feed it past work. They don't provide tone-of-voice documentation. They don't give examples of content that actually performed.

In Opal, this is where skills made the difference. Participants built skills from brand guidelines, compliance rejection letters, approved product language, and winning campaign examples. Once those were attached to agents, the output quality jumped dramatically — not because the AI got smarter, but because it finally had the right context.

One participant who works in pensions and insurance summed it up: "I shared it with my team and there was a bit of a split. One of them was like, you can just ask ChatGPT for that, can't you? So I made him open ChatGPT and try. It wasn't even 80% of the way there. That was the moment."

The practical lesson: if you have an hour to spend on AI, spend 50 minutes improving a skill and 10 minutes generating content. The quality of your inputs determines everything downstream.

Pattern 4: Start with one boring task. Not a transformation programme.

The participants who made the most progress during the week all had one thing in common: they picked something small, specific, and immediately useful. Not "reimagine our content supply chain." Not "automate the entire campaign workflow." Just one task that eats time every week and produces predictable output.

Here's what people actually built:

A GA4 growth analyst — A digital strategist working with 20+ healthcare clients built an agent that connects to Google Analytics, surfaces traffic and conversion patterns, and delivers a prioritised action plan with concrete next steps. Work that previously took her team 5–20 hours per project now runs in minutes. "It's probably going to save me at least 10 hours a week."

A competitor messaging mapper — A data and automation lead at a UK pensions firm built a multi-agent workflow that scrapes competitor URLs, screenshots pages, extracts messaging and tone, compares it side-by-side to his company's positioning, and produces monthly competitor profiles. He'd been trying to build this in Copilot for six months. In Opal, it took a week. "I've had a specific problem. I've been playing with Copilot for about a year now. There's one thing I've been trying to do, which is create a full competitor analysis agent, and I just haven't been able to do in Copilot. I've done it in a week."

A compliance rejection recovery advisor — A marketing information manager built an agent that analyses previously rejected content, maps the rejection reasons against brand and compliance guidelines, and produces a pre-submission checklist for new team members. Instead of junior marketers repeating the same compliance mistakes, the agent catches them before the content ever reaches legal.

A campaign landing page builder — A health system marketer built an agent that takes a campaign brief and generates headlines, benefit bullets, image direction, and UTM tags — all mapped against existing brand skills and persona documentation. The output became a ready-made discussion document for stakeholder buy-in, cutting the pre-production research phase from days to minutes.

A go-to-market positioning analyser — A marketing consultant built a workflow that takes client and competitor URLs, runs a full landscape analysis, produces a SWOT breakdown, and generates a positioning strategy — consolidating work that previously required multiple tools and manual research into a single Opal workflow.

The pattern: nobody who tried to automate everything in week one succeeded. Everyone who picked one painful, specific bottleneck walked away with something working.

Pattern 5: The things nobody expected — agents finding what humans missed

This was the most interesting finding across cohorts, and participants kept coming back to it.

When you build an agent to do a task you've always done manually — competitive monitoring, content auditing, compliance checking — the agent doesn't just replicate your work. It catches things you never had the bandwidth to notice.

"So many people look at that product every day, and literally no one had noticed. And that's a tool that I haven't even finished building yet."

A competitor analysis agent discovered pricing anomalies. A GEO audit agent flagged missing schema markup and broken tracking that had been live for months. A compliance pre-checker surfaced inconsistencies between how different members of the same legal team were interpreting the same rules.

The reason is simple: humans skip maintenance checks when they're busy. AI doesn't get busy. It runs the same checklist every time, with the same attention, whether it's Monday morning or Friday at 5pm.

For regulated industries, this is a quiet but significant shift. The value of AI agents in financial services isn't just speed. It's coverage. It's the things that fall through the cracks when teams are stretched thin.

What we're still figuring out

Honesty matters more than hype. Here's what's not solved yet:

The manual review wall. AI agents can pre-check compliance, flag risks, and suggest fixes. But in every regulated institution represented in these cohorts, legal still needs to manually sign off on anything customer-facing. The agent gets you to the review faster — but it doesn't eliminate the review.

Governance frameworks that keep moving. Several participants described governance policies that changed mid-implementation. New AI oversight bodies forming. New risk categories being defined. Agents that passed review in March needing re-evaluation by June. This isn't a technology problem. It's an organisational maturity problem, and it's industry-wide.

The reporting and audit gap. When you put an agent into production in a regulated environment, you need to prove what it did and why. Execution logs exist, but making them human-readable, auditable, and efficient to run without slowing down the agent itself is still a work in progress. One participant described building an entire second agent just to format and interpret the first agent's execution logs for her compliance team.

Connecting everything. Full end-to-end automation — from brief to published content — still requires manual steps in most environments. CMS connectors, CRM integrations, and marketing automation platform links are getting better, but for many teams, there's still copy-pasting between systems.

Consistency across compliance reviewers. Multiple participants noted that the biggest variable in their compliance process isn't the AI — it's the humans reviewing the AI's output. Different reviewers applying different interpretations of the same rules. One marketer said it plainly: "I struggled even finding consistent examples because our compliance team, depending on who is reviewing it, has different feedback."

What this means for financial services marketing leaders

The marketers who joined these cohorts weren't AI engineers. They were senior marketing managers, heads of digital, content leads, and marketing ops professionals. Most had played with ChatGPT. Some had rolled out Copilot. A few had never used AI beyond a basic prompt.

By the end of the week, every single one of them had built at least one working agent that addressed a real problem in their workflow. Many built three. Some built entire multi-agent workflows that now run on a schedule without them touching a thing.

The shift wasn't technical. It was a shift in framing: from "AI is something IT gives us" to "AI is something I build for myself." From waiting for the perfect enterprise rollout to starting with one skill, one agent, one task that saves an hour this week.

"What you build now as a marketer becomes your resume. Your agents, your workflows, your systems — they're basically how your next opportunity finds you."

The five-day format worked because it removed the two biggest barriers: fear and access. A safe space to ask questions without judgment. Free credits so nobody had to file a procurement request. Peers from the same industry navigating the same constraints. And a simple, structured path from "I don't know where to start" to "I just built something that works."

Financial services marketing isn't behind on AI. It's behind on permission — permission from governance, from leadership, and most importantly, from themselves.

The teams that are pulling ahead aren't waiting for that permission. They're building in the margins, proving value with small wins, and expanding from there.

Opal University runs dedicated Financial Services cohorts throughout the year. Five days, one hour live per day, three working agents by the end. Free. No technical skills required.

Sign up here if you'd like to join the next one!