What marketers in financial services are learning about AI — and what's surprising them

There's a version of the AI conversation in financial services that goes like this: leadership sends a memo, a tool gets approved, someone books a training, and six months later nothing has really changed. The productivity gains haven't materialized. The enthusiasm has faded. And the team is quietly using ChatGPT for emails while pretending they're not.

We've heard this story enough times that we decided to do something about it.

Earlier this year, we ran the first cohort of Opal University for Financial Services, a five-day live program built specifically for marketing and digital leaders working inside one of the most compliance-conscious industries in the world. Fifty seats. One hour of live training followed by one hour of hands-on agent building, every day, for a week. A global cohort — participants joined from the US, the UK, and Europe — all working through the same challenges in the same room.

The goal wasn't to teach people what AI is. It was to get them building things that work inside the constraints they actually operate in.

What came out of those five sessions was more interesting than we expected. Not because participants built extraordinary things — though some of them did — but because of the patterns that emerged across the group. The same tensions, the same blockers, the same breakthroughs, surfacing again and again regardless of institution size, geography, or seniority.

This piece is an attempt to name those patterns honestly. Not as a recruitment pitch, and not as a product showcase. As a synthesis of what we actually heard, from the people in the room.

Pattern 1: The confidence gap — wanting to move, waiting for permission that keeps moving

Almost every participant arrived carrying some version of the same tension. They wanted to use AI more. They could see where it would help. But their organization was still building the governance framework, still deciding what was permissible, still waiting for legal and compliance to sign off on something — anything.

One participant, a senior operations lead at a large US bank, described their organization's first AI governance meeting with a clarity that the whole room recognized:

Another, from a US-based credit union, put it more bluntly:

What was striking wasn't the conservatism, that's expected and, frankly, appropriate in financial services. What was striking was how isolated people felt about it. The LinkedIn algorithm has spent two years telling everyone that the AI revolution is already over, that the winners have been decided, that if you're not running 64 agents you're behind. The cohort was a corrective. Participants discovered that the person sitting metaphorically next to them — at a different bank, a different credit union, a different asset manager — was navigating the exact same governance maze. Nobody was as far behind as they thought.

The confidence gap, it turned out, wasn't really about capability. It was about context. Once people understood they were all at the same stage, the anxiety dropped and the building started.

The practical insight that helped most: stop waiting for permission to automate everything, and start building things that don't require it. Personal skills, private agents, tools that improve your own output without touching any shared data or crossing any compliance line. The wins were small at first. But they were real and real wins, it turns out, are the currency you need to start shifting the conversation with the people who hold the governance keys.

Pattern 2: The 80% problem — AI gets you there fast, the last 15% is all you

Every participant who had spent any time with AI before arriving already knew about the 80% problem, even if they didn't have a name for it. They'd asked a tool to write something, or generate a report, or draft a brief, and the output was... fine. Competent. Technically correct. And completely devoid of the judgment, nuance, and institutional knowledge that would have made it actually useful.

What the cohort surfaced was how much of the problem sits upstream, not in the output itself.

The most common failure mode wasn't bad AI. It was under-briefed AI. Participants who gave their agents thin context got thin results. Participants who invested time feeding in past work — real reports, real examples, real tone-of-voice documents — got outputs that were genuinely close to what they'd have produced themselves. The brief quality was the bottleneck. It always had been; AI just made the gap visible.

The harder conversation was about what happens at the 80% mark, when the output looks good enough to send but isn't quite right. Several participants admitted their organizations had already been through one failed AI pilot, usually because someone rolled out a tool, people used it to generate emails or reports, and the outputs were shared without enough human review. The quality was mediocre. People noticed. Trust eroded.

Getting from 80% to 95% is, paradoxically, where human expertise matters most. Not at the generation stage but at the iteration stage. Knowing when the output is wrong. Knowing why it's wrong. Knowing which part of the brief to change to fix it. These are judgment calls that require domain knowledge, and domain knowledge is exactly what experienced financial services marketers have in abundance.

The participants who made the most progress during the week were the ones who reframed what their job was. Not "can I get AI to do this?" but "what does AI need from me to do this well?" The former treats AI as a vending machine. The latter treats it as a very capable but context-blind collaborator which, at the moment, is much closer to what it actually is.

Pattern 3: Starting small, winning big — the specific beats the ambitious every time

On Day 1, participants were asked to think about what they wanted to build. The instinct, almost universally, was to go big. An end-to-end content production workflow. A competitive intelligence system covering 30 competitors. A full agent stack that would handle everything from research to publishing to reporting, automatically, on a schedule.

By Day 3, the participants making the most progress were the ones who'd abandoned the big idea in favor of something embarrassingly specific.

A one-person marketing department at a London consultancy — covering wealth management, asset management, and investment banking content with no team to support her — didn't try to automate her whole content operation. She built three research agents, one per audience, that surfaced relevant topics she could then choose from. Specific, contained, useful from day one.

A head of product and range planning at a UK gifting business didn't try to automate all competitor monitoring. She built an agent that tracked SKU count and price movement for one specific competitor, by category, once a week. That agent — still in test mode, not fully finished — surfaced something her entire team had missed.

A digital leader inheriting a CMS with two years of undocumented development didn't try to build a documentation system. He built an agent that took sprint notes and converted them into readable editor guides. One input, one output, repeatable.

The pattern held across every participant who reported a genuine win: they had started with something small enough to test in a day, specific enough to know whether it worked, and painful enough that they actually cared about fixing it. The ambition wasn't wrong, it just needed to come later, once the foundation was solid.

The participants who struggled were, almost without exception, the ones who tried to build the whole system at once. Not because the system was impossible — it wasn't — but because when something went wrong (and something always went wrong), they couldn't tell which part had broken.

Pattern 4: The unexpected dividend — agents finding problems humans had missed

This was the finding that surprised us most, and the one that generated the most energy in the room.

Several participants discovered that their agents weren't just doing the task they'd been built for. They were surfacing information that no human had thought to look for or had simply never had the capacity to find.

The clearest example came from a head of product range planning, who built a competitor product comparison agent primarily to track pricing. The agent — while still being refined, before she'd even finished building it — flagged that her company's single best-selling product had a significant availability gap compared to the main competitor. A booking restriction that meant customers couldn't complete a purchase on the most popular shopping day of the week. In a business turning over significant revenue annually, this had been invisible.

A similar pattern emerged for a content strategist who asked an agent to identify timely topic opportunities for an under-resourced section of her company's site. The agent surfaced three ideas. All three were validated by internal stakeholders as not just relevant but actively useful — one of them even becoming the basis of a conversation about whether an expensive video production that had been planned for six months out was actually worth doing.

The unexpected dividend isn't about AI being smarter than humans. It's about attention. Humans are finite. We prioritize. We have meeting agendas and sprint deadlines and things that are more urgent than the thing we know we should probably check. Agents don't have that problem. An agent can be set to look at 50 competitor pages, every week, and notice when something has changed consistently, without fatigue, without the competing demands that cause humans to skip the things that feel like maintenance.

For financial services organizations, where the cost of missing a regulatory change, a competitor move, or a market shift can be significant, this is not a trivial capability.

What we're still figuring out

It would be dishonest to wrap this up with only the wins.

There are real gaps that the cohort surfaced — problems we don't have clean solutions to yet.

The manual review wall. You can build an agent that pre-checks compliance content before it reaches legal. What you cannot do, at least not yet, is change the fact that legal still needs to review it. The agent saves time on your side of the process. It doesn't change the downstream requirement. For organizations where the legal review bottleneck is the primary constraint, that's a meaningful limitation.

Governance frameworks that keep moving. Several participants were building in genuine uncertainty, their organization's AI policy was still being written. That's not a technology problem. But it does mean that anything they build now may need to be rebuilt or adjusted once the policy lands.

Connecting everything. The vision — data flows in automatically, agents run, insights appear — is real, and getting closer. The current reality still involves some copy-pasting and manual steps, especially for platforms and data sources that don't yet have native connectors. For participants who were hoping to fully automate a reporting workflow, this was sometimes a disappointing constraint.

These aren't reasons not to start. But they are reasons to be honest about what "AI in financial services marketing" looks like right now, versus what it will look like in 18 months.

A note on who this is for

Opal University for Financial Services runs five days, with fifty seats per cohort, limited to participants from within the industry. The deliberate mix — different institutions, different markets, different team sizes, all with the same regulatory context — is part of what makes it work. People can talk honestly about the constraints they're operating under without having to explain the basics. They can share what's working and what isn't without worrying about giving away competitive intelligence.

Every cohort produces a community. Several participants from this cohort ended up sharing prompts, comparing agent outputs, and exchanging ideas through the week and beyond. One participant put it well, describing the moment she showed her team the output from an agent she'd built:

If you're a marketing or digital leader in financial services who recognizes any of the patterns above — the governance uncertainty, the quality gap, the feeling that the ambition is there but the roadmap isn't — that's exactly what the next cohort is designed for.

Not to tell you AI is easy. But to help you find the one thing worth building first.

Opal University for Financial Services runs as a live, cohort-based program with limited places per intake. Each session is designed for senior practitioners, not beginners — and the work happens during the sessions, not just in theory.

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