The Approval Queue Is the Bottleneck: Why Bank Marketing Still Takes Weeks Despite AI

Generative AI has arrived in bank marketing departments. 63 percent of marketing teams at US banks and credit unions use it to produce content, according to a Cornerstone Advisors study commissioned by Persado, published at the end of September. Only 9 percent, by contrast, use AI agents that carry out an action themselves within approved rules. In between sit the 32 percent who apply AI to analytics and decision support.

The distribution is telling: adoption peaks in the step that generates words and drops off in the steps that decide and act. That is exactly where the bottleneck sits. The study names compliance review as one of the biggest obstacles to faster execution; at some institutions it takes four to six weeks. A campaign an AI drafted in minutes then waits in the same queue as one written by hand, while customer needs and competitive openings come and go within days.

The figures come from the US market. The mechanics are the same in Europe, with an additional layer of review through GDPR, consent and, since this year, supervisory oversight of customer-facing AI.

Agreement without priority

Marketing leaders know where the leverage is. Nearly two in three expect to be comfortable within two years with letting agents act on their own inside approved boundaries. 27 percent name campaign and journey orchestration as the highest-impact use, 25 percent offer and product recommendations. The expected payoff lies in deciding what goes to whom and when, not in writing it.

At the same time, 47 percent treat agentic AI as an experimental topic, and only 5 percent call it their top strategic priority. Between the expectation for the day after tomorrow and the budget for today there is a gap. For the compliance officer who would have to define those "approved boundaries", the definition is most of the work. As long as nobody does it, the agent stays a pilot.

The data underneath is not ready

A second finding explains why the gap is not just an organisational problem. In The Financial Brand's State of Financial Marketing research, Jim Marous reports that more than half of the institutions surveyed already use generative AI in marketing, yet not one said its customer data was immediately available and AI-ready. Roughly two thirds describe their data as scattered across systems or updated in batches. Almost nobody says real time.

The discrepancy becomes even clearer around the word personalisation. About one institution in a hundred says it truly delivers hyper-personalisation. Asked to define personalisation, around two thirds of the answers come down to selecting an audience for a message. That is segmentation, not personalisation. The customer never experiences the segmentation; they experience the offer, its timing, the onboarding and whether the bank seems to understand what is happening in their life.

The budget logic reinforces the problem. AI-driven content and creative rank last among the tactics respondents consider effective, yet receive one of the largest planned budget increases. Deposit growth is the number one priority, while primary-bank status, share of wallet, win-back and dormant accounts rank near the bottom. A great deal of energy goes into acquisition and comparatively little into what happens afterwards.

What is missing is attribution

Anyone who wants to claim that a campaign launched in hours earns more than one launched after six weeks has to be able to measure the effect. An earlier Cornerstone study from April, with 126 senior executives at US banks, shows that nearly six in ten say their core or CRM platform restricts how well they can measure marketing ROI. Not a single respondent could reliably attribute all six outcome measures surveyed, including account openings, funded accounts, balances and profitability.

For agents that depend on live signals, such as continuous segmentation or triggered outreach, this means they stall at the data layer before any compliance check is due. A core system can hold years of transaction history; if the bank cannot reach it at the moment a customer is ready to buy or about to leave, it cannot act.

Offers that ask for nothing

How this shows up in offer design is illustrated by Zafin's State of Offers research, picked up by The Financial Brand in early October. 48 percent of traditional financial institutions' offers use sophisticated mechanics such as partnership bundles, behavioural incentives or dynamically tailored rewards; at fintechs it is 65 percent, at community institutions 27 percent. Two thirds of traditional institutions' offers are barely targeted at all: they ask nothing more of the customer than a sign-up or occasional use.

The reason is less a lack of creativity than technical inertia. In many institutions, changing a product, rate or offer still means touching a decades-old core system; according to Zafin, time to market for a campaign runs three to nine months. The counter-movement is interesting: offers explicitly aimed at deepening existing relationships use sophisticated targeting in 53 percent of cases, such as segment-based reward tiers, life-event triggers or real-time adjustment based on transaction behaviour. Where the existing base is the focus, the work gets more precise.

Where a bank should start

The three studies suggest a sequence that does not require a major programme.

First, measure cycle time. Pick a recent campaign and count the days from the triggering customer signal to the send date, including the days spent in compliance review. That number is the bar every agent and every tool has to beat. Without it, "faster" remains a claim.

Second, give the customer record an owner. Marous recommends one accountable person with a measurable target and a unified view of account, card and loan behaviour refreshed daily. Daily is not yet real time, but it is a foundation to build on. A visible share of the planned increase in the generative AI budget belongs in that foundation.

Third, define the approved boundaries up front. Instead of reviewing every campaign individually, compliance and marketing jointly determine which offer components, wordings and contact rules are pre-approved for which customer groups. An agent that checks a draft against these rules shortens the loop; one that does not know them only lengthens it.

Fourth, measure impact with control groups. Hold back part of the audience in the next campaign and report the lift against that group. It is the cheapest way to find out whether marketing changed customer behaviour at all, and it supplies the argument for the data foundation at the same time.

For European institutions there is one more consideration: consent and purpose limitation are not a downstream check here but part of the data layer itself. Banks that carry them as attributes per customer from the start make later approval easier rather than harder.

Five takeaways

1. 63 percent of US bank marketing teams use generative AI for content, 9 percent use acting agents; adoption drops off exactly where decisions and actions happen.

2. Compliance review takes four to six weeks at some institutions; AI speeds up the step before it, and the bottleneck still sets the pace.

3. No institution surveyed has immediately AI-ready customer data; two thirds work with scattered or batch-updated data.

4. Two thirds of respondents define personalisation as selecting an audience; the customer experiences offer, timing and relevance instead.

5. The starting point is four steps: measure cycle time, give the customer record an owner, define approved boundaries up front, prove impact with control groups.

Illustration: AI-generated. AI-assisted content: We use AI technologies and automated agents in the creation of our articles, including from Microsoft, Google, OpenAI, Anthropic and other providers. Topics, editorial direction and final approval remain with our team.