If you run a digital bank in Nigeria, Indonesia, the Philippines, or any similar high-growth market, you’re probably hearing the same question from leadership on repeat: how, when, and where will you be implementing AI? After all, AI is predicted to increase productivity by 27% and improve front-office operations by somewhere between 27-35% in 2026, while potentially adding an additional $200 to $340 billion to the banking industry. Banks are racing ahead, with 92% of banks reporting an AI deployment in at least one core banking area by 2025 and the industry estimated to spend $73 billion on AI technologies by the end of 2025.
It can be tempting, when facing this challenge from leadership, to go shopping for tools. But to win the AI race in 2026, digital banks must instead focus on three things: picking focused use cases, building on the data they already have, and proving value in 90-day cycles.
We’ve covered the first two in previous pieces (how does AI help banks and which AI is best for digital banking), so this article focuses on the third by walking you through five AI projects you can realistically launch in the next three months, assuming you already have access to core data in a safe environment and a core banking platform that plays well with your tech stack.
Where relevant, we’ll also note which core and data capabilities make each project easier to deliver, because the hardest part of AI is almost always data and integration.
Before you start: what makes a 90-day AI project realistic?
Regardless of the use case, successful 90-day AI projects have three things in common:
A clear business metric
This looks like one of the following metrics:
- Reducing manual review time by 30%
- Cutting call centre volume by 20%
- Lifting approvals or repayments by X%
Production-grade data in a safe environment
You need this so that your teams can query recent and historical data off the live core so that experiments don’t risk downtime, as well as being able to move features and scores back into decision flows.
A narrow, testable scope
It’s important to be able to prioritise which delinquent customers to call first, be able to categorise and handle the top 20 FAQs via a chatbot, and score a specific fraud pattern on a specific channel.
If you already have three ingredients in place, the data and integration challenges are largely solved. Those ingredients are:
- Safe, off-core access to production-grade data
- Open APIs
- A configurable decision layer in your core
If you’re an Oradian customer, those capabilities are available through our open APIs, Custom Code, and Database Access. Database Access is a secure replica of your PostgreSQL production database that stays in sync without touching the live core.
With that all in mind, here are five projects your team can realistically ship in 90 days.
AI-assisted collections: who to call first tomorrow morning
Good for: Digital lenders, SME lenders, salary-linked products, BNPL, credit cards.
Problem: Your collections team can’t treat every delinquent account the same way. Some customers will self-cure with a gentle reminder. Others need proactive outreach before they roll into NPL. When every account gets the same script and cadence, you waste capacity and leave money on the table.
The 90-day AI project
Build a simple propensity-to-pay model that scores delinquent accounts and tells your team who is likely to self-cure as well as who is at risk of rolling into later buckets. You will need the following data already available and in the core:
- Repayment history
- Product type and tenor
- Days past due
- Salary and income
- Contact history
- Basic customer profile
How to deliver in 90 days
Days 0–30: Data & baseline
- Use your core’s database replica to pull 12–24 months of collections data.
- Define what success looks like: e.g. paid within 30 days without charge-off.
- Have your data team build a simple model to predict that outcome.
- Validate against a hold-out sample: does it rank-order accounts?
Days 30–60: Shadow mode
- Start scoring live delinquent accounts daily using that model, but only for analytics.
- Compare what your team actually did against what the scores suggested.
- Check for obvious biases and fix them.
Days 60–90: Controlled pilot
- For one cohort, route the top-risk accounts to collectors first each morning.
- Use your core’s APIs and Custom Code to surface priority tags inside the collections UI or daily worklists.
- Track things like any extra cash collected, time-to-cure, and agent productivity compared with the control group.
Be aware that if you can’t easily pull repayment history and push scores back into workflows, this 3-month cycle can easily turn into a 12-month data warehousing project. With a safe, analytics-ready replica of your core data and clear APIs, this stays a three-month sprint.
What you can expect after 90 days
- A live pilot where collectors start each day with a prioritised list instead of a flat queue
- Early evidence of improved recoveries, for example, higher cure rates or more cash collected in the pilot cohort
- Better use of collections capacity with more time spent on high-risk accounts and less on likely self-cures
- A first, documented collections model you can refine and expand to other portfolios
Smarter fraud alerts with anomaly detection
Good for: Banks seeing rising fraud on cards, accounts, or instant payments.
Problem: Rule-based fraud systems can be unreliable. You try to add more rules, but fraudsters keep routing around them, all while legitimate customers get blocked, leaving everyone unhappy.
The 90-day AI project
Layer a simple anomaly-detection model on top of your existing rules, focused on one high-risk channel, such as mobile transfers or card-not-present payments. Use it initially in shadow mode to spot suspicious behaviour without changing production decisions on day one. To achieve this, you will need the following data:
- Amount, merchant, channel, device, IP, location
- Time of day, day of week
- Historical pattern for that customer
- Outcome labels for known fraud cases
How to deliver in 90 days
Days 0–30: Pick a single focus area
- Choose one rail, such as mobile transfers, and define the fraud types you care about, for instance, you could focus on account takeovers or SIM swap follow-on fraud.
- Use your data replica to pull 6 to 12 months of transactions and confirmed fraud labels.
- Implement a straightforward model that scores transactions based on how unusual they are for that user.
Days 30–60: Shadow mode evaluation
- Start scoring every new transaction in the selected channel in real time or near-real time, but do not block anything yet.
- Compare scored high risk events with current rule triggers and confirmed fraud coming in via chargebacks or complaints.
- Tweak the thresholds to find a useful trade-off between true positives and false positives.
Days 60–90: Limited intervention
- Introduce step-up checks for the riskiest transactions only, for example, look at the top 1–2% of scores.
- Implement this via your core’s event system. For example, when a high-risk event is published, trigger extra authentication, such as requesting the consumer re-enters their PIN, and perform a soft block pending customer confirmation.
- Track reduced fraud losses on that channel against the impact on the customer experience.
To make this work, you need streaming or frequent sync of transactions into an analytics environment and the ability to trigger actions (step-up auth, notifications) from the core. A cloud-native, API-first core with an analytics-ready replica of your data gives you both.
What you can expect after 90 days
- A working anomaly-detection model in shadow or limited-traffic mode for one high-risk channel
- Clear metrics on how many additional fraud cases you could catch at a given false-positive rate
- A controlled step-up flow for the riskiest 1–2% of transactions on that channel
- A reusable blueprint to expand behavioural fraud analytics to other products or rails
An AI-assisted support copilot for your agents
Good for: Banks with growing chat volumes and complex products, for instance, multi-market, SME, credit.
Problem: Your call centre and back-office teams spend huge amounts of time looking up policy details, explaining basic product terms, and navigating multiple systems to answer simple questions, which means training new staff takes months and customers are forced to wait for longer.
The 90-day AI project
Deploy an AI copilot for support agents that suggests answers based on your own knowledge base, policies, and scripts, generates email and chat responses for review, and summarises long interactions and attaches notes to the core. For this, you will need the following data:
- Up-to-date FAQ documents
- Product terms & conditions, policy docs, internal playbooks
- A small set of real anonymised support transcripts for fine-tuning prompts
How to deliver in 90 days
Days 0–30: Content and guardrails
- Gather and clean your knowledge base content by removing outdated policies.
- Decide what the copilot can and cannot do.
- Work with your tech team to integrate a gen-AI service via API, using your content as the retrieval base.
Days 30–60: Pilot with a small group of agents
- Embed the copilot into the internal support tool UI to suggest an answer while the agent is typing and offer a summarise this chat button.
- Train a small group of agents to use it and collect the feedback.
- Track reduction in average handle time and quality scores.
Days 60–90: Expand and refine
- Roll out to a larger team if the metrics are promising.
- Add structured shortcuts, for example, you could add buttons to generate a repayment-plan explanation email as well as consistent SMS templates for payment reminders
- Feed outcomes back into your analytics layer.
The copilot doesn’t need direct access to the core to write good answers, but it does benefit from structured context to tailor replies. An API-first core makes it easy to pass that context safely into the assistant.
What you can expect after 90 days
- A support copilot embedded into your agent desktop for one or more queues
- Measured impact on handle time and quality scores for the pilot group
- Cleaner, more consistent customer communications
- A clearer view of which topics to automate next
Churn-risk scoring for your mobile users
Good for: Digital banks and wallets with big install bases but flat or declining active users.
Problem: Do you get great monthly total downloads and sign-ups, while monthly active users are stagnant? Are your marketing campaigns broad and reactive because you don’t know who is about to go dormant until they are already gone? This project is for you.
The 90-day AI project
Build a churn-risk model that identifies customers likely to lapse in the next 30–60 days and run targeted retention experiments against that group. For this, you’ll need this data:
- Logins and session activity
- Transaction types and frequency
- Salary and income
- Product holdings and tenure
- Past campaigns and engagement with them
How to deliver in 90 days
Days 0–30: Define churn and pull history
- Agree on your definition of churn.
- From your data replica, extract a year of behaviour data and label who churned.
- Build a model that predicts churn probability based on the previous 30–60 days of behaviour.
Days 30–60: Score and design offers
- Start scoring your active base weekly.
- Split high-risk customers into a test and control group.
- Design lightweight retention offers, such as fee waivers or cashback on next bill payment, top-up bonuses for savings, or pre-approved micro-loan offers for eligible salary customers.
Days 60–90: Activate the campaign and measure the results
- Use push notifications, in-app messages, SMS or email to serve offers to the test group.
- Measure the re-activation rate against control, incremental revenue against incentive cost, and the impact on longer-term retention beyond 30 days.
This project lives and dies on timely access to behavioural data. With an analytics-ready replica of your core data, for example, Oradian’s Database Access, and proper event feeds from your core, your data team can iterate quickly
What you can expect after 90 days
- A working churn-risk score that updates weekly on your active base
- At least one tested retention playbook for high-risk users
- A list of the 3–5 behavioural signals that best predict churn in your market, which you can feed into future product and marketing decisions
AI-driven SME lead and pre-qualification scoring
Good for: Banks and lenders targeting SMEs, merchants, or micro-entrepreneurs who focus strongly on their digital offerings.
Problem: Is your SME pipeline messy, with lots of leads and very few approvals? Do your relationship managers spend time on businesses that don’t qualify, while strong prospects wait? If youwant to be faster and more targeted without loosening risk standards, this one is for you.
The 90-day AI project
Deploy a lead pre-qualification scoring model that ranks inbound SME leads based on their likelihood of approval and profitability. This will help your sales teams know who to call first and what to offer. For this, you will need data from approved and declined SME customers, centering around:
- Business sector and size
- Turnover and cash-flow patterns
- Product combinations taken
- Time from lead to approval
- Performance outcomes
How to deliver in 90 days
Days 0–30: Build the score
- Use your core and CRM data to build a profile of SMEs that are a good fit.
- Train a model to predict the probability of approval and expected value.
Days 30–60: Pilot with a digital channel
- Start applying the score to new inbound leads from one channel.
- Present the score as a simple band in your dashboard.
- Capture feedback: does the score feel intuitive? Are the high band leads genuinely better?
Days 60–90: Embed this project into your offering channels
- Route high-score leads to your strongest relationship managers or fastest digital flows.
- For low-score leads, offer simpler or smaller products by default.
- Track approval rates, time-to-yes, and early-stage performance.
This project is where ecosystem growth connects back to your engine, because to score SME leads coming from partners, you really need a core that can ingest partner data through APIs and a data layer that can join it with your own history. That’s what makes tools like Database Access so vital.
What you can expect after 90 days
- A live SME lead score visible to sales or RM teams in at least one channel
- Improved focus on high-quality leads
- Early improvements in approval rates and time-to-yes for the pilot channel
- A clearer picture of your ideal SME customer based on data
Making these projects real on Oradian
If you’re already using Oradian, most of the hard parts of these 90-day projects are already in place. With Oradian, you already have a cloud-native core with APIs, which makes it straightforward to plug in AI scores and trigger actions without deep rewrites. You may even use Custom Code and events, which let you encode decision rules and connect to external AI services with minimal vendor dependency. If you are using Database Access, you have a secure, always-up-to-date replica of your production PostgreSQL database so data scientists and analysts can work safely off-core, and if you don’t, drop vanda.jirasek@oradian.com a message to find out how Database Access can take your Oradian usage to the next level.
That doesn’t mean activating an AI project becomes easy, after all, you still need clear goals, good governance, and the right people, but it does mean you can focus on designing the right experiments, not wrestling with missing data and ancient legacy systems.
Where to start on Monday
Here is exactly what to do on Monday: ask yourself which decision, if improved by 10-20%, would matter most in the next 12 months?
Then consider whether you have the data available to improve that decision, if you can access that data safely off your core, and if you can feed scores back into live journeys.
If the answer is yes to all those questions, you’re ready for your first 90-day AI sprint.
Ready to ship your first 90-day AI project?
If you’re serious about AI in 2026, the hardest step is deciding on a single concrete project with a deadline, an owner, and a business outcome.
If you’re already on Oradian, you’re closer to AI success than you think.
You have:
- A cloud-native core that can plug into AI services via APIs and events
- The ability to turn on access to Database Access, which will give your teams safe, SQL-level access to production-grade data
- The ability to encode new logic and workflows with Custom Code instead of waiting on long vendor cycles
If you’d like help mapping out which project to start with first, whether you’re a current client or not, drop vanda.jirasek@oradian.com an email with which project you’re considering, the market you’re operating in, and one metric you’d like to see growth with in the next 90 days.
We’ll respond with a simple outline of what a 90-day AI sprint could look like on Oradian in your context, including the data you’ll need, the teams to involve, and where your core and data layer can do the heavy lifting.
