AI for Banking in Kenya
Where every automated decision has to be explainable to a regulator.
What is banking?
AI for banking covers onboarding and document handling, alert triage, and customer service volume, applied inside a regulated environment where any decision affecting a customer must be explainable and reviewable.
Banking is the sector where the explainability requirement is not optional. A customer refused an account or flagged for review can ask why, and the answer cannot be that a model decided.
That rules out a whole class of otherwise attractive applications and leaves a large one. Onboarding document handling, alert triage, and service volume are all places where AI does real work without making the decision.
The largest practical gain in most Kenyan banks is alert triage. Transaction monitoring generates far more alerts than analysts can review properly, and most of them are noise.
What we will not let a system decline
Duara works with banks, microfinance institutions and SACCOs from Nairobi, on the operational side rather than on credit or risk models.
We will not build a system that declines a customer without an explainable reason. Where a decision affects somebody access to financial services, a model output that cannot be explained is a problem for your compliance team long before it is a problem for the customer.
Who this is for
- Onboarding is slow and document heavy
- Alert volume exceeds analyst capacity
- Customer service volume is unmanageable
- Regulators are asking about your use of AI
Industries we serve
- Commercial banks
- Microfinance institutions
- SACCOs
- Digital lenders
- Payment providers
- Investment firms
Problems this solves
These are volume problems in a regulated environment, which is a narrower design space than it first appears.
- Onboarding that takes days. Document collection and verification is manual, and customers abandon the process while waiting.
- Alert volume beyond capacity. Transaction monitoring produces more alerts than analysts can review, so review quality drops rather than volume.
- Service queues at peak. The same questions arrive in volume and staff cannot reach the customers with genuine problems.
- Regulator questions about AI. Supervisors are asking what you use and how it is controlled, and the answer is not documented.
What we deliver
Document handling and verification support during onboarding, alert triage that prioritises rather than closes, service automation for routine questions, and documentation of what the systems do and why.
Document handling and verification support
Alert triage that prioritises rather than closes
Routine service questions handled automatically
Explainable reasoning on every output
Documentation suitable for a supervisor
Human review preserved at every decision point
Features
- Nothing declines a customer. Systems prepare and prioritise. Decisions affecting access to financial services stay with people who can explain them.
- Triage prioritises, never closes. Alerts are ordered by likely significance and none are closed automatically, because a closed alert is an unreviewed one.
- Explainability as a requirement. Every output carries its reasoning, since a supervisor asking why will not accept that a model decided.
- Documented for supervision. What the system does, its limits and its controls, written for a regulator rather than for a brochure.
- Built for the regulated case. Applications that cannot be explained are excluded at design rather than discovered during an examination.
Benefits
- Faster onboarding. Document handling support shortens the period where customers abandon the process.
- Analysts review what matters. Prioritised alerts mean scarce analyst attention goes to the ones most likely to be genuine.
- Service reaches real problems. Routine volume handled automatically frees staff for customers who need a person.
- Supervisory questions answerable. Documentation exists before it is requested rather than being assembled under pressure.
Technologies we use
- Document processing. Reading and checking onboarding documents, with uncertain cases routed to a person.
- Alert prioritisation. Ordering monitoring alerts by likely significance without closing any.
- Service automation. Handling routine customer questions with escalation to staff.
- Explainability. Producing the reasoning behind every output in a form a person can review.
- Control documentation. Recording what systems do, their limits and their oversight, for supervisory review.
Alert triage is the highest value application in most Kenyan banks and the least discussed. The problem is not detecting suspicious activity, it is that the volume of alerts makes proper review of any of them impossible.
How the process works, step by step
Banking projects usually take eight to sixteen weeks, with compliance involved from the first conversation rather than at approval.
-
Step 1
Free operations review
We look at onboarding times, alert volumes and where service pressure actually sits.
-
Step 2
Compliance in early
Your compliance and risk teams are involved from the start, since late involvement means rework.
-
Step 3
Explainability designed in
How each output will be explained is settled before building, not retrofitted for an examination.
-
Step 4
Parallel running
The system runs alongside existing process so outputs can be compared before anything is relied on.
-
Step 5
Documented and handed over
Controls documentation is produced for supervision as part of delivery.
Quality assurance and how we work
The commitments attached to systems operating inside a regulated bank.
- No system declines a customer. Decisions affecting access to financial services stay with people who can explain them to the customer and to a supervisor.
- Alerts are never closed automatically. Triage orders them by likely significance. A closed alert is an unreviewed one and that is the wrong risk to take.
- Explainability is designed in, not added. Retrofitting reasoning for an examination produces documentation that describes what you wish the system did.
- Compliance is involved from the first conversation. Late involvement means rework, and in this sector the rework is usually total.
- We decline applications that cannot be explained. Where a use case only works as an unexplainable model output, we say it is unsuitable rather than building it.
Frequently asked questions about banking
Can AI decide whether to open an account?
No, and we will not build that. A customer refused access can ask why, and the answer cannot be that a model decided.
What can it do in onboarding then?
Document handling and verification support, which is where the days are lost, with uncertain cases routed to a person.
Where is the biggest gain?
Alert triage in most Kenyan banks. The problem is not detection, it is that alert volume makes proper review of any of them impossible.
Will it close alerts automatically?
Never. It orders them by likely significance. A closed alert is an unreviewed one and that is the wrong risk to accept.
What about credit scoring?
We do not build credit decision models. That is a specialist regulated area and adjacency to it would not make us competent at it.
How do you handle explainability?
It is designed in before building. Retrofitting reasoning for an examination produces documentation describing what you wish the system did.
When should compliance be involved?
From the first conversation. Late involvement means rework, and in banking the rework is usually total.
Will our regulator accept this?
We produce controls documentation for supervisory review and keep decisions with people. Whether a specific supervisor is satisfied is their judgement and we would engage early.
How long does a banking project take?
Eight to sixteen weeks, with parallel running before anything is relied on.
Can it handle customer service volume?
Yes, for routine questions, which frees staff for customers with genuine problems rather than replacing the service.
Does customer data leave our systems?
No. Banking deployments run inside your own infrastructure, which is usually a supervisory expectation as well as a preference.
Can SACCOs and microfinance use this?
Yes, and alert and onboarding volumes are proportionally just as pressing there, often with fewer staff to absorb them.
What about fraud detection?
We support triage and prioritisation. Determining that something is fraud stays a human decision with consequences we will not automate.
Does it work with our core banking system?
Most common ones, reading and writing through supported interfaces rather than around them.
What if the system is wrong?
A person was always making the decision, which is the design point. The system prepares and prioritises rather than concluding.
Can it read Kenyan identity documents?
Yes, with accuracy tested on your own document images rather than quoted from a benchmark.
Will it reduce our compliance headcount?
That is your decision. Most institutions redirect analyst time to better review rather than reducing it, given alert volumes.
Can it help with regulatory reporting?
With assembly and consistency checking. The submission and its accuracy stay your responsibility.
Who can see customer records?
Only your own authorised staff. Nothing is copied out and no public AI model is trained on customer data.
What does a banking project cost?
Scope and integration complexity determine it, quoted after the operations review.
Glossary of terms
Banking terms in plain language.
- Onboarding
- The process of taking on a new customer, including identity and document checks.
- Alert triage
- Ordering monitoring alerts by likely significance so analyst attention goes where it matters.
- Explainability
- Being able to state why a system produced a particular output, in terms a person can review.
- False positive
- An alert that turns out to be legitimate activity, which is most of them.
- Human in the loop
- A person making the decision, with the system preparing and prioritising only.
- Controls documentation
- The written description of what a system does and how it is overseen, for supervisory review.
Related services in Industries
Talk to us about banking
Ask how many monitoring alerts were raised last month and how many were genuinely reviewed. The gap is usually the strongest case.
Start your project